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.1005086 | A Non-enveloped Virus Hijacks Host Disaggregation Machinery to Translocate across the Endoplasmic Reticulum Membrane | Mammalian cytosolic Hsp110 family, in concert with the Hsc70:J-protein complex, functions as a disaggregation machinery to rectify protein misfolding problems. Here we uncover a novel role of this machinery in driving membrane translocation during viral entry. The non-enveloped virus SV40 penetrates the endoplasmic reticulum (ER) membrane to reach the cytosol, a critical infection step. Combining biochemical, cell-based, and imaging approaches, we find that the Hsp110 family member Hsp105 associates with the ER membrane J-protein B14. Here Hsp105 cooperates with Hsc70 and extracts the membrane-penetrating SV40 into the cytosol, potentially by disassembling the membrane-embedded virus. Hence the energy provided by the Hsc70-dependent Hsp105 disaggregation machinery can be harnessed to catalyze a membrane translocation event.
| How non-enveloped viruses penetrate a host membrane to enter cells and cause disease remains an enigmatic step. To infect cells, the non-enveloped SV40 must transport across the ER membrane to reach the cytosol. In this study, we report that a cellular Hsp105-powered disaggregation machinery pulls SV40 into the cytosol, likely by uncoating the ER membrane-penetrating virus. Because this disaggregation machinery is thought to clarify cellular aggregated proteins, we propose that the force generated by this machinery can also be hijacked by a non-enveloped virus to propel its entry into the host.
| Protein misfolding and aggregation compromise cellular integrity. Cells in turn deploy powerful molecular chaperones to promote protein folding, prevent aggregation, and in some instances, re-solubilize the aggregated toxic species to rectify these problems and maintain proper cellular function [1–3]. A cell’s ability to effectively mount a response to protein misfolding and aggregation despite acute or sustained environmental stresses has major implications in the development of protein conformational-based diseases [4,5].
The 110 kDa heat shock protein (Hsp110) family, including Hsp105, Apg1, and Apg2, are cytosolic chaperones that belong to the Hsp70 superfamily [6–10]. In addition to serving housekeeping roles during protein homeostasis, this protein family has been linked to wide ranging cellular processes including cell migration [11], spindle length control [12], and molecular scaffolding [13]. Importantly, as the Hsp110 family has also been implicated in many protein misfolding diseases, such as amyotrophic lateral sclerosis [14,15], prion disease [16], Alzheimer’s disease [17], cystic fibrosis [18], and polyglutamine disease [19,20], clarifying its precise mechanism of action in cells is paramount.
At the molecular level, Hsp110 acts as a nucleotide exchange factor (NEF) against Hsp70 and the constitutively expressed Hsc70 [7,8], which was used in this study. A NEF triggers nucleotide exchange of ADP-Hsc70, generating ATP-Hsc70 that displays a low affinity for its substrate [21]. This reaction reverses the effect of a J-protein, which uses its J-domain to stimulate the ATPase activity of ATP-Hsc70, forming ADP-Hsc70 that binds to its substrate with high affinity. Thus, a typical substrate-binding and release cycle by Hsc70 is coordinately regulated by a NEF and a J-protein.
Structurally, Hsp110 harbors an N-terminal ATPase domain similar to Hsc70, followed by a peptide-binding domain, an acidic loop, and a C-terminal helix domain thought to sub-serve a “holdase” function [6]. Strikingly, beyond simply acting as a NEF, reports suggest that Hsp110, in conjunction with the Hsc70:J-protein complex, can function as a disaggregase against model substrates [7,22–25]. However, whether Hsp110 and its chaperone activity acts on a physiologically relevant substrate as part of a cell’s protein quality control response, or is exploited to promote other unanticipated biological processes, is unclear. Here we demonstrate a novel and unexpected role of Hsp110 in driving membrane translocation of a virus.
To cause infection, the non-enveloped polyomavirus (PyV), typified by the classic simian PyV SV40, traffics from the host cell surface to the ER from where it penetrates the ER membrane to reach the cytosol [26–29]. In the cytosol, the virus moves into the nucleus to enable transcription and replication of the viral genome, causing lytic infection or cellular transformation. Our understanding of how SV40 is extracted into the cytosol from the ER is slowly unraveling.
Structurally, SV40 is composed of 360 copies of the major coat protein VP1 arranged as 72 pentamers, with each pentamer engaging either of the internal hydrophobic protein VP2 or VP3. The pentamers are assembled as a 45 nm-diameter icosahedral particle which in turn encapsulates its viral DNA genome [30,31]. Upon reaching the ER [32–34], the virus hijacks ER-resident isomerase and reductase that impart conformational changes to the viral particle to expose its hydrophobic proteins VP2 and VP3 without triggering massive disassembly [35–41]. These remodeling reactions generate a hydrophobic particle that binds to and integrates into the ER membrane [35,37,42]. However, the molecular mechanism by which the membrane-embedded intact hydrophobic virion is extracted into the cytosol and disassembled is not entirely clear.
A clue to unraveling this mystery emerged when three ER membrane J-proteins called DnaJB12 (B12), DnaJB14 (B14), and DnaJC18 (C18) were reported to support ER-to-cytosol translocation of SV40 and the related human BK PyV (BKV) [38]. Because the J-domain of B12, B14, and C18 orient towards the cytosol, we hypothesized that they recruit the cytosolic Hsc70 machinery to the ER membrane where it engages and ejects SV40 from the ER into the cytosol. In fact, we identified an Hsc70 co-chaperone called SGTA that forms part of a poorly-defined ER membrane-localized Hsc70 complex (which comprises of at least B14-Hsc70-SGTA) that supports extraction of PyV into the cytosol [43]. However, the precise molecular nature of this machinery as well as how SV40 is ultimately ejected into the cytosol from the ER is not completely defined. In this study, we identify the Hsp110 family member Hsp105 as a novel B14 binding partner. ER membrane juxtaposition of Hsp105 enables it to interact with and promote the extraction of membrane-embedded SV40 into the cytosol, thereby preparing the virus for infection. Our data support a scenario whereby Hsp105 synergizes with the B14-Hsc70-SGTA to extract SV40 into the cytosol, potentially by disassembling the virus.
As B14 mediates SV40 ER membrane translocation [38], we reasoned that pinpointing B14’s interacting partners might reveal how the ER membrane-penetrating SV40 is extracted into the cytosol. Combining immunoaffinity purification and mass spectrometry, we identified the cytosolic co-chaperone SGTA as a B14 steady-state binding partner that is part of an Hsc70 complex which mobilizes SV40 from the ER to the cytosol [43]. However, the precise biochemical mechanism by which the B14-Hsc70-SGTA complex promotes SV40 cytosol entry is unclear. To characterize additional interaction partners of this complex, we asked if novel cellular factors are recruited to B14-Hsc70-SGTA during SV40 infection. To this end, we used a HEK 293T cell line stably expressing 3x-FLAG tagged (hereon FLAG-tag is abbreviated as ‘F’) B14 (B14-3xF) that expresses a similar level of B14-3xF as endogenous B14 (Fig 1A). This human cell line, used previously to identify SGTA as a B14-interacting partner [43], supports SV40 infection when supplemented with the cell-surface SV40 receptor ganglioside GM1.
B14-3xF in the whole cell extract (WCE) derived from cells infected with or without SV40 was immunopurified and eluted (see Methods). The eluted material was subjected to SDS-PAGE and silver staining. Distinct bands were excised and analyzed by mass spectrometry. Interestingly, some of the band intensity appeared higher in samples obtained from SV40-infected cells when compared to the uninfected sample (Fig 1B). Results from mass spectrometry revealed the presence of B14, SGTA, and Hsc70/Hsp70, as anticipated. While numerous additional proteins were also identified, we focused our attention on the Hsp110 family members (Hsp105, Apg1, and Apg2 in Fig 1B, highlighted in bold) because they are established Hsc70/Hsp70 interacting partners [7,8]. To validate our mass spectrometry results, samples in Fig 1B were initially immunoblotted using an antibody against Hsp105 (see Table 1 for list of all antibodies used in this study), which confirmed Hsp105 co-precipitated with B14-3xF (Fig 1C, first panel). A parental 293T cell that does not express B14-3xF was used as a negative control (Fig 1C, lane 1). These findings demonstrate that B14-3xF interacts with Hsp105. We found a modest increase in the level of Hsp105 that co-precipitated with B14-3xF in SV40-infected cells when compared to uninfected cells (Fig 1C, first panel, compare lane 3 to 2), suggesting that Hsp105 is likely recruited to B14 during infection. We also established endogenous Hsp105-B14 interaction in the simian CV-1 cells by co-immunoprecipitation, with the association moderately enhanced during SV40 infection (Fig 1D, first panel, compare lane 3 to 2). This experiment was performed in the presence of the membrane permeable, amine-reactive, and thiol-cleavable crosslinker DSP (dithiobis(succinimidyl proprionate)) in order to stabilize transient or weak protein-protein interactions. CV-1 cells were analyzed because they are the normal permissive cells used to study SV40 infection. All studies hereafter were performed in this cell line unless otherwise indicated. We attempted but were not able to establish an interaction between B14 and Apg1/Apg2, nor with other proteins identified in the mass spectrometry results. For these reasons, we focused our efforts on Hsp105 in the rest of this study.
To assess if the B14-Hsp105 interaction requires an intact B14 J-domain, we used a B14 J-domain mutant (B14 H136Q) that was previously shown to be defective in Hsc70 binding [44]. Precipitation of transiently transfected FLAG-tagged B14 (F-B14) or H136Q B14 (F-B14 H136Q) in CV-1 cells revealed that the mutant does not interact with endogenous Hsp105 or Hsc70 (Fig 1E, first and second panels), even though more mutant protein was precipitated. This result suggests that B14’s interaction with Hsp105 requires an intact J-domain, but it does not indicate the interaction is direct or mediated by Hsc70. To test this, we depleted Hsc70 from CV-1 cells (using an siRNA directed specifically against Hsc70) and found that depletion of Hsc70 did not impair binding between endogenous B14 and transiently transfected FLAG-tagged wild-type (WT) Hsp105 (Hsp105 WT-F) (Fig 1F, first panel; see Table 2 for list of all primers used in this study). These observations suggest that B14 utilizes its J-domain to bind to Hsp105 directly, although we cannot rule out the possibility that the related Hsp70 would provide the physical link between B14 and Hsp105 when Hsc70 is absent. Knockdown of SGTA also did not affect Hsp105 WT-F’s interaction with endogenous B14 or Hsc70 (Fig 1G, first and second panels), indicating that this co-chaperone unlikely controls Hsp105’s ability to complex with B14 and Hsc70. To further evaluate if the Hsp105 interaction with B14-Hsc70-SGTA complex was specific, we transiently transfected S-tagged WT Hsp105 (Hsp105 WT-S) or a cytosolic non-Hsp110 family NEF called HspBP1 (HspBP1-S). Whereas endogenous Hsc70 was pulled down by precipitating either Hsp105 WT-S or HspBP1-S (Fig 1H, third panel, lanes 2 and 3), only precipitation of Hsp105 WT-S pulled down endogenous B14 and SGTA (Fig 1H, first and second panels). We conclude that Hsp105 is specifically recruited to the ER membrane by anchoring to B14.
To test whether Hsp105 is required for SV40 infection, CV-1 cells were transfected with either of two distinct siRNAs targeted against Hsp105 (Hsp105 #1 and #2), or a negative control siRNA (ctrl) (see Table 3 for list of all siRNAs used in this study). Immunoblotting of the resulting WCE demonstrated that the Hsp105-specific siRNAs efficiently knocked down endogenous Hsp105, without affecting the levels of the other cytosolic chaperones (Fig 2A, top panel) and without triggering XBP1 splicing (see Table 4; Fig 2A, bottom panel), a sensitive readout of cellular ER stress induction. Under the knockdown condition, we evaluated SV40 infection by scoring for presence of the virally encoded large T antigen (TAg) in the host nucleus by immunostaining. Hsp105 knockdown inhibited SV40 infection by 70–75% compared to control siRNA or a siRNA directed against HspBP1 (Fig 2B). A similar infection block was also observed for BKV when Hsp105 is knocked down (Fig 2C). Likewise, depletion of Hsp105 in the simian BSC-1 cells (S1A Fig) reduced SV40 infection by 60% (S1B Fig). The residual infection found when Hsp105 is down-regulated could be due to presence of other Hsp110 family members such as Apg1 or Apg2, or in the case of BSC-1 cells, the incomplete knockdown of Hsp105.
As previous studies reported that the Hsp110 family acts as a NEF against Hsc70/Hsp70 [7,25,45], we asked if Hsp105’s nucleotide exchange activity is important in supporting SV40 infection. For this, we generated two Hsp105 mutants based on previous characterization of mammalian Hsp110. A mutant Apg2, referred to as the N619Y/E622A variant, is defective in its nucleotide exchange activity because it cannot interact with Hsc70 [24], consistent with a crystal structure of the yeast homolog demonstrating that these two residues are positioned at the interface required for interaction with Hsc70 [9,46]. Because N619 and E622 are highly conserved residues in the Hsp70 family/superfamily (Fig 2D, bottom panel), we generated the corresponding mutant (N636Y/E639A) in Hsp105 with FLAG tag (Hsp105 NE*-F). In the second mutant, we took advantage of insights from studies on Grp170, the ER-resident Hsp110 family NEF [47–50]. In Grp170, mutating glycine at position 41 to leucine located within the ATP-binding pocket at its N-terminal ATPase domain renders the G41L mutant defective in its nucleotide exchange activity and impairs binding to the ER-resident Hsp70 BiP [51]. Since G41 is also conserved among Hsp70 family/superfamily (Fig 2D, top panel), we generated the corresponding mutant (G9L) in Hsp105 with FLAG tag (Hsp105 G*-F).
We first assessed the behavior of Hsp105 NE*-F and G*-F. To act as a NEF against Hsc70, Hsp105 must bind to ATP, engage Hsc70, and then stimulate nucleotide release from Hsc70 [48]. Accordingly, we purified FLAG-tagged proteins from transfected 293T cells (Fig 2E). To determine the ATP-binding affinity of the Hsp105 mutants, purified proteins were incubated with ATP-conjugated agarose beads and eluted samples were analyzed by immunoblot. As predicted, Hsp105 WT-F and NE*-F but not G*-F bind to ATP (Fig 2F, first panel, compare lanes 2 and 3 to 4). By co-immunoprecipitation, neither Hsp105 NE*-F nor G*-F exhibited any significant binding affinity for Hsc70 (Fig 2G, first panel). Thus, while Hsp105 NE*-F binds to ATP, it cannot interact with Hsc70 due to mutations at the Hsc70 binding interface. By contrast, Hsp105 G*-F does not bind to ATP, likely leading to a defective ATPase activity essential for its interaction with Hsc70.
To evaluate the nucleotide exchange activity of Hsp105 mutants, we used a radioactive nucleotide release assay. Briefly, Hsc70 was preloaded with labeled nucleotide [α-32P]ATP which spontaneously hydrolyzes to generate [α-32P]ADP-Hsc70. Individual FLAG tagged protein along with unlabeled ATP was added to [α-32P]ADP-Hsc70 to induce nucleotide release. The amount of labeled ADP that remain bound to Hsc70 was evaluated by thin layer chromatography. Using this assay, we found that only Hsp105 WT-F but not the Hsp105 mutants triggered [α-32P]ADP release from Hsc70 (Fig 2H, compare lane 2 to 3 and 4), demonstrating that both Hsp105 mutants are indeed defective in their nucleotide exchange activity.
We performed knockdown followed by rescue experiments to probe the requirement of Hsp105’s nucleotide exchange activity during SV40 infection. Cells were initially transfected with control or Hsp105 #1 siRNA, followed by transfection with GFP-F or an siRNA-resistant Hsp105 WT-F, NE*-F, or G*-F. Cells were subsequently infected, fixed, and scored for the presence of TAg only in FLAG-expressing cells. Importantly, when compared to expressing the control GFP-F, only re-expressing Hsp105 WT-F but not NE*-F or G*-F under the Hsp105 knockdown condition restored infection (Fig 2I, right graph). As controls, expressing Hsc70, SGTA, or HspBP1 when Hsp105 is down-regulated did not significantly restore infection. We conclude that the nucleotide exchange activity of Hsp105 is required to promote successful SV40 infection. We note that the expression level of siRNA resistant Hsp105 WT-F in a cell extract derived from the total pool of Hsp105-depleted cells is less compared to the level of endogenous Hsp105 in control cells (Fig 2I, left first panel, compare lane 3 to 1). Given that the transfection efficiency in CV-1 cells is low (~20%), the re-expressed Hsp105 level is likely comparable to endogenous Hsp105. This observation is consistent with our finding that SV40 infection is only restored (but not enhanced) in Hsp105-reexpressed cells when compared to control cells.
As Hsp105 localizes to the ER membrane and is important in SV40 infection, we hypothesized that it might promote the extraction of SV40 into the cytosol from the ER membrane. To test this, we monitored arrival of SV40 into the cytosol from the ER membrane using a semi-permeabilized cytosol arrival assay established previously [37,39]. In this assay, siRNA transfected CV-1 cells were harvested post-infection and treated with a low concentration of digitonin to semi-permeabilize the plasma membrane without affecting internal membranes. Subsequent centrifugation generates two fractions, a supernatant fraction that harbors cytosolic proteins and virus that reaches the cytosol (referred as “cytosolic” fraction), and a pellet fraction that contains membranes including the ER, as well as associated viral particles (referred as “membrane” fraction). The cytosolic Hsp90 was found predominantly in the cytosolic fraction (Fig 3A, second panel), while the ER marker protein disulfide isomerase (PDI) was found exclusively in the membrane fraction (Fig 3A, seventh panel), verifying the integrity of the fractionation procedure. Importantly, using this assay, silencing Hsp105 by either siRNA (Fig 3A, fourth panel) markedly decreased the VP1 level in the cytosol (Fig 3A, first panel, compare lanes 2 and 3 to 1; the VP1 band intensity is quantified in Fig 3B). These data indicate that Hsp105 exerts an important role in promoting cytosol arrival of SV40 from the ER. The residual VP1 observed in the cytosol fraction when Hsp105 is knocked down is consistent with the low infection observed when this chaperone is down-regulated (Fig 2B).
To test if trafficking of SV40 from the cell surface to the ER is controlled by Hsp105, we isolated ER-localized SV40 (which includes viral particles in the ER lumen and those integrated into the ER bilayer during membrane penetration) using a previously established, Triton X-100 extraction protocol [39,43]. Using this strategy, we found that the level of ER-localized SV40 was unperturbed by down-regulating Hsp105 (Fig 3C, top panel; the VP1 band intensity is quantified in bottom panel). Hence, the block in cytosol arrival of SV40 when Hsp105 is knocked down (Fig 3A and 3B) cannot be attributed to a disruption in trafficking of the viral particle from the plasma membrane to the ER. This finding strengthens our proposal that Hsp105 regulates cytosol arrival of SV40 from the ER.
Another toxic agent that uses the ER-to-cytosol membrane translocation pathway to cause disease is cholera toxin (CT) [52,53]. However, in this case, knockdown of Hsp105 (Fig 3D, third panel) did not affect cytosol arrival of CT’s catalytic CTA1 subunit from the ER (Fig 3D, first panel). This finding not only indicates that Hsp105 specifically controls cytosol entry of SV40 and not another toxic agent, but also suggests that silencing Hsp105 did not globally disable cytosol entry machineries from the ER.
In addition to the cell-based approach, we used imaging strategies to further support the idea that Hsp105 promotes extraction of ER-localized SV40 into the cytosol. In the ER, SV40 accumulates in discrete foci where specific ER membrane proteins important for virus infection are also recruited [37,43,54]. An example of the SV40-induced foci is presented in S2 Fig, in which infected cells were immunostained for the viral proteins VP1 and VP2/3 (first and second row, white arrows); these foci colocalize with the ER membrane protein BAP31 which is also essential in virus infection [37]. As all BAP31 foci colocalize with ER-localized SV40, BAP31 foci serves as a convenient marker for SV40-containing foci. Under higher magnification, the intensity surface plots of several representative foci indicate that each focus contains several discrete spots, likely corresponding to multimeric viral particles (S3 Fig, surface plots on the right).
We hypothesize that these SV40-induced foci represent the cytosol entry site for the virus based on several lines of evidence. First, the VP2/3-exposed, membrane penetration-competent SV40 colocalizes with the foci [54] (S2 Fig, second row). Second, specific ER membrane proteins that facilitate SV40 membrane penetration, including B14/B12 (third and fourth rows) and BAP31 (all rows), are recruited to the foci [37,43]. Third, foci formation kinetics temporally parallels SV40 cytosol entry [43]. And fourth, SV40 mutants that cannot penetrate the ER membrane to access the cytosol fail to induce foci [37,54].
We reasoned that if the foci harboring SV40 represent the site from where ER-localized virus enters the cytosol, depleting Hsp105 (which prevents SV40 extraction into the cytosol, Fig 3A and 3B) should trap the virus in the ER membrane, thereby enhancing the foci structure. To evaluate this, cells transfected with either a control or Hsp105 #1 siRNA were infected with SV40 (Fig 3E). When compared to control, silencing Hsp105 increased the number of cells with at least one BAP31-positive foci by approximately 2-fold (Fig 3E; quantified in Fig 3F); a similar 2-fold increase was also observed using Hsp105 siRNA #2 (Fig 3F). Additionally, in the Hsp105 knockdown cells, there were more foci per cell and the size of the foci in the knockdown cells appears larger when compared to control (Fig 3E, see insert for 2x enlarged dotted box; quantified in Fig 3G). These data suggest that depleting Hsp105 entraps SV40 in the ER because it cannot be extracted into the cytosol, consequently enhancing foci formation. These imaging approaches were consistent with the cell-based studies, strongly suggesting that Hsp105 plays a key role in extracting SV40 from the ER into the cytosol.
We used a gain-of-function strategy to assess Hsp105’s role in extracting SV40 into the cytosol. The CV-1 derived COS-7 cells were transfected with GFP-S, HspBP1-S, or Hsp105 WT-S, and subjected to the cytosol arrival assay as described above. COS-7 cells were used because of their ability to support a high DNA transfection efficiency required for this experiment. We found that over-expressing Hsp105 WT-S (Fig 4A, fourth panel, compare lane 3 to 2 and 1) but not HspBP1-S stimulated SV40 arrival to the cytosol (Fig 4A, first panel; the VP1 band intensity is quantified in Fig 4B). These findings demonstrate that Hsp105 stimulates SV40 extraction into the cytosol. Not surprisingly, Hsp105 WT overexpression also enhanced SV40 infection (Fig 4C, compare second to first bar). This stimulation requires Hsp105’s nucleotide exchange activity as overexpressing Hsp105 NE*-F or G*-F did not robustly enhance infection (Fig 4C, compare third and fourth bars to second bar), and is specific because overexpressing neither Hsc70, SGTA, nor HspBP1 stimulated infection.
Again we used an imaging approach to further strengthen the idea that Hsp105 overexpression promotes extraction of ER-localized SV40 into the cytosol. We reasoned that, if the foci harboring SV40 represent the site from where ER-localized virus enters the cytosol, overexpressing Hsp105 (which stimulates virus extraction into the cytosol, Fig 4A and 4B) should correspondingly decrease the virus-induced foci. To test this, CV-1 cells expressing Hsp105 WT-S or the control GFP-S were infected with SV40, fixed, and immunostained (Fig 4D). Strikingly, whereas VP1-positive foci that colocalize with the BAP31 foci are found in GFP-S expressing cells (Fig 4D, first row, see insert for 2x enlarged dotted box), a dramatic decrease in these foci was observed in cells overexpressing Hsp105-S (Fig 4D, second row). We quantified these effects by scoring for presence of BAP31-positive foci in cells expressing GFP-S or Hsp105 WT-S; in our quantification, any cell that displays at least one BAP31-positive focus is scored positive. Our analyses revealed that overexpressing tagged Hsp105 decreased the number of cells containing at least one BAP31-positive focus by approximately 3-fold (Fig 4E, first two bar graphs), suggesting that overexpressing Hsp105 impairs foci formation. As Hsp105 overexpression also stimulates SV40 cytosol arrival (Fig 4A and 4B) and infection (Fig 4C), the ability of overexpressed Hsp105 to decrease foci formation likely reflects efficient Hsp105-dependent extraction of SV40 into the cytosol essential for infection.
To ascertain if Hsp105’s nucleotide exchange activity is necessary to decrease foci formation, cells were transfected with either Hsp105 NE*-F or G*-F. Under these conditions, no significant reduction in the number of cells containing at least one BAP31-positive foci was found (Fig 4E, compare third and fourth bar graphs to second), suggesting that Hsp105’s nucleotide exchange activity plays a role in this process. This finding is consistent with the observation that overexpressing Hsp105 NE*-F or G*-F did not markedly stimulate infection (Fig 4C). As controls, overexpressing Hsc70, SGTA, or HspBP1 failed to diminish the number of cells harboring the BAP31-positive foci (Fig 4E, compare fifth and seventh bar graphs to second), in complete agreement with their inability to stimulate infection when overexpressed (Fig 4C). Hence, Hsp105 but not other chaperones promotes extraction of SV40 into the cytosol, a reaction that appears to impair the stable formation of the virus-containing foci.
While depleting Hsp105 enhanced foci formation (Fig 3E and 3F), we tested whether re-introducing Hsp105 under this condition can reverse the effect. Accordingly, Hsp105-depleted cells transfected with the siRNA-resistant Hsp105 construct were infected and processed as before. Using this strategy, we found that while Hsp105 knockdown increased the number of cells containing at least one BAP31-positive foci compared to control (Fig 4F, first two bar graphs), expressing WT Hsp105 under this condition reversed the effect (Fig 4F, second and third bar graphs). As controls, expressing either of the Hsp105 mutants, SGTA, or HspBP1 under the Hsp105 knockdown condition did not promote loss of foci, whereas Hsc70 overexpression modestly impaired foci formation (Fig 4F, fourth to eight bar graphs). Hence Hsp105’s nucleotide exchange activity is essential for regulating foci formation. Because expressing WT but not mutant Hsp105 in Hsp105-depleted cells also restored infection (Fig 2I), these data further strengthen the functional connection between foci formation and productive infection.
Mechanistically, we envision that Hsp105 binds to SV40 at the ER-cytosol interface to extract the membrane-embedded virus into the cytosol. To determine if the Hsp105-SV40 interaction is direct, we incubated purified SV40 with purified proteins (Fig 2E); SGTA-F or GFP-F serves as a positive and negative control, respectively. In this experiment, SV40 was pretreated with the reducing agent dithiothreitol (DTT) and the calcium chelator EGTA to partially mimic conformational altered virus [36,40]. Importantly, when the virus was precipitated from the sample (Fig 5A, second panel), Hsp105 WT and mutant co-precipitated (Fig 5A, first panel, lanes 3–5); as expected, SGTA-F was pulled down but not GFP-F (Fig 5A, first panel, compare lane 1 to 2). Thus Hsp105 can interact with SV40 directly; because the Hsp105 mutants also bind to SV40 in vitro, they are unlikely globally misfolded.
To establish an interaction between Hsp105 and SV40 during cellular entry, cells expressing S tagged proteins were infected with SV40 in a synchronized manner for 12 h. The tagged proteins were affinity isolated and analyzed by immunoblot. While affinity isolation of either Hsp105 WT-S or HspBP1-S (but not GFP-S) pulled down Hsc70 (Fig 5B, second panel), only precipitation of Hsp105 WT-S pulled down SV40 (Fig 5B, first panel). This finding indicates that an Hsc70 complex containing Hsp105 engages the virus during infection. Because SV40 reaches the cytosol 6–8 hours post-infection (hpi) [37,39], we performed a time-course experiment to evaluate when Hsp105 associates with the virus. Cells expressing Hsp105 WT-S were infected in a synchronized manner for 2, 8, or 12 h. Hsp105 WT-S was affinity isolated and subjected to immunoblotting. We found that Hsp105 begins to engage the virus at 8 hpi, with the interaction increasing at 12 hpi (Fig 5C, first panel). The observation that SV40-Hsp105 binding is detected at approximately the same time point when the virus enters the cytosol suggests that Hsp105 engages the virus at the ER-cytosol interface.
As a second approach to test the idea that Hsp105 initiates its interaction with the virus at the ER membrane, we asked whether membrane-associated Hsp105 binds to SV40. To this end, Hsp105 WT-S was immunoprecipitated from the membrane (as well as the cytosol) fraction derived from infected cells. Interestingly, despite precipitating a lower level of Hsp105 WT-S from the membrane fraction when compared to the cytosol fraction (Fig 5D, second panel), no significant difference in the VP1 level from each fraction was co-precipitated (Fig 5D, first panel). This finding suggests that membrane-associated Hsp105 binds to SV40, consistent with the notion that this chaperone initiates its interaction with SV40 at the cytosolic surface of the ER membrane.
For a third approach to evaluate whether Hsp105 interacts with SV40 at the ER-cytosol interface, we tested if Hsp105 NE*-F and G*-F interact with SV40 and found that they cannot (Fig 5E, first panel, compare lanes 3 and 4 to 2). It is possible that these Hsp105 mutants are not targeted to the ER membrane because they cannot bind to transmembrane protein B14 via Hsc70 (Fig 5E, second and third panels, compare lanes 3 and 4 to 2). Alternatively, if B14 binds to Hsp105 directly, these Hsp105 mutants may fail to bind to B14 due to conformational changes resulting from the introduced mutations. Regardless, these results support the hypothesis that Hsp105 binds to SV40 at the ER-cytosol interface to extract the virus into the cytosol. Because Hsp105 NE*-F and G*-F interacts with SV40 in vitro (Fig 5A), their inability to bind to SV40 in cells (Fig 5E) is likely due to mis-localization away from the ER membrane.
How might Hsp105 engage and extract a large and intact viral particle embedded in the ER membrane [39] into the cytosol? One possibility is that Hsp105, in concert with the Hsc70 complex, disassembles the membrane-penetrating virus. This would destabilize the structural integrity of the membrane-embedded virus, enabling it to be released into the cytosol more efficiently. To test this, a sucrose gradient sedimentation assay modified slightly from previous studies was used [36,39,55]. In this assay, Triton X-100 extracted ER-localized SV40 (including membrane-penetrating virus) was incubated with ATP and the indicated purified proteins (Fig 2E). The samples were layered on a discontinuous sucrose gradient and centrifuged (Fig 5F). Individual fractions were collected and subjected to immunoblotting with anti-VP1 antibodies. In this fractionation procedure, any VP1 liberated from disassembled viral particles should appear in the top fractions corresponding to light sucrose density, whereas dense viral particles should appear in the bottom fractions with a denser sucrose concentration. Incubation with BSA retained viral VP1 signal at the dense bottom fraction (Fig 5F, first panel), confirming that ER-localized virus is intact as previously reported [39]. By contrast, addition of SDS caused a significant pool of VP1 to shift to the top fractions (Fig 5F, second panel), reflecting the ability of this detergent to chemically disassemble the virus [36]. Strikingly, whereas addition of Hsc70+B14, Hsp105 alone, or Hsp105+Hsc70 did not alter the fractionation pattern when compared to incubation with BSA (Fig 5F, third-fifth panels), inclusion of all three components (Hsp105 WT-F+Hsc70+F-B14) caused a significant portion of VP1 signal to distribute to the top fractions (Fig 5F, sixth panel). These findings indicate that Hsp105 must operate in concert with B14 and Hsc70 to stimulate disassembly of ER-localized virus. The distribution of VP1 signal to the top fraction represents VP1 pentamers liberated from disassembled viral particle, consistent with previous reports. The Hsp105-driven disassembly reaction is energy-dependent as addition of hexokinase and glucose (used to deplete ATP) to the Hsp105+Hsc70+B14 sample reduced the appearance of the disassembled virus in the top fractions (Fig 5F, seventh panel); a low level of virus nonetheless appeared in the middle fractions (fractions 4 and 5), possibly reflecting the use of non-hydrolyzed ATP that generated a partially uncoated SV40 intermediate. When the Hsp105 NE* mutant was incubated with the Hsc70-B14 complex, minimal virus disassembly was observed (Fig 5F, eighth panel), implicating the nucleotide exchange activity of Hsp105 as crucial in the disassembly reaction. Together, these findings suggest that Hsp105 binds to the membrane-penetrating virus at the ER-cytosol interface, and can potentially promote disassembly of the viral particle to facilitate extraction into the cytosol.
Our findings here establish an unanticipated role of an Hsp110 family member in driving membrane translocation of a viral particle. Specifically, we demonstrate that SV40 co-opts Hsp105 to cross the ER membrane and reach the cytosol in order to promote infection. SV40 infection begins when it traffics from the cell surface to the ER (Fig 6, step 1). In the ER, specific ER-resident isomerase and reductase act on the viral particle, imparting conformational changes to expose the hidden VP2/VP3 which generates a hydrophobic particle (Fig 6, step 2). The structurally altered virus then binds to and integrates into the ER membrane where SV40 accumulates into discrete foci (Fig 6, step 3). In our model, Hsp105, anchored to the membrane J-protein B14 directly or indirectly through Hsc70/Hsp70, binds to the membrane-penetrating virus (Fig 6, step 4a, see insert). Next, we hypothesize that iterative binding-release of SV40 by Hsp105-Hsc70 initiates the extraction process, a step that may involve disassembly of the membrane-embedded viral particle (step 4b). Extraction is completed when SV40 is fully released into the cytosol (step 4c). Upon cytosol arrival, a sub-viral particle intermediate is likely further processed to transport into the nucleus to cause infection (Fig 6, step 5).
The discovery that an Hsp110 family member is responsible for a membrane translocation event might appear surprising given that this chaperone family has generally been studied in the context of its chaperone activity during protein quality control in the cytosol [18,56,57]. However, as our biochemical analyses suggest that a pool of Hsp105 is localized to the ER via binding to the transmembrane J-protein B14, Hsp105 might control ER-associated protein quality control. The most well characterized ER protein quality control process is called ER-associated degradation (ERAD), a pathway that in fact involves translocation of a misfolded ER substrate across the ER membrane to the cytosol via an elaborate machinery where the substrate is degraded by the ubiquitin-dependent proteasomal system [58,59]. ER-to-cytosol translocation of SV40 is reminiscent of the fate of misfolded proteins in the ERAD pathway. Although a myriad of cellular factors sub-serving a defined function during ERAD have been identified [58], including Hsp105’s membrane-binding partner B14 and the related B12 [44,60–62], a role of Hsp105 itself in ERAD remains unclear [18,63].
Recent findings revealed that elements of the ERAD machinery are hijacked by toxic agents including viruses and bacterial toxins [36,39,64]. During host entry, these toxic agents are thought to disguise as misfolded proteins, co-opting components of the ERAD machinery in the ER to gain access to the cytosol. And by evading the cytosolic proteasome [65], these toxic agents are able to avoid a degradative fate to continue their cellular journey. SV40 serves as a salient example of a virus that hijacks the ERAD pathway during entry. While this virus indeed exploits many ERAD components to prepare it for ER-to-cytosol translocation [36,37,40,51,53], how SV40 is ultimately ejected into the cytosol from the ER membrane represents its most enigmatic entry step. This is especially salient given that the cytosolic p97 ATPase that normally extracts typical cellular misfolded ERAD substrates to the cytosol [66] is not used to extract the membrane-embedded SV40 into the cytosol, and that the Hrd1 E3 ubiquitin ligase linking p97 to the ER membrane also appears not to subserve any role during viral translocation [37]. Hence, the identity of the extraction machinery responsible for SV40 cytosol entry is unknown. Interestingly, cholera toxin also reaches the cytosol from the ER using a p97-independent process [67,68].
In this context, the loss- and gain-of-function data presented in this study strongly argue that Hsp105 plays a major role in extracting SV40 into the cytosol from the ER membrane. We found that this step requires the nucleotide exchange activity of Hsp105, suggesting that Hsp105 must act in concert with Hsc70 to drive the release of the viral particle into the cytosol. Our previous analyses demonstrated that SV40 penetrates the ER membrane as a large and intact viral particle [39]. A separate study suggested that SV40 undergoes modest uncoating in the ER without experiencing massive disassembly, morphing from its native 45 nm to 34 nm when it is in the ER [37]. How then might the Hsp105-Hsc70-SGTA-B14 complex generate the force necessary to eject this large membrane-embedded viral species into the cytosol? As stated, Hsp110 in conjunction with Hsc70 and a J-protein can disaggregate an aggregated model substrate [7,22–25]. Paralleling this proposed disaggregation activity, we found that purified Hsp105, Hsc70, and B14 drive disassembly of the ER-localized SV40. SV40 in fact disassembles in the cytosol during the infection course [39,69]. While a previous in vitro study demonstrated that eukaryotic Hsc70 and J-protein can uncoat native murine PyV when incubated overnight [55], whether addition of a mammalian NEF to this system can stimulate the rate of uncoating requires further investigation.
As our data demonstrate that the Hsc70-coupled Hsp105 machinery extracts SV40 into the cytosol, we speculate that the physical force generated by this machinery disassembles the membrane-penetrating virus, thereby destabilizing the viral particle to enable more efficient extraction into the cytosol. If this is the case, this disassembly may involve disruption of the VP1 C-terminal arm that normally stabilizes SV40 interpentamer interactions [30,31]. This disruption, coupled with reduction of any remaining interpentamer disulfide bonds in SV40 by the highly reducing cytosolic environment, should lead to formation of disassembled VP1 pentamers. Interestingly, the use of the Hsc70-associated machinery to disassemble the large and intact ER-localized viral particle is reminiscent of the disassembly of the approximate 150–200 nm clathrin basket by the Hsc70 machinery [70–72].
In infected cells, SV40 accumulates in foci within the ER postulated to represent the cytosol entry site [37,43,54]. Our imaging analyses performed under Hsp105 loss- and gain-of-function conditions further support this idea. Specifically, overexpressing Hsp105 impaired the formation of the virus-containing foci leading to increased cytosol extraction, while down-regulating Hsp105 enhanced foci formation thereby blocking cytosol arrival. The correlation between Hsp105’s capacity to promote loss of the SV40-containing foci and cytosol arrival of the virus strengthens the notion that the foci represent a portal for viral cytosol entry.
While chaperones are vital for maintaining proper protein folding, they are equally crucial in driving protein translocation across biological membranes [73]. This is arguably most evident in protein translocation across the ER membrane. For instance, during post-translational forward translocation where nascent polypeptides are translocated from the cytosol into the ER, the ER-resident Hsc70 chaperone BiP “pulls” substrates into the lumen via a Brownian ratcheting mechanism [74], whereas during ERAD when typical misfolded proteins are translocated from the ER back to the cytosol, the cytosolic p97 “pulls” the substrates into cytosol [66]. Our discovery that an Hsp110 family member can “pull” a large protein complex (a viral particle) from the ER into the cytosol suggests that this chaperone family member may also exert an important role during ERAD. Clearly future experiments will be required to expand on this exciting possibility.
CV-1, BSC-1, COS-7 and HEK 293T cells (ATCC) were grown in complete DMEM (cDMEM; containing 10% fetal bovine serum, 10 U/ml penicillin, and 10 μg/ml streptomycin; Gibco, Grand Island, NY). Opti-MEM and 0.25% trypsin-EDTA were purchased from Gibco. Sources of the other reagents are as follows: dithiobis(succinimidyl proprionate) (DSP; Thermo, Rockford, IL), digitonin and S tag protein-conjugated agarose beads (EMD Millipore, San Diego, CA), protein A conjugated agarose beads and protein G conjugated magnetic beads (Life Technologies, Carlsbad, CA), anti-FLAG M2 antibody-conjugated agarose beads, phenylmethanesulfonylfluoride (PMSF) and Triton X-100 (Sigma, St. Louis, MO).
SV40 was purified using the OptiPrep (60% stock solution of iodixanol in water; Sigma) gradient centrifugation method described previously in [39]. Briefly, viral genome transfected CV-1 cells were lysed in a buffer containing 50 mM Hepes pH 7.5, 150 mM NaCl and 0.5% Brij 58 for 30 min on ice, and the supernatant was collected after centrifugation at 20,000x g for 10 min. The supernatant was placed on top of a discontinuous OptiPrep gradient of 20% and 40%, and centrifuged at 49,500 rpm for 2 h at 4°C in an SW55Ti rotor (Beckman Coulter, Indianapolis, IN). A white interface formed between 20% and 40% OptiPrep was collected, and aliquots were stored at -80°C for future use. Purified BKV and antibody against BKV large TAg (pAB416) were generous gifts from Dr. Michael Imperiale (University of Michigan).
All the plasmids used in this study contain pcDNA3.1 (-) as vector backbone and the sources of these plasmids are: F-B14, F-B14 H136Q, SGTA-F and GFP-F [43]. Protein tags (S- or F-) at the N- or C-terminus are depicted as prefix or suffix, respectively. Hsp105 WT-S was generated from the plasmid pET28a-Hsp105 (a gift from Dr. Eileen Lafer, University of Texas). The siRNA-resistant Hsp105 WT was generated by introducing silent mutations (underlined 343-GAGCAGATAACAGCCATGTTGTTGA-367) using two rounds of PCR and the FLAG tag was introduced to obtained Hsp105 WT-F. The mutant Hsp105 N636Y/E639A (NE*) and G9L (G*) were created based on [24,51] by overlapping PCR using siRNA resistant Hsp105 WT-F. The gene products were then inserted into the vector backbone with FLAG tag. The HspBP1-S and Hsc70-S were amplified from a HEK 293T cDNA library and inserted into a vector with an S tag. The list of primers used to generate all the above plasmids is given in Table 2.
For overexpression studies, 50% confluent CV-1 cells in 6 cm, 10 cm or 15 cm plates were transfected with plasmid using the FuGENE HD (Promega, Madison, WI) transfection reagent at a ratio of 1:4 (plasmid to transfection reagent; w/v). Cells were allowed to express the protein for at least 24 h before experiments. For COS-7 cells in 6 cm plate, polyethylenimine (PEI; Polysciences, Warrington, PA) was used as the transfection reagent.
For the knockdown of Hsp105, SGTA and HspBP1, custom stealth siRNAs were generated and purchased from Invitrogen (Carlsbad, CA). Hsp105 siRNA #1 target sequence was adapted from [18], and HspBP1 siRNA target sequence was adapted from [75]. Hsc70 siRNA was purchased from Dharmacon (Lafayette, CO; catalogue: J-017609-08). The list of siRNAs used in this study is given in Table 3.
For microscopy experiments, 2x104 CV-1 cells were seeded on coverslips in 12-well plates and for all other experiments, 2x105 cells were seeded in 6 cm plates and were reverse transfected with 10 nM Hsp105 #1 or 12.5 nM Hsp105 #2 siRNA using Lipofectamine RNAiMAX reagent (Invitrogen) for at least 24 h. The ratio of siRNA to transfection reagent was maintained at 1:4 v/v. For knockdown followed by rescue experiments, 24 h post-siRNA transfection, cells were washed with cDMEM and transfected with siRNA resistant plasmids. Cells are allowed to express the protein for at least 24 h before experimentation.
Detection of XBP1 splicing was performed as described previously [76]. The primers used were listed in Table 4.
Transfected cells were harvested using trypsin and cell pellets were washed three times with cold phosphate buffered saline (PBS, Gibco). Washed cells were lysed in TSEp buffer (50 mM Tris-Cl pH 7.5, 150 mM NaCl, 1 mM EDTA and 1 mM PMSF) with 0.2% digitonin at 4°C for 10 min. For protein crosslinking studies, harvested cells were incubated with freshly prepared 2 mM DSP (dithiobis(succinimidyl proprionate)) for 30 min at room temperature with intermittent shaking. This membrane permeable, amine-reactive, and thiol-cleavable crosslinker was used to stabilize transient or weak protein-protein interactions. Excess cross-linker was quenched with 200 mM Tris pH 7.5. Cells were then lysed with 1% Triton X-100 in TSEp buffer at 4°C for 10 min. Cell lysate were clarified by centrifugation at 20,000x g for 10 min at 4°C. The resulting supernatant was immunoprecipitated with anti-FLAG conjugated agarose beads or affinity purified with S tag protein-conjugated agarose beads for 2 h at 4°C. Samples were eluted with 1x SDS sample buffer with 1.25% β-mercaptoethanol (Sigma) and boiled for 5 min at 95°C before subjected to SDS-PAGE and immunoblotting.
Flp-In 293 T-Rex cells (Invitrogen) transfected pcDNA5-B14-3xF was used to immunopurify B14-3xF as in [43]. Briefly, cells selected in media containing blasticidin and hygromycin (Invitrogen) were induced overnight with freshly prepared 5 ng/ml tetracycline (Sigma) to express B14-3xF to near endogenous level. 10 μM ganglioside GM1 (Matreya, Pleasant Gap, PA) was supplemented to the culture media prior to infection. Next day, near confluent cells were infected with SV40 (MOI ~50) for 16 h. HEK 293T cells were used as negative control. Post infection, cells were harvested with cold PBS and centrifuged at 500x g for 5 min. Cells pellets were lysed in 2.5 ml buffer containing 0.1% digitonin in TSEp buffer for 30 min in ice. Lysate was centrifuged at 20,000x g for 15 min and the supernatant was incubated with anti-FLAG agarose conjugated beads for 2 h at 4°C. Beads were washed three times with TSEp buffer, and the proteins were eluted twice using 3x FLAG peptide (200 μl, 0.25 mg/ml in PBS) (Sigma) for 1 h at 4°C. Eluents were concentrated using centrifugal filters (Amicon Ultra 3K, Cork, Ireland), and the concentrated samples were separated on SDS-PAGE and either visualized by silver staining (Invitrogen) or immunoblotted. For mass spectrometry analysis, protein bands were excised from the silver stained gel, and analyzed at Taplin Biological Mass Spectrometry Facility (Harvard Medical School). The data obtained were processed based on number of unique peptides identified, percentage sequence coverage, and the observed molecular weight.
Cells were prechilled at 4°C for 20 min before infecting with SV40 (MOI ~5) for 1 h at 4°C. Cells were washed once with cold cDMEM, warm cDMEM was added, and cells were incubated for 12 h at 37°C. Post infection, cells were lysed in HNp buffer (50 mM Hepes pH 7.5, 150 mM NaCl and 1 mM PMSF) containing 0.1% digitonin at 4°C for 10 min, and separated into supernatant (cytosol) and pellet (membrane) fractions by centrifugation at 20,000x g for 10 min at 4°C. To isolate ER-localized SV40, the pellet fraction was further treated with HNp buffer containing 1% Triton X-100 for 10 min at 4°C and centrifuged at 20,000x g for 10 min at 4°C. The fractions were then dissolved in 1x SDS sample buffer containing 1.25% β-mercaptoethanol, and boiled for 5 min at 95°C before immunoblotting. To assess cytosol arrival of cholera toxin A1 (CTA1) subunit, CV-1 cells were treated with 10 nM CT (EMD Millipore) for 90 min. Cells were harvested and fractionated as above.
Purified Hsc70 was purchased from StressMarq Biosciences (Victoria, Canada). For purification of F-B14, SGTA-F, GFP-F, Hsp105 WT-F, Hsp105 NE*-F and Hsp105 G*-F proteins, confluent HEK 293T cells in 15 cm plates were transfected with pcDNA3.1 (-)-FLAG tag plasmids using 1:4 ratio of DNA to PEI transfection reagent (w/w). After transfection for 24–48 h, cells were washed three times with PBS and harvested using trypsin. Cell pellets were lysed in HNp buffer containing 1% Triton X-100 for 20 min at 4°C. Cell lysates were cleared by centrifugation at 20,000x g for 10 min at 4°C and the supernatant was incubated with 25 μl of anti-FLAG M2 agarose conjugated beads for 2 h at 4°C. Beads were washed three times with HNp buffer and incubated with HKM buffer (20 mM Hepes pH 7.5, 50 mM KCl, 2 mM MgCl2) containing 0.1% Triton X-100 and 2 mM ATP for 30 min at room temperature. Experiments involving F-B14 contains 0.1% Triton X-100 throughout. Bound proteins were eluted twice using FLAG peptide (100 μl, 0.25 mg/ml; Sigma) for 30 min at 4°C. Eluents were concentrated using centrifugal filters (Amicon Ultra 3K) and the concentrated samples were separated on SDS-PAGE and visualized using Brilliant Blue R250 (Thermo Fisher) staining.
OptiPrep purified SV40 was treated with 3 mM DTT and 10 mM EGTA for 45 min at 37°C. The reaction was passed through Micro biospin P-30 Tri-chromatography columns (Biorad, Hercules, CA) and SV40 was eluted using PBS. Approximately 250 ng of SV40 was incubated with 0.2–0.5 μM of purified FLAG tag proteins. The reaction was made up to 50 μl with PBS and incubated for 1 h at 25°C, followed by overnight incubation with anti-VP1 at 4°C using an end-over-end rotor. The reaction was then incubated with 10 μl of protein G conjugated magnetic beads for 2 h at 4°C. Unbound fraction was collected and the beads were washed three times with PBS. Bound proteins were eluted with 1x SDS sample buffer with 1.25% β-mercaptoethanol, and boiled for 5 min at 95°C before immunoblotting.
Purified FLAG tag proteins (0.2–0.5 μM) were initially incubated with HKM buffer for 30 min at room temperature and then incubated with 10 μl of ATP-agarose beads (Innova Biosciences, Cambridge, UK) for 30 min at room temperature with gentle agitation. Unbound fraction was collected and the beads were washed three times with HKM buffer. Bound proteins were eluted with 1x SDS sample buffer with 1.25% β-mercaptoethanol and boiled for 5 min at 95°C before immunoblotting.
An assay, based on [77] and [78], was developed to test the nucleotide exchange activity of WT and mutant Hsp105 against Hsc70. Hsc70 (3 μM) was incubated with 50 μCi of [α-32P] ATP (3000 Ci/mmol; Perkin Elmer, Waltham, MA) in a final volume of 25 μl (50 μCi is equivalent to 0.66 μM ATP in this reaction) at 37°C for 30 min to form [α-32P] ADP Hsc70, and the sample subjected to a spin gel filtration column (GE Healthcare, Cleveland, OH) to remove the free nucleotides. [α-32P] ADP Hsc70 was incubated with 0.3 μM of FLAG tag proteins (GFP-F, Hsp105 WT-F, Hsp105 NE*-F or Hsp105 G*-F) in 23 μl of a buffer containing 20 mM Hepes (pH 7.5), 50 mM KCl and 0.1% Triton X-100 at 23°C for 20 min. Following incubation, each reaction was mixed with unlabeled 0.3 μM ATP and 2 mM MgCl2 and further incubated at 23°C for 1 min. After removal of free nucleotides using another spin gel filtration column, 2 μl of the reaction was spotted onto a PEI cellulose TLC plate (Sigma). The plates were developed in 0.6 M KH2PO4 pH 3.4 as the solvent system. Once dried, the plates are developed by exposing on a photographic film.
ER-localized SV40 was incubated in HKM buffer containing 2 mM ATP and 0.1% Triton X-100 with chaperones at the following concentrations: Hsc70 (~2 μM), F-B14 (1 μM), Hsp105 WT-F, or Hsp105 NE*-F (0.4 μM). BSA (1 μM) with or without 1% SDS was used as a control. Total volume of 20 μl was maintained throughout the reaction. Reaction was incubated at 37°C for 1 h and the sample placed on top of a discontinuous sucrose gradient consisting of 20%, 30% and 40% sucrose. Samples were centrifuged at 50,000 rpm in a TLA100 rotor (Beckman Coulter) for 30 min at 4°C, and individual fractions collected from the top of the gradient. Samples were then subjected to immunoblotting using antibodies against VP1.
CV-1 cells were grown and transfected on sterile cover slips. Cells were infected with MOI ~0.5 (for TAg expression studies) or MOI 20–30 (for foci formation studies) for 24 h and 16 h, respectively. Infected cells were fixed with 1% formaldehyde for 15 min at room temperature followed by permeabilization with 0.2% Triton X-100 in PBS for 5 min. Cells are then covered for 15 min with blocking buffer containing 5% milk in TBST (Tris buffered saline with 0.02% Tween 20). Cells were then immunostained with primary antibody diluted in blocking buffer for 1 h at room temperature and then washed five times with blocking buffer. Cells were then incubated with fluorescence dye conjugated secondary antibody for 30 min and then washed three times with blocking buffer, PBS, and water before air drying and mounting on glass slides (Fisher) using ProLong gold (Invitrogen) with or without DAPI (Molecular Probes, Eugene, OR). Slides were then allowed to dry in dark at room temperature for at least 12 h before imaging. Images were taken using an inverted epifluorescence microscope (Nikon Eclipse TE2000-E, Melville, NY) equipped with 40x, 60x and 100x 1.40 NA objectives and standard DAPI (blue), FITC (green) and TRITC (red) filter cubes. Images were processed using the ImageJ software version 1.48i (NIH). Cells were counted either under a microscope with an eyepiece or with the help of ImageJ program (Plugin: Cell counter). In each experiment, >1000 cells were scored for TAg positive cells or >100 were scored for BAP31 foci positive cells, in the indicated channel. Foci intensity and area profile are plotted with the assistance of ImageJ program (Plugin: Surface plot).
Data obtained from at least three independent experiments were combined together for statistical analyses. Results were analyzed using Student’s t test. Data are plotted using GraphPad Prism software, version 5.0b. Data are represented as the mean values and error bar represents standard deviation (SD) (n ≥3) where indicated. * p < 0.05, ** p < 0.01, *** p < 0.001 were considered to be significant unless otherwise noted.
|
10.1371/journal.pgen.1006427 | Human Oocyte-Derived Methylation Differences Persist in the Placenta Revealing Widespread Transient Imprinting | Thousands of regions in gametes have opposing methylation profiles that are largely resolved during the post-fertilization epigenetic reprogramming. However some specific sequences associated with imprinted loci survive this demethylation process. Here we present the data describing the fate of germline-derived methylation in humans. With the exception of a few known paternally methylated germline differentially methylated regions (DMRs) associated with known imprinted domains, we demonstrate that sperm-derived methylation is reprogrammed by the blastocyst stage of development. In contrast a large number of oocyte-derived methylation differences survive to the blastocyst stage and uniquely persist as transiently methylated DMRs only in the placenta. Furthermore, we demonstrate that this phenomenon is exclusive to primates, since no placenta-specific maternal methylation was observed in mouse. Utilizing single cell RNA-seq datasets from human preimplantation embryos we show that following embryonic genome activation the maternally methylated transient DMRs can orchestrate imprinted expression. However despite showing widespread imprinted expression of genes in placenta, allele-specific transcriptional profiling revealed that not all placenta-specific DMRs coordinate imprinted expression and that this maternal methylation may be absent in a minority of samples, suggestive of polymorphic imprinted methylation.
| Differences in gamete DNA methylation is subject to genome-wide reprogramming during preimplantation development to establish an embryo with an epigenetic state compatible with totipotency. DNA sequences associated with imprinted differentially methylated regions (DMRs) are largely protected from this process, retaining their parent-of-origin epigenetic marks. By comparing the methylation profiles of human oocytes, sperm, blastocysts and various somatic tissues including placenta, we observe hundreds of CpG island sequences that maintain methylation on their maternal allele in blastocysts and placenta indicative of incomplete reprogramming. In some cases this maternal methylation influence transcription of nearby genes, revealing transient imprinting in embryos after genome-activation and in placenta. Strikingly, these placenta-specific DMRs are polymorphic between placenta samples with a minority of samples being robustly unmethylated on both alleles.
| In mammals, DNA methylation of CpG dinucleotides has been shown to play critical roles in many developmental processes including cellular differentiation, X chromosome inactivation and genomic imprinting. DNA methylation patterns are initially established by the de novo DNA methyltransferase DNMT3A [1], with the methylation profile faithfully maintained during DNA replication by the maintenance methyltransferase DNMT1-UHRF1 complex [2].
It has recently been shown that the gametes from both mouse and humans possess large intervals of opposing methylation [3–7]. Within a few hours after fertilization, a wave of global epigenetic reprogramming ensures that methylation at the blastocyst stage is at their lowest level, erasing the majority of this gametic epigenetic information [3, 5, 7]. However, some specific sequences survive this demethylation process, specifically those located within imprinted regions and certain repeat subtypes. Imprinted genes are only transcribed from one parental allele leading to parent-of-origin specific expression, with allelic expression directly controlled by allelic methylation [8]. To date all imprinted domains contain at least one differentially methylated region (DMR) that acquires methylation during gametogenesis (germline DMR, or gDMR), and maintained throughout development. Some imprinted loci also contain DMRs that become allelically methylated in the embryonic diploid genome (somatic DMRs, or sDMR) which are under the hierarchical influence of gDMRs [4, 9, 10]. Recently, transiently methylated germline DMRs (tDMRs) have been identified in mice that are indistinguishable from ubiquitous imprinted gDMRs in gametes and preimplantation embryos [11]. The maternally methylated tDMRs described in mouse subsequently gain methylation on their paternal alleles at implantation, having first survived the post-fertilization demethylation process. This reprogramming to a totipotent state starts in the male pronucleus with TET3-mediated conversion of 5-methylcytosine (5mC) to 5-hydroxymethylcytosine (5hmC) [12] with subsequent replication-dependent dilution of methylation of both maternal and paternal genomes occurring during the first 2 days of human development [5, 7, 13]. Unlike mice, in humans it is currently unknown how many germline differences survive embryonic reprogramming and persist in humans, either as ubiquitous imprinted gDMRs or tDMRs. However initial screens suggest that oocyte-derived tDMRs may be present to the blastocyst stage [7, 14].
Here, we present the data describing the fate of germline-derived methylation in humans. Using publically available methyl-seq datasets from gametes, preimplantation embryos, placenta and somatic tissues, we identify 53,549 methylation differences between gametes, the majority being methylated in the sperm and not in oocytes. With the exception of a few paternally methylated gDMRs associated within known imprinted domains, we demonstrate that sperm-derived methylation is reprogrammed by the blastocyst stage. In contrast a large number of oocyte-derived methylation differences survive to the blastocyst stage, persisting as maternally methylated DMRs in the placenta only, expanding the number of placenta-specific DMRs reported using high-density array based screens [10, 15–17]. Furthermore, we demonstrate that this phenomenon is exclusive to humans and non-human primates since no placenta-specific maternal methylation was observed in other mammalian species. Utilizing single cell RNA-seq datasets from human preimplantation embryos [18] we show that following embryonic genome activation the maternally methylated gDMRs orchestrate imprinted expression in preimplantation embryos. However, despite showing imprinted expression of many genes, transcriptional profiling revealed that not all placenta-specific maternally methylated DMRs coordinate imprinted expression suggesting differential reading of this epigenetic mark during human embryonic development.
Transient maternally inherited monoallelic methylation has been previously observed in mouse. To identify candidate loci in humans we searched for regions that are differentially methylated between sperm and oocytes. Using defined criteria (see methods) we identified 5, 438 oocytes and 48, 111 sperm-derived DMRs.
A high proportion of regions methylated in sperm and hypomethylated in oocytes were intergenic or map to repeat elements, consistent with previous observations [5]. In contrast, oocyte-specific DMRs were more uniformly distributed throughout the genome, often overlapping promoter CpG islands. Eighty percent of the oocyte-derived DMRs (n = 4, 352) remain partially methylated at the blastocyst stage, which is consistent with methylation dynamics during the progression of cleavage stage embryos to blastocysts [6, 7] with very few sperm-derived DMRs surviving to the blastocyst stage (1%, n = 517) (Fig 1A and 1B). This reprogramming is particularly evident when the size of the gDMRs surviving to the blastocyst stage is taken into consideration. In total ~7 Mb of the human genome encompasses oocyte-derived gDMRs of which 74% is hemimethylated in preimplantation embyros, whereas ~2.7 Mb is covered by sperm-derived gDMRs of which only 11% is hemimethylated at the same developmental stage. Therefore, maternal gDMRs are lost after the blastocyst stage whereas the methylation at paternal gDMRs is largely removed during preimplantation stages, possible occurring before the first cleavage division arguing against a simple replication-dependent demethylation of the maternal genome during preimplantation development.
Numerous studies have shown that gDMR that persist uniformly in somatic tissues act as imprinting control regions. To date 49 ubiquitous imprinted DMRs have been identified in humans using high-density methylation arrays [10]. To determine if additional imprinted DMRs are present in the human genome, we determined the methylation profile of the oocyte and sperm-derived gDMRs that are present with preserved methylation in methyl-seq datasets in blastocysts, placenta and 14 different somatic tissues. We observe only one sperm-derived region mapping to a known paternally methylated DMR in > 12 tissues, the H19 gDMR on chromosome 11. The only additional known paternally methylated DMR originating from sperm in humans, the IG-DMR on chromosome 14, was differently methylated between gametes but was partially methylated in blastocysts and five somatic tissues only. Using the same criteria we observe 60 oocyte-derived DMRs in >12 tissues, including 25 known maternally methylated imprinted DMRs (S1 Table). Of the remaining intervals not associated with known imprinted gDMRs, we confirm FANCC and SVOPL as being novel ubiquitous imprints (S1 Fig; Fig 1C and 1D). Using allele-specific RT–PCR that incorporated a coding SNP within exon 5, we observed maternal expression of SVOPL in placenta and monoallelic expression in brain and leukocytes (Fig 1E). Unfortunately we could not identify any informative samples to allow for the allelic expression of FANCC to be ascertained.
To determine if germline-derived DMRs are maintained in a tissue-specific fashion we screened for loci partially methylated in only one tissue (Fig 2A and 2B). This analysis revealed that 551 of the partially methylated regions in blastocysts inheriting methylation from the oocyte survived only in the placenta, whereas only 38 regions inheriting methylation from sperm were identified in this extra-embryonic tissue (S2 Table). Since standard bisulphite conversion based technologies cannot distinguish between 5mC and 5hmC, we utilized methylation-sensitive genotyping assays that can distinguish these two forms based on the addition of a glucose moiety to yield glucosyl-5-hydroxymethylcytosine. This, combined with allele-specific bisulphite PCR, revealed no novel paternally methylated placenta-specific gDMRs since all candidates were mosaically methylated, but maternal placenta-specific gDMRs were abundant and specifically associated with 5mC (S2, S3 and S4 Figs; S2 and S3 Tables).
The fate of 5mC at maternal placenta-specific gDMRs in somatic tissues was largely influenced by sequence content. The confirmed placenta-specific maternal methylation regions were almost always high CG content intervals robustly unmethylated in somatic tissues, whereas hypermethylated loci in somatic tissues were often false positives being partially methylated regardless of the parental allele, with the exceptions of four loci which we confirm as maternally methylated (TMEM247, GPR1-AS1, ZFAT, and C19MC) (S5 and S6 Figs) [19, 20]. This reflects the general methylation status of the placenta, which is relatively hypomethylated across the genome, including repeat elements [21, 22]. For example the GRID2 gene is associated with two maternally methylated gDMRs with different genomic content. The promoter CpG island is robustly methylated on the maternal allele in placenta and is unmethylated in somatic tissues, whereas an intergenic region within intron 3, consisting of an Alu/SINE repeat, is a gDMR with a mosaic methylation profile in placenta that is fully methylated in all somatic tissues (Fig 2C and 2D).
We observe robust maternal methylation associated with multiple members of two large gene families, the fibroblast growth factors (FGF8, FGF12 and FGF14) and calcium channel, voltage-dependent channel subunits (CACNA1A, CACNA1C, CACNA1E and CACNA1I) as well as several gene involved in epigenetic regulation (JMJD1C and DNMT1) and microRNA processing (LIN28B and EIF2C1) (S3 and S4 Figs). These results therefore reveal that placenta-specific gDMRs are much more abundant in the human genome than previously reported.
Using nested-multiplex bisulphite PCR, we confirmed the methylation profiles of four ubiquitous imprinted gDMRs (H19, MCTS2, FANCC, SVOPL) and 13 placenta-specific gDMRs in sperm and blastocysts micosurgically separated into inner cell mass (ICM) and trophectoderm (TE) (Fig 3 and S4 Fig). The R3HCC1 loci on chromosome 8 exemplified the fate of opposing germline methylation difference as this gene has adjacent oocyte and sperm-derived gDMRs. Using bisulphite PCR we show that the maternally methylated gDMR is observed in ICM/TE and term placenta, whereas the paternally methylated gDMR, which was not identified in our initial genome-wide screen since it does not reach our screening criteria having < 25 CpGs, resolves to a mosaic methylated state at the blastocysts stage (Fig 3C and 3D).
We have previously shown that 14 orthologs of maternally methylated placenta-specific DMRs are devoid of methylation in the mouse placenta [10]. Using methyl-seq datasets from mouse placenta with bisulphite PCR confirmation we show that no human placenta-specific DMRs are conserved in mice (S7A and S7B Fig). Similarly the mouse orthologous regions corresponding to the SVOPL and FANCC DMRs also lack allelic methylation and are biallelically expressed in multiple tissues (S7B Fig). Several studies have shown that maternally methylated gDMRs mark different loci in mouse compared to humans [3, 5], suggesting that the mouse genome may possess a unique set of placenta-specific DMRs inherited from the female germline. We therefore determined the fate of oocyte-derived gDMRs in hybrid mouse placenta. Consistent with our previous observation, no maternal gDMRs persist as placenta-specific DMRs, reinforcing that this phenomenon is not observed in mice (S7C Fig). Recently methyl-seq datasets have been produced from different mammalian species, including rhesus macaque, horse, cow and dog [23]. Similar to mouse, the orthologues of the vast majority of human placenta-specific gDMRs do not have a methylation profile consistent with imprinting in non-primate species (S7A Fig). Using bisulphite PCR on DNA derived from rhesus placenta, we confirm evolutionary conservation of 63% placenta-specific DMRs as well as those associated with the ubiquitously imprinted MCTS2, GRB10 and L3MBTL1 genes (S7D Fig).
The main biological significance of promoter methylation is thought to be transcriptional repression of tissue-specific genes, with methylation levels negatively correlated with expression following genome activation at the 8 cell stage [6]. To determine if maternal-specific methylation at placenta-specific gDMRs dictates paternal expression we performed allele-specific RT-PCR in placenta. Paternal expression was confirmed for nine genes including AGO1, USP4, SH3BP2, FAM149A, MOCS1, R3HCC1, JMJD1C, PAK1 and PAPLN-AS (Fig 4A; S4 Table). Curiously however, we observe that not all informative placenta samples exhibited monoallelic expression despite maintaining robust maternal methylation (Fig 5A, S4 and S8 Figs; S4 Table). We also observe paternal expression of a ~10 kb non-coding (nc)RNA overlapping a placenta-specific gDMR located 12 kb 3’ to TET3 (Fig 4B–4D). To determine if this ncRNA influences expression in cis, we performed allelic RT-PCR for TET3. We observe biallelic expression of TET3 suggesting that the neighboring ncRNA does not possess enhancer or repressive function in term placenta (Fig 4D). In total this bringing the total number of confirmed placenta-specific paternally expressed genes to more than 30 [8, 16].
Polymorphic imprinting has been described for only a few loci in humans, including the IGF2R [24, 25] and nc886/vtRNA2-1 [26, 27], with the latter consistent with being a metastable epiallele. To determine if the placenta-specific gDMRs that we identified show variable methylation on the maternal allele, we performed pyrosequencing to quantify a larger cohort of normal placenta samples from uncomplicated pregnancies. We identified hypomethylated samples for 12 of the regions (Fig 5B), with the most affected loci being LIN28B and AGBL3. For samples with informative polymorphisms this lack of methylation is associated with biallelic expression (Fig 5C), an observation consistent with some placenta-specific maternal gDMRs being a stochastic polymorphic trait [17].
It has previously been reported that a significant proportion of transcripts are monoallelically expressed in cleavage embryo [17] indicating that maternally methylated placenta-specific gDMRs may regulate allelic expression at this earlier developmental time point. To ascertain if the placenta-specific gDMRs orchestrate imprinted expression, we determined allelic expression in publically available single cell embryo RNA-seq datasets for which paternal genotypes were available [18]. Gene expression profiles were analyzed in individual embryos to determine the progression of expression levels and their allelic origin. To compare embryos at different stages it is important to take into consideration two events, embryonic genome activation and oocyte-derived transcript degradation. Zygotic genome activation (ZGA) occurs soon after fertilization (pre-major ZGA) and processed in successive waves of activation with the major changes reported at the 4–8 cell stage [28]. Maternal transcript stores in the oocyte cytoplasm are diminished after fertilization by a combination of degradation and recruitment to the polysome and translated prior to ZGA [29]. Transcripts highly abundant at the pronuclear stage and decreasing as developmental proceeds will not be expressed from the embryonic genome and will appear maternally derived. Embryonically transcribed genes that maintain high expression levels from the pronuclear stages would appear maternally expressed before 8-cell stage, switching to imprinted paternal expression with RNA synthesis from the unmethylated allele if the gDMRs are functional. Some instances of biallelic expression maybe wrongly classified since embryonic paternal expression and oocyte-derived transcripts may co-exist until late cleavage stage. Finally, genes that are activated during cleavage embryo development, but not originally expressed in the zygote are predicted to be from the paternal allele. Therefore functional paternal expression can only be categorized after genome activation (Fig 6A). Using these criteria we screened all transcripts near the oocyte-derived gDMRs for imprinting and observed, as proof of principal, the paternal expression of ZHX3 in 8-cell and morula and confirm preferential paternal expression arising from a maternally methylated promoter in multiple term placenta biopsies (Fig 6B–6E).
In addition to the reprogramming that occurs immediately after fertilizations from which imprints are protected, reprogramming in primordial germ cells (PGCs) of the developing fetus includes all ubiquitous imprints ensuring the transmission of genetic information with the correct epigenetic profile in the gametes [30]. Recently, the methylomes of human PGCs of both sexes have been generated, which confirm that human PGCs at 7–9 weeks gestation are hypomethylated similar to those in the mouse at embryonic day 13.5 [31, 32]. Using these datasets, we confirm that placenta-specific DMRs are devoid of methylation in both male and female PGCs at 10 weeks gestation and are indistinguishable from ubiquitous gDMR imprints (S5 Table). Similar to the ubiquitous gDMR imprints, the majority of placenta-specific gDMRs (78%) are frequently associated with CpG-rich sequences with an intragenic location with evidence of a transcriptional event initiating from upstream promoters (S5 Table). This intragenic location has been shown to be important in facilitating the acquisition of methylation during female germline development [33, 34].
In this study DNA methylation in human gametes, embryos, placenta and multiple somatic tissues were used to identify gDMRs that may act as imprints. Using high-density methylation arrays, our group and others have recently identified ~150 maternally methylated DMRs in placenta [10, 15–17, 35], for which we confirm the majority are bona fide germline difference in methylation. A comparison of the oocyte-derived DMRs reported by Smith and colleagues revealed largely overlapping datasets in blastocysts [7]. Using different bioinformatics criteria, 25 continuously CpGs rather than 100bp tiles, our analysis identified ~64% of previously identified loci, with missing regions possible due to inferior sequence coverage of reduced representation bisulphite sequencing or the size of the windows analyzed. Furthermore, using methyl-seq datasets, we identify an additional 551 loci that could represent placenta-specific gDMRs, however only 11% had high informative polymorphisms to allow for allelic discrimination. With the exception of only four regions, these placenta-specific gDMRs are associated with CpG islands or promoter intervals devoid of methylation in somatic tissues. Those regions fulfilling our criteria of partially methylation and hypermethylated in other tissues may simply reflect the relatively hypomethylated nature of the placenta genome that had previously hindered us from performing imprinted DMR analyses in placenta methyl-seq datasets [10]. Recently, Schroeder and colleagues described that the placenta genome has unique partially methylated domains (PMDs) that are larger (>100 kb) and have lower levels of DNA methylation than the rest of the genome, which are stable throughout gestation [21, 36]. The placenta-specific gDMRs we describe are much smaller than PMDs having an average size of 2.2 kb with only two (CACNA1I and ZNF385D) mapping to PMDs.
While allelic DNA methylation at ubiquitous gDMR imprints is associated with monoallelic expression, our analysis reveals that only half of all placenta-specific gDMRs orchestrate paternal expression suggesting that despite being maternally methylated, the maternal alleles may not be associated with a compact chromatin state or decorated with repressive histone modifications sufficient to influence transcription. A recent genome-wide screen using diandric and digynic triploid conceptions and RRBS datasets also identified placenta-specific gDMRs, many overlapping with the loci we identify [17]. However, these authors did not perform any allelic expression analyses for their candidates and so the functional relevance of this tissue-specific methylation was not addressed. Furthermore this study revealed epigenetic stochasticity for many of the placenta-specific DMRs described, similar to what we also observe for many of the regions we quantified using pyrosequencing (Fig 5). However it remains to be determined whether lack of methylation at these loci reflects a random selection of cells not maintaining methylation after embryonic reprogramming or alternatively, exposes loci that fail to establish methylation in the female germline in a polymorphic fashion.
We show that allelic methylation is present in the inner cell mass and trophectoderm of human blastocysts, revealing that 5mC is selectively protected from embryonic reprogramming and that it maintained following the first differentiation step. Furthermore, our data suggest that an additional small wave of targeted demethylation exists following implantation in cells specified for the somatic lineages that is absent during placenta differentiation. Very few studies have assessed allelic expression of imprinted genes in human embryos with only paternal expression of IGF2, SNRPN and MEST being previously reported [37–39]. We show that the placenta-specific gDMRs can influence allelic expression immediately following embryonic genome activation as highlighted by ZHX3. Unfortunately no additional paternally expressed genes were identified in the embryo datasets due to the lack of informative polymorphisms. Extrapolating this observations means that there are potentially thousands more transiently imprinted genes in the blastocysts associated with the loci which get that reprogrammed after implantation which may have a physiological role in embryonic development.
By directly assessing methylation in placenta-derived DNA from different mammalian species we observe that oocyte-derived gDMRs in placenta are largely restricted to primates, being most abundant in humans. These observations are inconsistent with recent reports that oocyte-derived methylation regulates trophoblast development in the mouse [40]. However, this study did not assess allelic methylation per se, but inferred it from various Dnmt3a/Dnmt3b knockout crosses. The developmental phenotype observed could be due to the deregulation of only a few genes such as the maternally expressed Ascl2 (previously known as Mash2) that is regulated in cis by the maternally methylated ubiquitous KvDMR1[41, 42]. Furthermore strand-specific bisulphite PCR of several of the proposed genes responsible for this developmental phenotype failed to identify methylation specifically on the maternal allele in mouse hybrid placenta (S9 Fig).
There are no unifying explanations of how imprinted genes evolved, but there are several theories hypothesized that underscore the importance of the placenta. The most popular theory is associated with the parental conflict and nutrient supply and demand hypothesis [43, 44]. However with the recent identification of developmentally important genes, including the FGFs that regulate trophoblast survival and placental angiogenesis [45], and key epigenetic regulators, such as JMJD1C which is involved in regulating early preimplantation development of bovine embryos [46], we favor the hypothesis that maternal silencing is a mechanism to prevent ovarian teratomas that arise from parthenogenetically activated oocytes [47, 48].
Our study has shown that oocyte-derived methylation can uniquely be maintained as DMRs in the extra-embryonic lineages, with many placenta-specific DMRs coordinating paternal expression following embryonic genome activation. Our data corroborates the observations that these placenta-specific gDMRs can be polymorphic, with a minority of samples being unmethylated [17]. It remains to be seen if the lack of these placenta-specific DMRs influences pregnancy outcomes and whether they are involved in implantation and preimplantation embryo viability.
Ethical approval for the use of human placenta samples was granted by the Institutional Review Boards at the National Center for Child Health and Development (project 234), Hospital St Joan De Deu Ethics Committee (35/07) and Bellvitge Institute for Biomedical Research (PR006/08). The use of surplus human embryos for this study was evaluated and approved by the scientific and ethic committee of the Instituto Valenciano de Infertilidad (IVI) (1310-FIVI-131-CS), Bellvitge Institute for Biomedical Research Ethics Committee (PR292/14), the National Committtee for Human Reproduction (CNRHA) and the Regional Health Counsel of Valencia.
Mouse work was approved by the Institutional Review Board Committees at the National Center for Child Health and Development (approval number A2010-002).
A single placenta sample from rhesus macaque was obtained from the breeding colony of the Biomedical Primate Research Center, Rijswijk, Netherlands using protocols approved by the Committee on the Ethics of Animal Tissue Collection at BPRC (Permit # 730). The EUPRIM-Net Bio-Bank is conducted and supervised by the scientific government board along all lines of EU regulations and in harmonization with Directive 2010/63/EU on the Protection of Animals Used for Scientific Purposes.
A cohort of 72 human term placenta biopsies (gestational age 35–41 weeks gestation, average 37 weeks) from uncomplicated pregnancies with their corresponding maternal blood samples were collected at Hospital St Joan De Deu (Barcelona, Spain) and the National Center for Child Health and Development (Tokyo, Japan). Written informed consent was obtained from all participants. All placenta biopsies were collected from the fetal side around the cord insertion site. The placenta-derived DNA samples were free of maternal DNA contamination based on microsatellite repeat analysis. Both DNA and RNA extractions and cDNA synthesis were carried out as previously described [22].
Three surplus human blastocysts were recruited at the Fundación Instituto Valenciano de Infertilidad (FIVI) in Valencia. The blastocysts were thawed using the Cryotop method following manufacturer’s instructions [49] and incubated in CCM medium (Vitrolife, Göteborg, Sweden) for 6–12 hours before microdissection in order to allow their full expansion and the inner cell mass (ICM) and trophectoderm (TE), that were subsequently separated by micromanipulation using laser technology (OCTAX, Herborn, Germany). The separated ICMs and TEs were individually placed in PCR tubes containing 2.5 μL of PBS and immediately snap frozen at -80°C until processing.
Wild type mouse embryos and placentas were produced by crossing C57BL/6 with Mus musculus molosinus or Mus musculus castaneous mice. Animal husbandry and breeding were conducted according to the institutional guidelines for the care and the use of laboratory animals.
A single placenta sample from rhesus macaque (animal 95023) was obtained from the breeding colony of the Biomedical Primate Research Center, Rijswijk, following a C-section procedure.
We analysed twenty-eight publicly available methylomes obtain from GEO or NBDC repositories. Two datasets were derived from human oocytes (JGAS00000000006), 5 from human sperm (JGAS00000000006 and GSE30340), 3 from brain (GSM913595, GSM916050, GSM1134680) 3 from CD4+ lymphocytes (GSE31263), 2 from liver (GSM916049, GSM1134681) and individual datasets from preimplantation embryos (JGAS00000000006), placenta (GSM1134682), muscle (GSM1010986), CD34+ cells (GSM916052), sigmoid colon (GSM983645), lung (GSM983647), aorta (GSM983648), esophagus (GSM983649), small intestine (GSM983646), pancreas (GSM983651), spleen (GSM983652), adrenal (GSM1120325) and adipose tissue (GSM1010983). Methylation calls were mapped to the hg19 genome. CpG methylation values were calculated using reads from both strands as (methylated / (methylated + unmethylmated). Only CpGs covered by at least 5 reads were considered for the analysis. For samples with duplicates, the average of methylation was used except for oocyte samples that present a low coverage. For this sample the methylated and unmethylated calls of the two experiments were sum to calculate the methylation ratio. Using the cut off of 5 reads per CpG, the coverage of all experiment vary from 89.6% up to 96.9% of all the CpGs, except for the oocyte methylomes that cover 54.8% of CpGs sites.
The methylomes for oocyte and sperm were screen with a sliding windows approach to identify methylated and umethylated intervals. Windows were defined as 25 consecutive CpGs and was only considered if the methylation levels was present for at least 10 CpG sites. This windows was classified methylated if mean25CpGs—1SD25CpGs > 0.75 and unmethylated if mean25CpGs + 1SD25CpGs < 0.25. Overlapping windows with the same classification were merge and allowed us to identify 40025 unmethylated (Us) and 177787 methylated (Ms) region in sperm and 118853 unmethylated (Uo) and 102858 methylated (Mo) regions in oocyte. A germline DMR was identify when opposite methylated regions in sperm and oocyte overlap for more than 25 CpGs and the position defined by the overlapping difference between methylated regions in sperm and oocyte.
Intermediately methylated region in blastocysts, placenta and somatic tissues were identify using the sliding windows approach with the following criteria 0.2 < mean25CpGs +/- 1.5SD25CpGs < 0.8. Consecutive windows on each sample were fused to generate only a single region. A gDMR was considered to be conserved in preimplantation embryo if the gDMR overlap with a partially methylated region in the blastocyst dataset. To identify the gDMR that persist in somatic tissues, all partially methylated region obtain in the 15 tissues were merge and the number of samples partially methylated for each region is attribute to each region. Only regions > 500 bp were considered to generate the partial methylation region in tissues. To be considered as a ubiquitous gDMR, the partially methylated regions have to persist in the blastocyst and in at least 12 somatic tissues. Placenta-specific gDMR were identified when the partially methylated region is conserved at blastocyst stage but is not observed in additional tissues methylomes. All positional annotations (CpG islands, repeats and gene locations, etc) were obtained from UCSC web browser and genome build hg19.
We used the methyl-seq datasets from GSE63330 that contains placenta methylation information from rhesus macaque, dog, horse, cow and mouse [23]. The orthologous genomic intervals associated with the 551 human oocyte-derived gDMR that maintained an intermediate methylation profiles throughout embryonic reprogramming and in placenta were extracted using the UCSC LiftOver function.
The abundance and genotypes of highly informative exonic SNPs within the transcripts flanking the gDMRs that maintained an intermediate methylation profile in blastocysts were called using Tophat v1.4.0 [50] (for the alignment) and Samtools v1.2 [51] (for the filtering and allelic count) in two published single cell RNA-seq datasets for preimplantation embryos (GSE44183 [18]; GSE36552 [28]). For the purpose of this study the data from individual cells were merged to reconstruct each embryo. In the case of the GSE44183 dataset the embryonic genotypes were compared to the accompanying paternal exome-seq data from the sperm donor’s blood sample.
Genotypes of potential SNPs identified in the UCSC hg19 browser were obtained by PCR and direct sequencing. Sequence traces were interrogated using Sequencher v4.6 (Gene Codes Corporation, MI) to distinguish heterozygous and homozygous samples. Heterozygous sample sets were analyzed for either allelic expression using RT-PCR, methylation-sensitive genotyping or bisulphite PCR, incorporating the polymorphism within the final PCR amplicon so that parental alleles could be distinguished (for primer sequences see S6 Table).
5hmC- 5 μg of heterozygous placenta DNA was subject to DNA Glucosylation using the Epimark kit (New England Biolabs) and the DNA subject to digestion with 100 units of Msp1 for a minimum of 8 hours at 37°C. The DNA was subject to proteinase K digestion prior to PCR.
5mC- Approximately 1 μg of heterozygous placenta DNA was digested with 10 units of HpaII restriction endonuclease for 6 hours at 37°C. The digested DNA was subject to ethanol precipitation and resuspended in a final volume of 20 μl TE. Approximately 50 ng of digested DNA was used in each amplification reaction using Bioline Taq polymerase for 35–40 cycles (for primer sequences see S6 Table). The resulting amplicons were sequenced and the sequences traces compared to those obtained for the corresponding undigested DNA template.
For standard bisulphite conversion approximately 1 μg DNA was subjected to sodium bisulphite treatment and purified using the EZ DNA methylation-Gold kit (ZYMO, Orange, CA). Approximately 2 ul of bisulphite converted DNA was used in each amplification reaction using Immolase Taq polymerase (Bioline) at 45 cycles and the resulting PCR product cloned into pGEM-T easy vector (Promega) for subsequent subcloning and sequencing (for primer sequence see S6 Table).
Surgically separated ICM and TE biopsies were subject to bisulphite conversion using the EZ DNA Methylation-Direct kit (ZYMO, Orange, CA). We employed a multiplex nest PCR approach to maximize data generation. Two sets of primers were designed to each locus and robustly optimized in placenta-derived bisulphite DNA to ensure efficient amplification of both methylated and unmethylated strands at a single annealing temperature without contamination or the formation of primer dimer.
All subsequent outer primers (for ~20 separate loci) were co-amplified in the first reaction using Immolase Taq polymerase (Bioline) for 45 cycles. Second round of amplifications specific to each region, also 45 cycles, utilized locus-specific inner primers using 1ul of first round PCR as template. All second round nested PCR products were subcloned into pGEM-T easy vector for direct sequencing (for primer sequence see S6 Table).
Approximately 50 ng of bisulphite converted DNA was used for pyrosequencing following previously described protocols [22]. Standard bisulphite PCR was used to amplify the imprinted DMRs with the exception that one primer was biotinylated (for primer sequences see S6 Table). For sequencing, forward primers were designed to the complementary strand. The pyrosequencing reaction was carried out on a PyroMark Q96 instrument. The peak heights were determined using Pyro Q-CpG1.0.9 software (Biotage).
|
10.1371/journal.pntd.0000200 | DNA-Sequence Variation Among Schistosoma mekongi Populations and Related Taxa; Phylogeography and the Current Distribution of Asian Schistosomiasis | Schistosomiasis in humans along the lower Mekong River has proven a persistent public health problem in the region. The causative agent is the parasite Schistosoma mekongi (Trematoda: Digenea). A new transmission focus is reported, as well as the first study of genetic variation among S. mekongi populations. The aim is to confirm the identity of the species involved at each known focus of Mekong schistosomiasis transmission, to examine historical relationships among the populations and related taxa, and to provide data for use (a priori) in further studies of the origins, radiation, and future dispersal capabilities of S. mekongi.
DNA sequence data are presented for four populations of S. mekongi from Cambodia and southern Laos, three of which were distinguishable at the COI (cox1) and 12S (rrnS) mitochondrial loci sampled. A phylogeny was estimated for these populations and the other members of the Schistosoma sinensium group. The study provides new DNA sequence data for three new populations and one new locus/population combination. A Bayesian approach is used to estimate divergence dates for events within the S. sinensium group and among the S. mekongi populations.
The date estimates are consistent with phylogeographical hypotheses describing a Pliocene radiation of the S. sinensium group and a mid-Pleistocene invasion of Southeast Asia by S. mekongi. The date estimates also provide Bayesian priors for future work on the evolution of S. mekongi. The public health implications of S. mekongi transmission outside the lower Mekong River are also discussed.
| Schistosomiasis is a disease caused by parasitic worms of the genus Schistosoma. In the lower Mekong river, schistosomiasis in humans is called Mekong schistosomiasis and is caused by Schistosoma mekongi. In the past, Mekong schistosomiasis was known only from the lower Mekong river. Here DNA-sequence variation is used to study the relationships and history of populations of S. mekongi. Populations from other rivers are compared and shown to be S. mekongi, thus confirming that this species is not restricted to only a small section of one river. The dates of divergence among populations are also estimated. Prior to this study it was assumed that S. mekongi originated in Yunnan, China, migrated southwards across Laos and into Cambodia, later becoming extinct in Laos (due to conditions unsuitable for transmission). In contrast, the dates estimated here indicate that S. mekongi entered Cambodia from Vietnam, 2.5–1 Ma. The pattern of genetic variation fits better with a more recent, and ongoing, northwards migration from Cambodia into Laos. The implications are that Mekong schistosomiasis is more widespread than once thought and that the human population at risk is up to 10 times greater than originally estimated. There is also an increased possibility of the spread of Mekong schistosomiasis across Laos.
| Schistosomiasis in humans along the lower Mekong river (specifically Cambodia and southern Laos) was first recognized in 1957 [1] and has proven a persistent public health problem in the region [2]. The species involved is the parasitic blood fluke Schistosoma mekongi Voge, Buckner & Bruce 1978, which uses the caenogastropod snail Neotricula aperta (Temcharoen, 1971) (Gastropoda: Pomatiopsidae: Triculinae) as intermediate host. Published records identify the following foci of S. mekongi transmission: Ban Hat-Xai-Khoun, Khong Island, southern Laos [3] ; Kratié in Kratié Province, northeastern Cambodia, approximately 180 km downstream of Khong Island [4]; and San Dan, Sambour District, also in Kratié Province [5] (Fig. 1A). Prior to 1994 up to 40% of the admissions to Kratié hospital were schistosomiasis-related and deaths were common place [5]. Following mass treatment with the anthelmintic Praziquantel, the prevalence in school children in Kratié Province fell from 40% in 1994 to 14% in 1995 [6]. In Laos, at Khong Island, a nine year Praziquantel intervention programme reduced the prevalence among village children from 51% to 27% [2].
There has been recent optimism regarding the possible complete control of S. mekongi infection [7],[8]; however, this may be unfounded, not only because of the persistence of infection in reservoir hosts [9],[10],[11], but also because the range of N. aperta (and therefore the potential range of the disease) has been underestimated. No new cases of severe morbidity have been reported in Cambodia since 2002, but three new cases of human infection were reported in 2005 [12]; this highlights the resilience of transmission in the face of seven years of mass chemotherapy (beginning 1996). To date all published molecular studies on S. mekongi transmission, and much of the control efforts, have been restricted to little more than a 200 km stretch of the lower Mekong river. Recent surveillance of tributaries of the Mekong river, that drain the Annam highlands (Fig. 1A), and of other river valley systems, have revealed new N. aperta populations. Recently 11 new N. aperta populations, involving six new river systems in Cambodia and Laos, were reported [13]. Many of the new populations lay outside the Mekong river valley, most were reported from the upper Xe Kong river valley. Prior to these studies the total population at risk was estimated to be 120,000 people [2]; however, after taking into account these new populations in areas beyond the lower Mekong river, the potential affected population rises to over 1.5 million. The findings also suggested that areas cleared of N. aperta by control efforts in the Khong Island area or in Kratié could be rapidly recolonized by snails from inaccessible populations in the tributaries draining the Annam mountains. In 2004 S. mekongi was detected in an N. aperta population at Sa Dao in the Xe Kong river of Cambodia (Fig. 1A) [13]; this was the first published direct evidence for the ongoing transmission of S. mekongi outside the Mekong river and further suggests that control of Mekong schistosomiasis will be problematic.
Phylogeographies (incorporating data on DNA-sequence variation) have been used to study the evolutionary radiation of Asian Schistosoma Weinland, 1858 [14]–[16], their relationships with other Schistosoma [17], and their snail intermediate hosts [18]–[20]. Earlier studies suggested that Neotricula arrived in Laos after dispersing southwest from Hunan (China) via the Red river (Fig. 1A) [20]; this at a time when the Yangtze and Red rivers shared a common course some 400 km further West than at present (prior to Pleistocene tectonic events affecting this region [21]). Davis (1992) [18] used snail phylogenies, and Attwood et al. (2002) [16] used DNA-sequence data for Schistosoma, to argue that Schistosoma japonicum Katsurada, 1904 and subsequently S. mekongi diverged from an antecedent resembling Schistosoma sinensium Bao, 1958 in the Shan region of China and Myanmar (Fig. 1B). In this case S. mekongi would be expected to occur in northern and central Laos and even in Vietnam.
Recent DNA-sequence based phylogenies show Schistosoma malayensis Greer et al., 1988 as a sibling species of S. mekongi [22]–[24]. Indeed, the two species differ only in terms of intermediate host, life cycle parameters (e.g. length of pre-patent period), and biogeography [25]. Examination of DNA-sequence variation between the intermediate hosts of S. malayensis and S. mekongi estimated their divergence at 5 Ma (million years ago) [26]; however, palaeogeographical models suggested that the two species were separated less than 1.5 Ma [26]. Després et al. [27] used an ITS2 molecular clock rate of 0.3–0.8% per Myr to date the divergence of S. japonicum from African species at 24–70 Ma. In contrast, Attwood et al. [28] estimated a divergence date of only 12 Ma on the basis of palaeogeography, available dispersal tracts and the radiation of definitive host groups. Després et al. [27] suggested that the divergence of S. haematobium in Africa, from other species infecting animals, was triggered by the colonization of savanna areas by hominids (1–10 Ma). Similarly, the wide host range of S. japonicum (which is a true zoonosis) has been explained as a result of very recent transfers from animals to modern humans [29].
In the present study samples of S. mekongi were taken from all published foci of infection and from a previously unknown population in Lumphat District of Northeast Cambodia; these enabled the first intraspecific study of S. mekongi. Surveys were performed in Lumphat because the region was accessible and there have been suggestions of past transmission in Rattanakiri Province [30]. The Lumphat taxon showed morphological differences (larger eggs and cercariae) from other populations. The work was undertaken to confirm the status of the Lumphat and Sa Dao taxa as S. mekongi, to provide the first divergence date estimates for the radiation of S. mekongi in Southeast Asia (that can be used as priors in future studies), and to estimate a phylogeny for Southeast Asian Schistosoma which can be compared with phylogenies and historical biogeographical hypotheses for the intermediate hosts. The public health implications of the reported data are also considered.
Samples were taken in Cambodia, Laos and Pahang State, West Malaysia. Table 1 gives details of sampling sites, laboratory lines, dates of collection, sample codes, whilst Table 2 details other sources of DNA sequence data. Adult worms were obtained following published methods [16] using the hamster (Mesocricetus auratus), as the laboratory definitive host, and cercariae from naturally infected snail intermediate hosts, but with the following exceptions. The MAL sample was obtained from field trapped rodents by perfusion [31] and the JAP sample was obtained from laboratory lines. Tegumental features (tubercles, spines, etc.), gross internal anatomy and egg morphology were used to identify the worms. DNA was preferentially extracted from females or from separated worm pairs for which eggs had been observed and identified in corpo. Species identification followed relevant publications for S. japonicum [32]–[34] and for S. mekongi [32],[33],[35].
DNA was extracted from single adult worms using a standard method [36]. Sequence variation was assessed at two loci, being partial sequences of the mitochondrial (mt) cytochrome oxidase subunit I gene (cox1) and the small ribosomal-RNA gene (rrnS), here denoted as COI and 12S loci respectively. Sequences of the oligonucleotide primers used in the PCR for the amplification of rrnS locus are published elsewhere [16]. The rrnS region amplified corresponded approximately to positions 11433–11760 in the complete mt genome sequence of Schistosoma spindale Montgomery, 1906 (see Littlewood et al. [37]). The COI locus was amplified using the HCO-2198 and LCO-1490 primer pair [38]; the region amplified using this primer pair corresponded approximately to positions 10224–10851 on the same complete mt sequence. Further details of the data set (including sample sizes and GenBank accession numbers) are given in Table 3. The efficiency of the PCR varied considerably between populations and, in some cases, this effect and the small number of worms available to us, led to a low number of replicates for some populations.
Two mt genes were selected because, with their maternal pattern of inheritance, and smaller effective population size, they were considered to represent potentially better recorders of phylogenetic events at the intra-specific to sibling species level. In addition, the loci targeted were those within regions previously shown to exhibit ideal levels of variation in Schistosoma for this type of study [16], and those which had been used in earlier studies so that data were already available for the outgroup and for comparisons with related taxa.
Total genomic DNA was used as a template for PCR amplification on a Progene thermal cycler (MWG) employing standard PCR conditions [39]. Unincorporated primers and nucleotides were removed from PCR products using the QIAQuick PCR purification kit (QIAGEN). Sequences were determined bidirectionally, directly from the products by thermal-cycle-sequencing using Big Dye fluorescent dye terminators and an ABI 377 automated sequencer (Perkin-Elmer), following procedures recommended by the manufacturers. DNA extracts were not pooled and one DNA sequence thus represented one worm. Sequences were assembled and aligned using Sequencher (version 3.1 Gene Codes Corp. Ann Arbor, Michigan). DNA sequences for both strands were aligned and compared to verify accuracy. Controls without DNA template were included in all PCR runs to exclude any cross-over contamination.
Consensus sequences for the populations sampled were grouped together into sets of aligned sequences of equal length (one set for each locus), such that all taxa were represented in each set (Table 3). In addition, the COI and 12S sequences for each population were concatenated and aligned to form a combined data set. No intrapopulation variation was found among the sequences. Outgroup sequences were taken from the GenBank for Schistosoma incognitum Chandler, 1926 from Central Thailand. Phylogenetic analysis was conducted using both a solely maximum likelihood (ML) approach and a Bayesian method (BM). The present data showed significant variation in the rate of substitution among sites, together with considerable bias among the six different types of nucleotide substitutions. In such cases, ML-based methods are considered more robust than most other commonly used phylogenetic methods, as they permit a better optimized model of substitution [40]. The three data sets were analysed separately by ML and BM.
A suitable substitution model was selected using an hierarchical test of alternative models as implemented in Modeltest v. 3.06 [41]. A General Time Reversible model, with estimates for among site rate heterogeneity (GTR+G), was the model selected for the COI data (the-ln likelihood for this model was 2038.1306, whereas the–ln likelihood for the next more complex model was 2037.0510; X2 = 2.1592, P = 0.0709). The Hasegawa, Kishino and Yano model, again with estimates for among site rate heterogeneity (HKY+G), was the model selected for the 12S data (the-ln likelihood for this model was 850.6237, whereas the–ln likelihood for the next more complex model was 850.1796; X2 = 0.8882, P = 0.1730). The data were partitioned and the appropriate model applied to each partition during the analyses. The data were tested for substitution saturation using plots of the numbers of transitions and transversions against the ML genetic distance (following DeSalle et al. [42]). The indications of these plots were further evaluated using the entropy-based test [43] as found in the DAMBE (v. 4.5.29) software package [44], which provides a statistical test for saturation. Statistics relating to polymorphism (see Table 4) were computed using DNAsp (v. 3.51) [45]. The incongruence length-difference (ILD) test [46], as implemented in PAUP* (v. 4.0b10; [47]), was used to test for homogeneity between the COI and 12S data partitions prior to combining them; the test was applied to informative sites only [48]. In all analyses, gaps were treated as missing data and all characters were run unordered and equally weighted.
For the ML method heuristic searches were performed (under the respective model and starting parameters indicated by Modeltest) using PAUP* with random addition of sequences (10 replicates) and tree-bisection-reconnection branch-swapping options in effect. Nodal support was assessed by bootstrap with 5000 replicates. Starting parameters for the BM were taken from Modeltest; these were then “optimized” using a ML method with the Brent Powell algorithm in the phylogenetics software suite P4 [49]. The values from these optimizations were used as starting parameters for the first Bayesian analyses. A Metropolis-coupled Markov chain Monte Carlo sampling process (McMcMC) [50] was used to search the parameter space of our evolutionary model and compute the posterior probability density.
Although a direct ML method was used in this study this was mainly to afford comparisons with earlier work. The final inferences were made using a BM; this is in accordance with a growing opinion that Bayesian phylogenetic analysis is not only faster in terms of computing time (for analyses with an equivalent level of confidence) but also statistically superior to a solely ML method [51]. For example, such methods do not assume approximate normality or large sample sizes as would general ML methods [52]; they also allow the incorporation of prior information about the phylogenetic process into the analysis. In this study P4 was used to apply the BM; this employs the same method as MrBayes [53] but allows consideration of unresolved trees (i.e. polytomies) and provides an automated (iterative) procedure for tuning the McMC acceptance rates to acceptable levels. The McMC was thereby tuned to give proposal acceptance rates between 10 and 70% for each data partition (this required over 5,000 replicates). The P4 analyses (except for those using the polytomy prior) were repeated in MrBayes (3.1.2) to reveal any topological disagreement.
The priors specified for the BM generally followed the default values found in MrBayes; a flat Dirichlet distribution was set as the prior for the state frequency and for the rate set priors (e.g., revmat, tratio), the branch lengths were unconstrained. A polytomy proposal was set as either zero (i.e., no favouring of multifurcations) or as e, e2 or 10 to examine the effect this has on the posterior probabilities of the clades found; this implements a move (proposed by Lewis et al., 2005) to counter the problem of the spuriously high posterior clade probabilities returned by MrBayes relative to corresponding ML analyses [54]. During the Bayesian analysis, model parameters and relative rates were set to be freely variable; there were four discrete rate categories for the Γ-distribution.
Convergence of the McMC was assessed by plotting split support (for the S. malayensis/mekongi partition) for consensus trees over different generation time windows; the generation of convergence was considered to be that at which the support reached a plateau. In this way, a burnin of 400,000 generations was found to be adequate for all the analyses in this study. Posterior probabilities were then estimated over 900,000 generations beyond the assumed point of stationarity. Four simultaneous Markov chains were run (one cold, three heated) and trees were sampled every 10 generations, two such runs were performed simultaneously. After 900,000 generations (post-stationarity) the average standard deviation of the split frequencies (between the two runs) was checked; the McMC was considered complete if this SD was <0.01.
Likelihood ratio tests (LRTs) were performed to assess the applicability of a molecular clock across the whole phylogeny [55]. The program BEAST (1.4.3) [56],[57] was used to estimate the rates. BEAST implements a Bayesian method for the simultaneous estimation of divergence times, tree topology and clock rates; this method is currently considered superior to other approaches (e.g., non-parametric methods such as NPRS [58] or penalized likelihood methods [59], particularly for phylogenies with a low time depth, because it can allow for uncertainty in dates assigned to calibration points and does not require untested assumptions about the pattern of clock rate variation among lineages [60]. The procedure involves the user specifying both a phylogenetic model (a model of evolutionary history; the tree model) and a clock model (of substitution and rate variation); however, the likelihood calculation is based on the clock model only. Rate variation between adjacent branches is assumed to be uncorrelated, as these rates did not show autocorrelation in recent studies [61]. BEAST can implement several combinations of tree and clock models, but for several combinations it was not possible to obtain a stable result (between replicate McMC chains) or a sufficient effective sample size (ESS) for parameter estimates (sufficient being >200). The program TRACER (1.3) [62] was used to check convergence of the chains to the stationary distribution by visual inspection of plotted posterior estimates and to summarize parameter estimates, errors and confidence intervals. For those models which gave stable results, the ratio of the marginal likelihoods (with respect to the prior) of alternative models (i.e., the Bayes Factor) was used to choose between them [63] (who used importance sampling and the harmonic mean of the sampled likelihoods as an estimator); this does not maximize the likelihoods but averages them over the parameters involved. The calculation was implemented using BEAST (1.5 alpha) following [64]. Divergence dates (Table 5) were taken from the Bayesian posterior distribution of the divergence of the taxa concerned.
The greatest benefit of using a Bayesian method for dating is that the specification of prior distributions can be used to ensure that the analysis realistically incorporates the uncertainty associated with the calibration points used [65]. The models and the priors for the BEAST analyses were set as follows. The tree model prior assumed that divergence patterns followed a Yule process where symmetrical trees are considered more probable (i.e. a simple uniform probability of speciation); this prior and a basic coalescent model (which assumed a constant population size over the time period concerned) were used to obtain the starting tree for the analysis. The clock rates were drawn from either a log normal distribution or an exponential distribution, which were then used to specify the probability of a certain substitution rate on a particular lineage during the McMC. The GTR+G model was applied to the COI partition and HKY+G to the 12S (GTR+ss (ss, site specific rates) could not be used owing to a paucity of polymorphic sites at the first codon position, which causes the BEAST analysis to stall). A normal clock rate prior was specified (0.035±0.0071 substitutions per site per Myr); this was based on rates for S. mansoni and S. incognitum estimated elsewhere [66]). A normal prior (5.0±0.1 Ma) was applied to the TMRCA for the ingroup; this corresponded to the second major Himalayan orogeny which could have isolated central Asian taxa from those of the Orient [28]. For the final parameter estimates three independent runs of 130 million generations were combined to give a final set of 390 million states; the burnin was set to 10%.
Table 4 provides basic statistics for the two loci and the combined data. The COI data appeared the most informative having a greater proportion of parsimony informative polymorphic sites (12.3% of the total number of aligned sites, excluding gaps, compared with 5.4% for 12S). Similarly, 34.6% of positions were polymorphic in the COI data set (of these 35.5% were informative sites, the remaining 64.5% being singletons) and only 23.8% in the 12S set (of which 22.7% were informative). For the COI data 201 mutations were inferred of which 121 (60.2%) were synonymous and 80 (39.8%) were amino acid replacements. The test of Xia et al. [43] suggested that there were no significant levels of substitution saturation at either locus (ISS<ISS.C, P<0.0001, a lack of statistical significance here would imply a poor phylogenetic signal). Table 4 also shows that the nucleotide diversity (D) was greater for the COI data. The haplotype diversities for the full taxa set (H, Table 4) show that not all taxa had unique haplotypes. In the COI set the HXK, SDN and SDO samples shared a common “lower Mekong river” haplotype. Among the 12S sequences SDN, SDO, LMP and S. malayensis shared a common haplotype; that S. malayensis was indistinguishable at this locus highlights its close relationship with S. mekongi. In the combined COI+12S data set each taxon was represented by a unique haplotype, except for SDN (which was identical in state to SDO) which was excluded from the Bayesian analysis. In all cases the test of Tajima (1989) [67] failed to refute the hypothesis of neutral evolution. LRTs for all data sets failed to support the hypothesis that the different lineages had been evolving at the same rate (-ln likelihood with a clock enforced 2932.5967, without clock 2921.57447; X2 = 22.04, P = 0.0005).
Phylogenies estimated using ML showed the same topology with all three data sets, aside from differences due to the number of distinct haplotypes. An LRT comparing the GTR+G and GTR+ss models for the COI data indicated a significant difference between them (X2 = 8.71 P = 0.0128, d.f. = 2) favouring GTR+ss, consequently this model was used for the COI partition in the Bayesian analysis of the combined data set (but GTR+G was used in the BEAST analyses, see METHODS final section). Figure 2 shows the tree resulting from phylogenetic estimation using BM and the COI+12S data set; this tree is identical to that of the ML analysis except that with ML there is an unresolved trichotomy for the three S. mekongi populations. Performing the BM with the polytomy prior turned off resulted in posterior probabilities >0.97 (except for the HXK/SDO node at 0.39) (Fig. 2), increasing the prior to e led to a slight drop in the probabilities, further increases to e2 and 10 had little further effect. The topology and split support using MrBayes was very close to that of P4 with the polytomy prior turned off. Schistosoma malayensis is confirmed as a sibling species of S. mekongi and the S. mekongi populations form a monophyletic unit on the tree; the statistical support for these groupings is high (1.00). The S. mekongi population of LMP appears as sister to a clade comprising the HXK and SDO populations, in the S. mekongi lineage, but this relationship is less well supported (posterior probability only 0.39).
Aside from the Yule process several more complex models of past population dynamics can be implemented using BEAST (e.g. past exponential, logistic or expansive growth and the Bayesian skyline model); however none of these gave stable results (after multiple runs of several 100 million states with tuning and prior-adjustments, or varying starting trees) or they had very low likelihoods. LRT indicated that a strict molecular clock model was inappropriate for these data (P = 0.0005). Consequently the Yule model was used in the final analyses in this study and the log normal and exponential clocks were compared. Table 5 shows the results of a Bayesian estimate using a Yule tree model and an uncorrelated log normal relaxed clock. Comparison of the posterior log likelihood of this model with that for the next best model (Yule process with an uncorrelated exponential clock) gave a Bayes factor of 22.35 which strongly favoured the log normal model. The TMRCA values given in Table 5 are summarized from the Bayesian posterior distribution of the divergence times of the taxa involved in the partition. The exponential clock model gave much lower TMRCAs than the statistically “preferable” log normal model; for example, TMRCA (mekongi) 9,760 years before present (YBP), TMRCA (malayensis) 45,512 YBP, TMRCA (japonicum) 242,690 YBP. Plots for the posterior distribution of estimates of mutation rate and the TMRCAs in the Yule/log normal analysis were bell-shaped and showed no cut-off at the upper or lower bounds; this suggested that the priors used were not restricting the range of values implied by the data [68] (this restriction was found with other model/prior combinations). The wide range of the HPDs in for the divergence time estimates in Table 5 reflects the uncertainty inherent in all molecular date estimates; this is not unusual and is a realistic feature of this method of analysis.
The Schistosoma indicum to ingroup divergence date of around 4.6 Ma (see Table 5) implies a 2.5% (of sites varying per Ma) clock for COI and a 2.0% clock for the 12S locus; these figures appear to be moderate values and compare well with published rates of 1–2% for African and South American Schistosoma at mt loci [27], of 3% for S. indicum-group taxa [28], and of 1% averaged across metazoan groups in general [69]. Attwood et al. [26] suggested a divergence date of 5 Ma for the intermediate hosts of S. mekongi/malayensis, whereas the TMRCA corresponding to this divergence for the parasites themselves in the present study was approximately 2.5 Ma. The S.D. of this estimate is only 4% but the confidence interval is wide, from c.a., 200 KYBP (thousand years before present) to 5 Ma; these wide 95% confidence intervals are common for Bayesian date estimates, they are wider than those of ML based point estimates but this is only because other methods fail to account fully for the uncertainty in the estimation procedure. Attwood et al. [26] used a simple point estimate of divergence times based on pairwise genetic distances (following [70]) and relied on a general invertebrate clock for calibration. Such methodological differences may explain the incongruence between snail and parasite phylogeographies.
The phylogeny in Figure 2 shows all of the Schistosoma mekongi populations, including that of LMP, as lying within a monophyletic clade and this hypothesis is well supported (posterior probability = 1.00). Consequently, it appears that the Schistosoma found in the Srepok river is indeed S. mekongi; this finding has implications for schistosomiasis surveillance in Vietnam. The Srepok river originates in Vietnam and flows westwards into Cambodia. Initial studies suggested that Neotricula aperta evolved in northern Laos/Thailand from a lineage dispersing from India, via Tibet and Yunnan (China), along the Miocene extended upper Irrawaddy and Mekong rivers; the same historical biogeography was assumed for S. mekongi diverging from S. japonicum [71]. However, more recent work suggested an origin for both proto-S. mekongi and proto-N. aperta in Hunan or Guangxi Provinces, China, with a Yangtze-Red river radiation into Cambodia via Vietnam [2]. At least five species of Neotricula Davis, 1986 are known from Hunan but only one from Laos and none from Yunnan; therefore it is more likely that Neotricula and an antecedent of S. mekongi arrived in Vietnam and Cambodia directly from Hunan and not from Yunnan, via Thailand and Laos [20]. Palaeogeographical evidence appears to favour the Vietnam-Cambodia dispersal hypothesis. Much of the Annam mountain chain (which today forms a barrier between Hunan and northern Laos and Vietnam) is Mesozoic and at 1.3 Ma the only trans-Annam dispersal corridor would be the 900 km long valley of the Red river fault, which in the past ran up to 400 km closer to Laos than today [21]. The Pliocene Yangtze is also reported to have flowed along a common course with the Red river [72]. The present data yielded an estimated date for the radiation of S. mekongi in Cambodia of around 1 Ma; this is just before the uplift of (volcanic) highlands in Southeast Cambodia when it would have been possible for S. mekongi to enter Cambodia from Vietnam, just South of the Kontum range. The Srepok river population (LMP) in southern Cambodia is seen as a sister taxon to the other (Xe Kong and lower Mekong river) populations in Figure 2 and may have been early divergent. A phylogenetically basal Srepok river population would be in agreement with the idea of an S. mekongi radiation beginning in Southeast Cambodia; however, the support for this clade was low (posterior probability = 0.39) and only three endemic geographical regions are available for comparison.
The North to South tract, from Yunnan to northern Thailand/Laos and then Cambodia, as proposed in an earlier publication [71] cannot readily explain the absence of S. mekongi from suitable transmission habitats in central Laos. The only known foci of transmission are on the border with Cambodia, around Khong Island at the southern tip of Laos. In contrast, a South to North dispersal together with the Pleistocene (i.e., relatively recent) divergence date estimated here, explains the current range of S. mekongi as a consequence of the limited time available for dispersal from Cambodia into Laos. The Dangrek escarpment lies immediately East of HXK (Fig. 1A); these Mesozoic hills are a likely effective biogeographical barrier between Cambodia and Laos.
S. malayensis has been regarded as a geographical isolate derived from the S. mekongi radiation in Cambodia [2]; however, Figure 2 shows S. malayensis as sister to the S. mekongi clade and the divergence dates of 2.5 Ma estimated for S. malayensis/mekongi and around 3.8 Ma for S. japonicum/Southeast Asian Schistosoma suggest that S. malayensis is basal in the true phylogeny rather than a derivative of S. mekongi. The ancestral definitive hosts of Asian Schistosoma were probably rodents [20]. S. malayensis appears to have retained this ancestral condition, with S. mekongi showing derived character states, that is the ability to utilise humans and Neotricula aperta as definitive and intermediate hosts, respectively. N. aperta is a snail of larger faster rivers than the springs and primary streams to which S. sinensium and all other Neotricula spp. are restricted. The Pliocene Dong-Ngai-Mekong river could have introduced an S. malayensis/mekongi antecedent to the whole Sundaland drainage, with later range contraction, fragmentation and divergence. The divergence time of 2.5 Ma coincides with a major intensification of monsoon winds affecting rainfall and flow patterns in the rivers of the region [73]; this would have impacted on the distribution of the intermediate hosts and could have isolated Cambodian proto-S. mekongi from Malaysian S. malayensis.
The mean date estimates obtained here agree well with palaeogeographical data and hypotheses based on snail phylogenies. For example, the radiation of S. mekongi in Cambodia (dated at 1.3 Ma) correlates well with Pleistocene tectonic upheavals in the region. The severity of late Cenozoic tectonic events in Sundaland strongly suggests that the lower Mekong river (in the area of SDN and Kratié) did not occupy its present course until 5–6 KYBP [74]. Consequently, all known extant S. mekongi populations must have been established mid- to late Pleistocene. The Pleistocene Mekong river itself flowed further west, along the Dangrek escarpment then southwards along the Tonlé Sap of today, and across the Sunda shelf from Kampot (Cambodia) to the present day West Malaysia [75] (Fig. 1A). The divergence of the S. sinensium group from Central Asian lineages (here represented by S. incognitum) dated at 4.6 Ma agrees with the published hypothesis [20], based on snail phylogenies, that the divergence of the S. sinensium group was triggered by isolating events linked to the second major Himalayan uplift (5 Ma). Consequently, the date estimates obtained here are useful priors upon which further studies based on independent data may be undertaken.
The work has demonstrated that transmission at all of the known foci of human schistosomiasis in the lower Mekong Basin involves S. mekongi, including the apparent zoonotic focus in the Srepok river. The phylogeny and divergence dates estimated, although not conclusive, correlate well with the idea of a Vietnam to Cambodia entry of S. mekongi into the lower Mekong region, with a subsequent South to North radiation from Cambodia into Laos. The study also demonstrates the transmission of S. mekongi in the Srepok river close to Vietnam. Such observations and inferences have certain public health implications. The likelihood of finding S. mekongi in Vietnam is increased in the light of these results. The inferred South to North dispersal of S. mekongi implies that it is not ecology but history which is limiting the current distribution of Mekong schistosomiasis. Further work is required into this problem, as, if we have no reason to assume that ecological conditions in Laos are unsuitable for transmission, we may expect the future spread of this disease northwards into Laos. Recent work has already demonstrated that the range of N. aperta is far greater than previously thought (particularly in Central Laos) [13]. The loci used here were chosen for a population phylogenetic study, with no expected intra-population variation, and not for population genetic work. Consequently, the genetic divergence among the S. mekongi populations was relatively small. Further work should involve additional loci and possibly also microsatellites; however, microsatellites are costly to develop and use in endemic countries and are less ideal for dating because they rely on genetic distance estimates of less certain reliability. In spite of low divergence levels, the date estimates obtained were biologically reasonable in the context of independently derived time frames and will be useful priors in future studies. |
10.1371/journal.pntd.0006163 | Seroprevalence of antibodies against chikungunya virus in Singapore resident adult population | We determined the seroprevalence of chikungunya virus (CHIKV) infection in the adult resident population in Singapore following local outbreaks of chikungunya fever (CHIKF) in 2008–2009.
Our cross-sectional study involved residual sera from 3,293 adults aged 18–79 years who had participated in the National Health Survey in 2010. Sera were tested for IgG antibodies against CHIKV and dengue virus (DENV) and neutralizing antibodies against CHIKV.
The prevalence of CHIKV-neutralizing antibodies among Singapore residents aged 18–79 years was 1.9% (95% confidence interval: 1.4%– 2.3%). The CHIKV seroprevalence was highest in the elderly aged 70–79 years at 11.5%, followed by those aged 30–39 years at 3.1%. Men had significantly higher CHIKV seroprevalence than women (2.5% versus 1.3%, p = 0.01). Among the three main ethnic groups, Indians had the highest seroprevalence (3.5%) compared to Chinese (1.6%) and Malays (0.7%) (p = 0.02 and p = 0.01, respectively). Multivariable logistic regression identified adults aged 30–39 years and 70–79 years, men, those of Indian ethnicity and ethnic minority groups, and residence on ground floor of public and private housing apartments as factors that were significantly associated with a higher likelihood of exposure to CHIKV. The overall prevalence of anti-DENV IgG antibodies was 56.8% (95% CI: 55.1%– 58.5%), while 1.5% (95% CI: 1.1%– 2.0%) of adults possessed both neutralizing antibodies against CHIKV and IgG antibodies against DENV.
Singapore remains highly susceptible to CHIKV infection. There is a need to maintain a high degree of vigilance through disease surveillance and vector control. Findings from such serological study, when conducted on a regular periodic basis, could supplement surveillance to provide insights on CHIKV circulation in at-risk population.
| The prevalence of neutralizing antibodies against chikungunya virus (CHIKV) was low at 1.9% among resident adults in Singapore after local outbreaks in 2008–2009. Adults aged 30–39 years and 70–79 years, men, those of Indian ethnicity and ethnic minority groups, and residence on ground floor of public and private housing apartments were significantly associated with a higher likelihood of exposure to CHIKV.
| Chikungunya fever (CHIKF) has re-emerged as an important mosquito-borne disease caused by the Chikungunya virus (CHIKV), an Alphavirus belonging to the Togaviridae family [1], and transmitted by two main vectors, Aedes aegypti and Aedes albopictus in the urban cycle [2]. It is characterized by fever, joint pain, headache and myalgia [3]. The disease was first described during an outbreak in southern Tanzania in 1952 [3,4]. Since then, CHIKV outbreaks had been identified in countries in Africa, Asia, Europe, and the Indian and Pacific Oceans [5]. In late 2013, the first evidence of local CHIKV transmission in the Americas emerged when France reported two laboratory-confirmed autochthonous cases in the French part of the Caribbean island of St Martin [6]. Subsequently, local transmission has been identified in 45 countries or territories throughout the Americas, and more than 2.9 million suspected and confirmed cases and 296 deaths have been reported to the Pan American Health Organization from affected areas as of late July 2016 [5,7].
In Asia, CHIKV was first isolated in Bangkok, Thailand, in 1958, and outbreaks of CHIKF have been reported since the 1960s [8,9]. More recent reports of CHIKF outbreaks in Southeast Asia include those of Indonesia in 2001–2003 [10], Malaysia in 2006–2009 [11,12], Thailand in 2008–2009 [13] and Singapore in 2008–2009 [14]. While the earlier outbreaks were associated with the Asian genotype, the recent resurgence was associated with the East/Central/South African (ECSA) genotype [15], which had also caused the preceding outbreaks in the Indian Ocean islands in 2005 [16], India in 2006–2008 [17] and Sri Lanka in 2006–2007 [18–20].
In Singapore, a tropical city-state, dengue is endemic with all four dengue virus (DENV) serotypes in circulation [21,22]. In response to the regional resurgence of CHIKF and to prevent its introduction into the country, an active laboratory-based surveillance system to detect CHIKV infection was established in late 2006 [23]. General practitioners were requested to consider Chikungunya as a diagnosis when dengue was suspected, and blood samples found to be negative for DENV by polymerase chain reaction (PCR) were routinely tested for CHIKV by PCR and serology. Sporadic imported cases were detected in November 2006. The first confirmed indigenous case infected with ECSA genotype was reported in a 13-year-old Taiwanese student who returned home from Singapore on November 20, 2006 [24]. A localized outbreak of 13 cases, which occurred in an Aedes aegypti predominant urban area, was rapidly contained from January to February 2008 (Fig 1) [25]. However, larger outbreaks subsequently occurred from July 2008 to January 2009 in other rural and suburban areas where Aedes albopictus was the predominant vector (Fig 1) [14]. The local transmission was attributed to the introduction of a mutated ECSA CHIKV with A226V substitution in the E1 gene [19,22]. E2-I211T substitution was also observed in CHIKV isolates from CHIKF-suspected sera used for full genome sequencing [19,20]. These two strains have been associated with efficient transmission by Aedes albopictus [26]. With aggressive vector control measures, the outbreak was finally brought under control in 2009. A total of 1,072 laboratory-confirmed CHIKF cases (260 imported and 812 indigenous) were reported between 2006 and 2009. In 2010, only 26 sporadic laboratory-confirmed cases were reported with 76.9% classified as imported cases [27].
There has been limited information on the seroepidemiology of CHIKV in Singapore and in many countries in South-east Asia. In a serosurvey conducted in Singapore among 531 healthy young adults aged 18–29 years in 2002–2003, two (0.4%) tested positive for IgG antibodies against CHIKV [28]. To assess the impact of the introduction into and spread of CHIKV in Singapore, we undertook a comprehensive serological study to determine its prevalence in the adult resident population.
We used residual sera from the National Health Survey (NHS) in 2010. The NHS 2010 was a population-based cross-sectional survey conducted by the Ministry of Health to determine the prevalence of major non-communicable diseases and their associated risk factors among Singapore adult residents (Singapore citizens and permanent residents) [29]. Selection of the general population was by a combination of disproportionate stratified sampling and systematic sampling. The survey fieldwork was carried out from 17 March to 13 June 2010 in six sites geographically distributed across the country. A total of 4,337 Singapore residents aged 18–79 years participated in the survey, giving a response rate of 57.7%. Only sera from NHS participants who had given informed consent to allow use of their residual sera for further research were included. Ethical approval was given by the Institutional Review Board Ethics Committee of the Health Promotion Board, Singapore (006/2010).
Residual sera from 3,293 (75.9%) of NHS respondents with sufficient amount leftover were tested. All samples analyzed were anonymized. The socio-demographic profile of these survey respondents in our study and the Singapore resident population aged 18–79 years was found to be similar [30].
All residual serum samples were first tested for IgG antibodies against CHIKV and DENV by anti-enzyme-linked immunosorbent assay (ELISA) using commercial test kits (EUROIMMUN, Germany) according to manufacturer's recommended procedure. Titres ≥ 20 RU/mL were considered to be reactive for both tests.
Samples tested positive for CHIKV IgG antibodies were further evaluated for CHIKV-specific neutralizing antibodies using plaque reduction neutralization tests (PRNT). Neutralizing activity of antibodies from human sera samples were tested in duplicates and analyzed by immunofluorescence-based cell infection assay in HEK 293T cells. CHIKV was mixed at a MOI of 10 with diluted (1:1000), heat-inactivated human sera and incubated for two hours at 37°C with gentle agitation (350 rpm). Virus-antibody mixtures were then added to HEK 293T cells seeded in 96-well plates and incubated for 1.5 hours at 37°C. Medium was removed, and cells were replenished with Dulbecco’s modified Eagle’s medium supplied with 10% FBS and incubated for six hours at 37°C. Live cells were determined by staining with a Live/Dead determination dye (Invitrogen) for 20 minutes according to the manufacturer’s protocol, before fixation with 4% paraformaldehyde and immunofluorescence staining. Cells were permeabilized with PBS containing 0.1% Tween-20, 0.1% Triton X-100, 3% BSA, 5% FBS and incubated for 30 minutes at room temperature. Cells were stained with mouse antibody recognizing CHIKV antigen [31] diluted in PBS for one hour at room temperature. This was followed by incubation with goat anti-mouse secondary antibody conjugated to Alexa Fluor 488 (Invitrogen) for one hour at room temperature. Data were acquired using MACSQuant Analyzer (Miltenyi Biotec) and results were analyzed by the FlowJo v10 software (FlowJo, LLC).
Percentage of infection was calculated according to the equation [% infection = 100 x (% infection from neutralization group/% infection from virus infection group)]. In this study, healthy donors lacking anti-CHIKV antibodies were included as negative controls, and infection ≤85% indicated presence of neutralizing activity to CHIKV [32,33]. Strong CHIKV-specific neutralizing activity was defined as ≤50% of 293T cells were infected by CHIKV post-incubation with the sera, moderate as >50% to 75% and weak as >75% to 85%.
Testing of residual sera for IgG antibodies and neutralizing antibodies against CHIKV was approved by National Healthcare Group's Domain Specific Review Board, Singapore (B/2015/01124).
To ensure that the characteristics of the NHS 2010 sample conformed to that of the general population, post-stratification weights were computed based on the age, gender, ethnic group and dwelling type attributes of the Singapore resident population. The overall sample weight was the product of weights for unequal probability of selection and non-response from the household enumeration exercise and survey fieldwork, respectively, and post-stratification weight.
The chi-square test or Fisher’s exact test, where appropriate, was used to test for group differences. Crude odds ratios (cOR) and adjusted odds ratios (aOR) with their 95% confidence intervals (CI) were estimated using univariable and multivariable logistic regression models. Listwise deletion was used for missing data of independent variables in the models. Multivariable logistic regression was used to determine independent factors associated with seropositivity, using forward stepwise selection based on maximum partial likelihood estimation with p < 0.20 for entry of variables and p < 0.05 for removal of variables. All p values reported were two-sided and statistical significance was taken as p < 0.05. Statistical analyses were performed using SPSS software, version 23 (Armonk, NY: IBM Corp., USA).
CHIKV IgG was detected in 71 (2.2%, 95% CI: 1.7%– 2.7%) out of 3,293 survey respondents. Of these 71, 61 had CHIKV-specific neutralizing antibodies–the overall prevalence was 1.9% (95% CI: 1.4%– 2.3%).
Compared with seronegative adults, a higher proportion of those tested seropositive were of age 70–79 years, men, ethnic minority groups categorized under ‘others’, retirees, and resided on ground floor of public housing apartments, private flats and condominiums (Table 1).
The seroprevalence was highest in the elderly aged 70–79 years at 11.5%, followed by those aged 30–39 years at 3.1% (Table 2). Men had significantly higher seroprevalence than women (2.5% versus 1.3%, p = 0.01). Among the three main ethnic groups, those of Indian ethnicity had a higher seroprevalence (3.5%) compared to that of Chinese (1.6%) (p = 0.02) and Malays (0.7%) (p = 0.01). The seroprevalence was highest in the ethnic minority groups categorized as ‘others’ (6.4%), which comprises Eurasians, Caucasians, Japanese, Filipino, Vietnamese, etc. The seroprevalence was also highest among retirees at 6.9%. Adults staying on landed residential properties had the highest seroprevalence at 2.5%. Among the adults living in housing apartments, residents on the ground floor had higher seroprevalence than those residing on second or higher levels (p = 0.002).
In the multivariable regression model, independent factors significantly associated with seropositivity were age group, gender, ethnic group and floor level of residential premises (Table 2). Compared to the age group of 18–29 years, adults in the age group of 30–39 years [adjusted OR (aOR): 6.58, 95% CI 1.90–22.80] and those aged 70–79 years (aOR: 34.28, 95% CI 9.74–120.73) were more likely to be CHIKV seropositive (Table 2). Men were more likely to be seropositive compared to women with an aOR of 2.07 (95% CI 1.21–3.56). Compared to the Chinese, adults of Indian ethnicity (aOR: 2.14, 95% CI 1.03–4.43) and ethnic minority groups categorized as ‘others’ (aOR: 4.02, 95% CI 1.70–9.51) were significantly associated with a higher likelihood of exposure to CHIKV. Compared to adults residing on the 10th floor or higher levels of housing apartments, those who stayed on the ground floor were at higher odds of being CHIKV seropositive (aOR: 6.55, 95% CI 2.30–18.72).
Among the 61 adults with neutralizing activity to CHIKV, 65.6% had strong neutralizing activity while the rest had moderate neutralizing activity and none had weak neutralizing activity. The proportion of CHIKV seropositive respondents having strong neutralizing activity ranged from 61.9% in the age group of 30–39 years to 100% in the age group of 60–69 years (Fig 2).
The overall prevalence of anti-DENV IgG antibodies was 56.8% (95% CI: 55.1%– 58.5%) [30]. A total of 51 adults (1.5%, 95% CI: 1.1%– 2.0%) had both neutralizing antibodies against CHIKV and IgG antibodies against DENV. Ten adults (0.3%) had neutralizing antibodies against CHIKV only, while 1,821 (55.3%) had IgG antibodies against DENV only.
This was the first nationally representative study to describe the seroepidemiology of CHIKV and determine the magnitude of exposure in the Singapore resident population. Our study showed that about 1.9% of the adults in Singapore had likely been exposed to CHIKV, which was much lower than that of DENV with a prevalence of 56.8% in the same population studied [30]. This is not surprising as dengue has long been endemic in Singapore since the first outbreak reported in 1960 [34,35].
The CHIKV seroprevalence in adults 18–29 years of age (0.5%) after the 2008–2009 outbreaks was similar to that of a smaller study [28] in the same age group (0.4%) five years before these outbreaks (p = 0.84). However, the findings of these two studies may not be directly comparable due to differences in laboratory methods. In the earlier serological study in 2002–2003, IgG antibodies were detected using CHIKV-infected cells on teflon-coated glass slide as antigen (Ooi EE, Duke-National University of Singapore Medical School, Singapore, personal communication). Nevertheless, this showed that transmission of infection among young adults in the community was relatively low before and during the outbreak.
The post-outbreak seroprevalence in Singapore was lower compared to that of studies in some countries. In north-eastern Italy, the prevalence of IgG antibody against CHIKV by the indirect immunofluorescence method was 10.2% in 325 residents across all ages surveyed after an outbreak in 2007 [36]. In four Malaysian outbreak-free states, 5.9% in 945 healthy adults aged 35–74 years recruited in 2008 tested positive for CHIKV IgG by ELISA [37]. In Cebu City, the Philippines, 22.0% in 150 individuals across all ages were seropositive for CHIKV from a cross-sectional study using neutralization assays in 1973, while a prospective fever cohort study in 2012 found that 28.3% in 853 residents ≥6 months of age had PRNT titers at baseline indicating a history of CHIKV infection [38]. In central and southern Thailand, a study involving serum samples of 835 individuals aged between 6 months and 60 years obtained in 2014 and analyzed by commercial ELISA test kits found that 26.8% were seropositive for CHIKV, while 24.4% possessed both anti-CHIKV and anti-DENV IgG antibodies [39].
In our study, age group, gender and ethnic group were independently associated with CHIKV seropositivity in the multivariable logistic regression analysis. This generally corresponded with demographic characteristics of reported indigenous laboratory-confirmed cases of CHIKF among Singapore residents, with higher incidence rates in adults aged 30–39 years, men, and ethnic minority groups categorized as ‘others’. The incidence rate of indigenous cases was consistently highest in ethnic minority groups at 7.8 and 4.2 per 100,000 Singapore resident population, followed by Chinese at 6.1 and 3.1 per 100,000 population in 2008 and 2009, respectively [14]. In 2013, another outbreak year, ethnic minority groups also had the highest incidence rate of indigenous cases among Singapore residents (26.9 per 100,000) followed by Chinese (9.4 per 100,000) [40].
CHIKF was known to have swept through Southeast Asia in the 1960s and 1970s [8]. Even though CHIKV was not widely tested then, one patient was incidentally tested positive for CHIKV infection by complement fixation tests and neutralization tests in 1960 during an outbreak of dengue hemorrhagic fever in Singapore [41]. Considering that the vector for dengue and chikungunya is the same Aedes mosquitoes, and the Aedes house index (percentage of houses infested with Aedes larvae/pupae) was in the range of 30–50% at that time, it is likely that Singapore was not spared from disease transmission during the regional CHIKF outbreak, which would have contributed to the highest seroprevalence observed among adults aged 70–79 years. The lower proportion of CHIKV seropositive respondents in this age group having strong neutralizing activity (Fig 2) could be due in part to the decline in neutralizing antibody titers over time.
The gender difference for CHIK seroprevalence in our study had also been observed for dengue-specific IgG prevalence [30]. Higher CHIKV seroprevalence in men was also reported in serological studies in Malaysia [37] and the Indian Ocean island of Mayotte [42]. It has been postulated that the gender differential in the risk of CHIKV infection could be attributed to specific behaviour that results in greater exposure to bites by Aedes mosquitoes, and less tendency toward individual protection [36,42,43].
The ethnic difference in seroprevalence could be partly due to exposure during travel to highly endemic countries in the region. The high incidence among Indians corresponded with the highest rate of imported cases among Singapore residents travelling to India during the 2008–2009 outbreaks [14]. Phylogenetic data revealed that the first three reported episodes of local transmission in 2008 were due to three genetically distinct viruses of different geographic origins, suggesting that these episodes may be due to independent importations of CHIKV, most likely from India, Malaysia, and Sri Lanka [22].
The seroprevalence was highest among adults staying on landed residential properties. The incidence rate of indigenous cases of CHIKF and dengue has consistently been the highest for those living on landed residential properties, where there are more potential Aedes breeding habitats.
The proportion of adults aged 18–79 years with both neutralizing antibodies against CHIKV and IgG antibodies against DENV detected was low at 1.5%. Among the 1,872 samples tested positive for IgG antibodies against DENV in our serological study, 51 (2.7%) also had neutralizing antibodies against CHIKV. In Thailand, a seroprevalence study to evaluate evidence of past infection against CHIKV and DENV found that 79.2% (661/835) of individuals aged between 6 months and 60 years were DENV-seropositive, of whom 30.9% (204/661) also had IgG antibodies against CHIKV [39].
The first concurrent isolation of CHIKV and dengue type 2 virus (DENV-2) was from a single blood specimen taken from a patient in the acute phase of a dengue-like illness in southern India in 1964 [44]. Since then, a number of cases of co-infection with DENV and CHIKV have been detected in countries/territories such as Angola, Gabon, India, Madagascar, Malaysia, Myanmar, Nigeria, Saint Martin, Singapore, Sri Lanka, Tanzania, Thailand and Yemen; these constitute only 13 out of the 98 countries/territories where both chikungunya and dengue epidemic/endemic transmission have been reported based on literature search conducted until May 2015 for all relevant articles [45]. This included report of one case of imported co-infection of CHIKV and DENV-2 who had returned to Taiwan from Singapore in 2010 [46]. In Singapore, the active laboratory-based surveillance initiated in late 2006 was confined to DENV-negative blood samples for detection of CHIKV, hence cases co-infected with CHIKV and DENV could be potentially missed out.
The vectors for CHIKF and dengue, Aedes albopictus and Aedes aegypti, are distributed throughout Singapore. Aedes aegypti thrives in urban areas while Aedes albopictus inhabits in higher proportion in less urbanized areas with greenery in Singapore. A well-established nationwide Aedes surveillance and control programme incorporating source reduction, public education, community participation, and law enforcement, has been in place over the last four decades [47]. The overall Aedes house index has been maintained at around 1–2%. Despite the aggressive vector control efforts, CHIKF re-emerged in 2013 with a total of 1,059 laboratory-confirmed cases (95.5% indigenous and 4.5% imported), 1.5 times the 718 laboratory-confirmed cases reported in 2008 (Fig 1) [40]. The recurrence of the outbreak of CHIKF coincided with the largest dengue epidemic in the same year, indicating that similar factors may have facilitated the upsurge in the number of cases of these two mosquito-borne diseases in Singapore [48]. Phylogenetic analysis revealed that while locally transmitted CHIKV strains in 2013 formed a monophyletic group within the ECSA genotype, they possessed a signature of two synonymous substitutions (C639T + C816A) in E1 gene, making them a genetically distinct group [40]. These findings, together with the long-term absence of CHIKV transmission on an outbreak scale in Singapore, supported a viral introduction event prior to the establishment of indigenous transmission during the CHIKV outbreak in 2013. Outbreak strains possessed E1-A226V substitution, which further supported the potential role of Aedes albopictus as a predominant vector in CHIKV transmission in the 2013 outbreak. Imported virus strains belonged to the ECSA and Asian genotypes and did not possess E1-A226V substitution [40]. The ECSA strains all shared an Indian sub-continent ancestry. While imported strains with the ECSA genotype clustered separately from outbreak strains and were sporadically detected during 2009–2013, those with the Asian genotype were mainly from the Philippines and Indonesia [40].
There are a few limitations in our study. Some of the positive tests for CHIKV infection could have cross-reacted with other arboviruses such as O'nyong-nyong virus (ONNV), Ross River virus (RRV) and Barmah Forest virus (BFV) [33,49–51]. However, we have no data on the prevalence of other alphaviruses in our local population. There has been no or limited data comparing the relative sensitivity or specificity of the available CHIKV diagnostic assays [52]. The low seroprevalence in our study was consistent with sporadic detection of clinical cases. To establish past exposure to CHIKV, we used PRNTs which are deemed to be specific for alphaviruses and serve as the gold standard for confirmation of serological test results [53]. As our study was carried out based on residual sera and not specifically for CHIKV infection, clinical signs and travel history were not recorded.
The re-emergence and spread of CHIKF have been attributed to several factors, including vast immunologically naïve human populations, viral adaption, enhanced efficiency of mosquito transmission, drastic increase in international travel, as well as climate and environmental changes [9,48,54]. As Singapore remains highly susceptible to CHIKV infection, there is a need to maintain a high degree of vigilance through disease surveillance and vector control. Findings from such serological study, when conducted on a regular periodic basis, could supplement surveillance to provide insights on CHIKV circulation and profile of at-risk population.
|
10.1371/journal.pntd.0001824 | Age-Related Patterns in Human Myeloid Dendritic Cell Populations in People Exposed to Schistosoma haematobium Infection | Urogenital schistosomiasis is caused by the helminth parasite Schistosoma haematobium. In high transmission areas, children acquire schistosome infection early in life with infection levels peaking in early childhood and subsequently declining in late childhood. This age-related infection profile is thought to result from the gradual development of protective acquired immunity. Age-related differences in schistosome-specific humoral and cellular responses have been reported from several field studies. However there has not yet been a systematic study of the age-related changes in human dendritic cells, the drivers of T cell polarisation.
Peripheral blood mononuclear cells were obtained from a cohort of 61 Zimbabwean aged 5–45 years with a S. haematobium prevalence of 47.5%. Two subsets of dendritic cells, myeloid and plasmacytoid dentritic cells (mDCs and pDCs), were analyzed by flow cytometry.
In this population, schistosome infection levels peaked in the youngest age group (5–9 years), and declined in late childhood and adulthood (10+ years). The proportions of both mDCs and pDCs varied with age. However, for mDCs the age profile depended on host infection status. In the youngest age group infected people had enhanced proportions of mDCs as well as lower levels of HLA-DR on mDCs than un-infected people. In the older age groups (10–13 and 14–45 years) infected people had lower proportions of mDCs compared to un-infected individuals, but no infection status-related differences were observed in their levels of HLA-DR. Moreover mDC proportions correlated with levels of schistosome-specific IgG, which can be associated with protective immunity. In contrast proportions of pDCs varied with host age, but not with infection status.
Our results show that dendritic cell proportions and activation in a human population living in schistosome-endemic areas vary with host age reflecting differences in cumulative history of exposure to schistosome infection.
| A characteristic feature of most helminth infections is the convex age infection profile, where infection levels rise to peak in early childhood and decline in adulthood, a pattern thought to result from the development of protective acquired immunity. Thus, several investigations characterizing protective responses to inform vaccine research have focused on responses present in older people, who despite continued exposure to infection carry
little or no infection. To date, such studies have identified key responses which are correlates of resistance. However, there is a paucity of information on cell types that are mediators rather than effectors of the immune responses. One such group where there are limited studies in human schistosome infections is dendritic cells which are important for the polarizations of CD4+ T cell responses. Therefore, we characterized the age profile of dendritic cells in Zimbabweans exposed to Schistosoma haematobium infection. We found an age-related pattern in the proportions of myeloid dendritic cells (a subset of dendritic cells) in this population. Furthermore, in the case myeloid dendritic cells, the age profile differed between schistosome infected and un-infected people. Thus our study suggests that activation and migration of myeloid dendritic cells also develop in an age-related pattern consistent with the cumulative history of exposure to schistosome parasites.
| Schistosoma haematobium helminth parasites cause urogenital schistosomiasis which affects about 112 million people mainly in rural areas of subtropical countries [1]. Infections with S. haematobium are most common in school-age children. Populations in endemic areas show a characteristic age-infection profile with infection levels increasing to peak in early to late childhood, typically around 9–15 years and then declining in adulthood [2]–[4]. This profile is believed to be largely reflective of the gradual development of protective acquired immunity reducing re-infection levels [2], [5], [6]. The consequence of this age profile is that individuals with comparable infection levels who have resided in a schistosome endemic area since birth (e.g. egg negative children versus egg negative adults) can differ significantly in their immune response against the parasite and thus their levels or resistance to re-infection. Several studies characterizing human immune responses to schistosome infections have shown age related differences in antibody levels [7]–[9], plasma cytokines [10], parasite-specific cytokines [11] and regulatory T cell proportions [12]. This concept is supported by theoretical modelling of the development of acquired immunity, which predicts that correlations between immunological responses and infection is positive in younger age groups and subsequently decreasing and potentially turning in a negative correlation [13]. Moreover this pattern is not restricted to schistosome infections, but is a characteristic of immune responses associated with protection from a variety of helminth species in endemically-exposed populations [14], [15].
As investigations of the nature and development of protective acquired immunity progress, there has been a move to decipher the mechanisms and pathways behind these observed age-related patterns. Human field studies show that the immune response against schistosomes is characterized by a very complex interaction of TH1, TH2 and regulatory responses [10]–[12], [16]–[18] with differences between natural human infections and experimental mouse models [19], [20]. Nevertheless, the induction of a TH2 response is important in the immune response against schistosomes [21] and in particular for the development of protection [22]. As a major antigen-presenting cell population dendritic cells (DCs) are responsible for acquiring, processing and presenting parasite antigens to T cells and the latter interaction leads to activation and polarisation of the acquired immune response. The importance of DCs in the induction of the TH2 immune responses in the context of schistosome infections has been recently highlighted in an experimental mouse model [23]. This study indicates that DCs are required for the induction and development of TH2 responses during schistosome infections supporting other studies showing the importance of DCs in the context of TH2 induction [24]–[27].
It has been shown that schistosome derived compounds can activate dendritic cells by toll-like receptors (TLR)-2 and 3 [28], [29] and early TH2-promoing activities can be diminished by TLR-3 [30]. Overall studies are rare addressing the role of the innate immune cells in the immune response against schistosomes [31].
In particular, studies analysing DCs directly from humans exposed to schistosomes are limited. The first study addressing this question, published by Everts and colleagues in 2010, elegantly showed changes in DCs during chronic schistosomiasis (S. haematobium) [32] suggesting that chronic schistosomiasis can suppress DCs, which might play a role in immune modulation by schistosomes.
Peripheral blood DCs can be divided in two major subsets [33], [34], myeloid dendritic cells (mDC; or conventional DC) and plasmacytoid (pDC). Myeloid DCs express CD11c, but low levels of CD123 and have a more pronounced role in antigen processing and initiation of T cell response [35]. Plasmacytoid DCs express CD303 (BDCA-2), ILT7, high levels of CD123 (IL-3Rα), but are negative for the classical marker CD11c [33], [36]–[38]. These pDCs play a critical role in anti-viral immunity as well as in immune tolerance [39], [40]. Everts et al's [32] study focusing on people aged 17–39 years showed that the frequencies of mDCs and pDCs are reduced in infected people compared to un-infected and that mDCs are functionally impaired in response to toll-like receptor ligands and in driving T cell responses ex vivo. Since in most schistosome endemic areas people are exposed to infection from as young as 6 months old [41], and may already be carrying heavy infections within the first decade of life [8], these early infection events may have a profound effect on the proportions of the DCs present in the host. Characterising the role of these cells during natural immune responses to helminth infection in the different age groups is vital for vaccine development, since people targeted by anti-schistosome vaccines in endemic areas will have been previously exposed to infection and their cellular immune responses will already be primed by repeated exposure to parasite antigens. In this context variations in antigen presenting cell populations, including DCs, might directly affect the efficacy of vaccination in different age groups. Lessons from the discontinued human hookworm vaccine trials, illustrate the importance of characterising existing natural immune responses in endemic populations and designing vaccine to avoid undesirable pathological outcomes of vaccination [42]. Therefore, the aim of this study was to characterise the relationship between age and DCs during natural schistosome infections and determine if these patterns are affected by host infection status across different age groups reflecting different dynamics in the acquisition and loss of schistosome infection.
This study was performed in Chipinda village which is located in the Mashonaland East Province in Zimbabwe (31°94′E; 17°67′S). The village was selected because health surveys regularly conducted in the region showed little or no infection with soil-transmitted helminths (STH) and a low S. mansoni prevalence (<2%), which elicit immune responses that cross-react with those against S. haematobium [43], [44], [45]. The low STH and S. mansoni prevalence is consistent with earlier surveys in this area of Zimbabwe [46]–[48]. Villagers are subsistence farmers who have frequent contact with water sources posing a risk of infection (as assessed by questionnaires) due to insufficient safe water provision and low coverage of sanitation facilities as is typical in rural Zimbabwe. Drinking water is collected from open wells while bathing and washing is conducted in perennial rivers surrounding the village. This area has not been included in any Schistosome Control Programmes and therefore participants had not received any anti-helminthic treatment for schistosomiasis or other helminth infections prior to this study. Thus their natural immune responses could be assessed in the absence of drug-altered schistosome-specific responses [49], [50]. Urine and stool samples were collected on three consecutive days and examined microscopically for S. haematobium (urine filtration, Mott method [51]) or S. mansoni and intestinal helminths (Kato-Katz method [52]) respectively using standard procedures. Participants were screened for malaria by microscopic examination of Giemsa stained blood smears and HIV status was determined by immunochromatography (DoubleCheckGold™ HIV 1&2, Orgenics) with HIV positive samples subsequently re-tested by a second rapid assay (Determine HIV 1/2 Ag/Ab Combo, InvernessMedical) to confirm HIV status [53]. Study participants had to meet the following criteria: (1) be life-long residents in this area (assessed by questionnaires) so that age would be a proxy for duration of exposure to S. haematobium infection, (2) should not have received anti-helminthic treatment prior to this study, (3) should have provided at least two urine and two stool samples on consecutive days for parasitological diagnosis, (4) should have tested negative for intestinal helminths including S. mansoni to focus on single infections with S. haematobium, (5) should be negative for HIV and malaria (Prevalence of both HIV and malaria were too small for inclusion in the statistical analysis (HIV prevalence in Chipinda village was 8% and no-one was positive for malaria infection at the time of sampling) and (6) have provided a sufficient volume of blood to isolate peripheral blood mononuclear cells (PBMC). The selected cohort comprised 61 individuals and details to the cohort are provided in Table 1.
Permission to conduct the study in the region was obtained from the Provincial Medical Director and institutional and ethical approval was received from the University of Zimbabwe's Institute Review Board and the Medical Research Council of Zimbabwe respectively. Only compliant participants were recruited and they were free to drop out at any point during the study. At the beginning of the study, participants and their parents/guardians (in case of children) had the aims and procedures of the project explained fully in the local language, Shona, and written consent was obtained from participants and parents/guardian before parasitology and blood samples were obtained. After collection of all samples, all participants were offered anti-helminthic treatment with the recommended dose of praziquantel (40 mg/kg of body weight).
Depending on age of the participants up to 25 ml of venous blood was collected in heparinized tubes of which approximately 5 ml was used for serological assays as well as microscopic detection of malaria parasites. The remaining blood was used for the isolation of peripheral blood mononuclear cells (PBMC) through density gradient centrifugation using Lymphoprep (Axis-Shield, Cambridgeshire, UK). These PBMC were subsequently enumerated, cryo-preserved in 10% DMSO, 90% fetal calf serum and stored in liquid nitrogen in Zimbabwe prior to shipping to Edinburgh in dry shippers for assaying.
Thawing of cryo-preserved PBMC was performed by rotating cryovials in a 37°C water bath until a small crystal was remaining in the cell suspension. Cells were then slowly re-suspended in RPMI 1640 supplemented with 10% FCS, 2 mM L-glutamine and 100 U Penicillin/Streptomycin (all Lonza, Verviers, Belgium). Cells were washed twice with media, counted and viability assessed using trypan blue (Sigma-Aldrich, Dorsert, UK). The median viability of the PBMCs was 71.4% which is within the range of published values [54]. In addition it was confirmed that including a viability marker did not change analysis of subsets. Afterwards, cells were washed with Dulbecco's-PBS (Lonza) and surface stained with the following antibodies: Qdot-605-conjugated anti-CD14 (clone TUK4, Invitrogen), FITC-conjugated anti-CD11c (clone Bu15, Invitrogen), APC-H7-conjugated anti-HLA-DR (clone L243, BD Biosciences), PE-Cy5. anti-CD123 (clone 9F5, BD Biosciences) and V450 BD Horizon-conjugated anti-CD86 (clone 2331, BD Biosciences), PE-conjugated anti-BDCA-2 (clone AC144, Miltenyi Biotec) and APC-conjugated anti-BDCA-4 (clone AD5-17F6, Miltenyi Biotec). Stained cells were acquired on a FACSCantoII (BD Biosciences) and analyzed using FlowJo software software (TreeStar, USA).
Serum antibody levels were measured by enzyme-linked immunosorbent assays (ELISA) following established protocols [8], [55]. Whole worm homogenate (WWH) was obtained from the Theodor Bilharz Research Institute (Giza, Egypt). In short, microtiter plates were coated overnight at 4°C at 10 µg/ml. Serum samples were diluted at 1∶20 for WWH-IgE and 1∶100 for WWH-IgG and IgM and incubated for 2 hours at 37°C. Horse-radish peroxide conjugated antibodies were diluted 1∶1000 for IgG, IgE and IgM and incubated for 1 hour at 37°C. ELISAs were developed using ABTS (Southern Biotech) and stopped after 15 minutes for IgG and IgM, and 30 minutes for IgE. Absorbance was read at 405 nm. Serum was available from 45 out of 61 individuals.
The proportions of DC cell subsets were square root arcsine transformed whereas expression levels of HLA-DR and CD86 were square root transformed to allow the use of parametric tests in subsequent analyses [56]. To analyse which factors influence the proportion of DC subsets, a univariate analysis of variance using sex (male/female), host age groups (group 1: 5–9 years; group 2: 10–13 years; group 3: 14+ years) and infection status (un-infected = 0 mean egg count per 10 ml and infected >0 mean egg count per 10 ml; all variables categorical) as independent variables was performed. The three different age groups were selected to reflect epidemiological groupings (i.e. where infections are acquired, peak and decline) by age and infection intensity and to obtain comparable sample sizes between the groups. Post hoc tests between infected and un-infected individuals were conducted in each age group. In addition post hoc analysis was performed to determine differences in pDCs between the three age groups. To determine if the mDC populations showed an age profile consistent with those for protective immune responses as predicted by quantitative studies [13], correlation analyses between infection intensity (log10 (mean egg count+1) transformed) and proportions of mDCs was conducted after allowing for the effect of sex. The correlation coefficients were then tested for homogeneity using the Fisher's r-to-z transformation [57]. Correlation analyses between WWH-specific IgG and DC cell subsets were performed after allowing for the effects of sex, age and infection intensity. All statistical tests were conducted using the software package SPSS v14 and p values were taken to be significant at p<0.05.
This study was designed to focus on the immune modulation in a population exposed to S. haematobium. The northeast part of Zimbabwe is endemic for S. haematobium with many regions having a moderate to high prevalence of S. haematobium, but low prevalence of S. mansoni and soil-transmitted helminths [47], [48]. The overall prevalence in Chipinda was (39%). This is defined by the World Health Organisation (WHO) as a moderate transmission area [58]. The overall prevalence of S. haematobium in the selected study population (N = 61) was 47.5% which is higher but not significantly different from (χ2 = 1.688, df = 1, p = 0.097) the village prevalence of 39% and still within the WHO definition of moderate transmission. The difference in infection prevalence between the youngest age group (35.3%) compared to more than 50% in the older age groups was not significant (χ2 = 0.806, df = 1, p = 0.185; details in Table 1). However, youngest individuals (5–9 years) who were infected carried high infection levels (Figure 1 and Table 1). Individuals aged 10–13 years still showed high infection levels, but with a higher prevalence than the first age group. In contrast, in the oldest age group (14–45 years), the prevalence remained high, but most individuals show lower infection intensities (Figure 1). The consequence of the infection profile is that individuals who are life long residents in the area and have never received treatment with anti-helminthic drugs (see selection criteria in Methods) differ in their cumulative histories of exposure despite carrying comparable infection intensities. This was supported by an analysis of the serum levels of adult worm specific (whole worm homogenate – WWH) antibody levels. As shown in Figure 2A and B levels of WWH-specific IgE and IgG, which are associated with history of infection and resistance to infection [59]–[61], increase significantly with age. In contrast IgM against WWH as marker of current infection starts to decrease in the oldest age group (Figure 2C).
To analyze mDCs and pDCs, PBMC were gated on CD14negHLA-DR+ cells (Figure 3A, B). Gated cells were stained with CD123 and CD11c to distinguish mDCs (CD11c+CD123neg/low) from pDCs (CD123hiCD11cneg; Figure 3C). To further verify specificity of pDC gated CD123hiCD11cneg were analysed for the expression of BDAC-2/CD303 and BDCA-4/CD304 [62] and granularity of gated cells was analysed in an FSC/SSC plot (Figure 3D). Therefore other cell types such as basophiles or B cell potentially able to express CD123 are excluded from the analysis. mDC did not express either BDCA-2 or BDCA-4 (Figure 3E). Subsets were expressed as percentages of PBMC or as percentage of CD14negHLA-DR+ cells. Both subsets were subsequently analysed for the expression of CD86 (Figure 3D, E) and levels of HLA-DR within the two different DC populations as indicators of DC activation status and their ability to present antigen and co-stimulatory signals to T cells.
When the study population was portioned by infection status (un-infected versus infected), neither mDCs (Figure 4A) nor pDCs (Figure 4B) showed a significant difference between both populations. Comparable results were obtained if DC subsets were expressed as percentage of CD14negHLA-DR+ (data not shown). Myeloid DCs showed a significant association with sex (Table 2), in which female had slightly higher proportions than males. This was then statistically accounted for in all subsequent analyses.
The age in the total population ranged from 5–45 years. Age group significantly affected the proportions of pDCs (Table 2), with a significant increase between 5–9 year olds (age group 1) and 14+ year olds (age group 3; p = 0.033). In contrast for mDCs the relationship with age varied depending on infection status (significant interaction between age group and infection status; Table 2). Based on these results post hoc analysis of mDC proportions by infection status was performed after partitioning the study population into the three different age groups (details in Table 1). In the youngest age group (Figure 5A) infected people had significantly higher mDC percentages than un-infected people. Results were comparable if expressed as percentage of CD14negHLA-DR+ (Figure S1A). This pattern differed in the second age group, where infected individuals showed fewer mDCs than un-infected individuals (Figure 5B) a difference which was even more pronounced in the oldest age groups (Figure 5C). A comparable analysis could be made by correlating infection intensity (rather than infection status) to percentages of mDCs in the three different age groups. In the youngest age group both parameters were positively correlated (b = 0.707, p = 0.001), whereas in the second age group the correlation was instead negative (b = −0.441, p = 0.018). This negative correlation was more pronounced in the third group (b = −0.540, p = 0.007; Figure 6). A test for the homogeneity of the correlations coefficients showed a significant difference between age group one and two (z = 3.89, p = 0.0001) and between the first and the third age group (z = 4.17, p<0.0001).
In contrast, to the clear picture in the case of mDCs, there was no significant difference in the proportion of pDCs (expressed either relative to live PBMC or CD14negHLA-DR+) between un-infected and infected people in any of the three age groups as shown in Figure 7, Figure S1B and Table 2.
Up-regulation of CD86 is hallmark of maturation and activation of DCs and their ability to provide co-stimulatory signals to T cells. HLA-DR expression can be used as an indicator both of DC activation status and their potential to activate antigen-specific T cells. Neither infection status nor the interaction between infection status and age group influenced expression of CD86 on mDC as determined by ANOVA (Table 2), which was confirmed by post hoc tests comparing CD86 expression between un-infected and infected people after partitioning into age group (Figure 8A). However HLA-DR expression was significantly influenced by an interaction of age group and infection status (Table 2). Post hoc test showed that HLA-DR expression differed between un-infected and infected individuals, in the youngest age group. In this age group HLA-DR was significantly lower on mDCs from infected people compared to un-infected (Figure 8B). In contrast no difference was observed in HLA-DR expression in people 10–13 years of age or in the oldest age group. Neither CD86 nor HLA-DR expression on pDCs were dependent on age group, infection status or the interaction of these variables (Table 2).
Adult worm specific IgG can be used as a marker development of resistance against infection [59]–[61]. WWH-specific IgG increases with age and in our population especially between the first and second age group (Figure 2B) and to the same time changes in mDC proportions between un-infected and infected individuals occurred. Therefore the interaction between mDC and WWH-specific IgG was analysed. As indicated in Table S1, WWH-specific IgG is influenced by sex, age and infection intensity, but addition also by proportions of mDC. mDC are directly correlated to WWH-specific IgG after allowing for the effects of sex, age and infection (Figure 9A). Although showing the same tendency pDCs and WWH-specific IgG were not significantly correlated (Figure 9B). This correlation was most obvious in the oldest age group (Table S2) and overall more significant in un-infected (protected) individuals. Correlation between mDC proportions and WWH-specific IgE and IgM were not significant.
Urogenital schistosomiasis caused by S. haematobium shows a characteristic age-infection profile with increasing infections intensities during early childhood, peaking usually between the age of 9–15 years and than slowly declining towards adulthood [2]–[4]. This study investigated whether the age-related changes in schistosome infection pattern impacted on the proportions and phenotype of DCs.
We provide evidence that indeed there is an age-related pattern in the proportions of mDCs, if infected and un-infected people were compared. The proportion of mDCs were found to be present in significantly higher proportions in infected individuals compared to un-infected individuals in the youngest age group carrying heaviest infection levels. This subsequently changes in older age groups, were older infected individuals had lower relative numbers of mDCs compared to un-infected people. The gradual change between the three age groups became more apparent when the correlation between infection intensity (rather than infection status) and mDCs proportions were considered. In this case both parameters were positively correlated in the youngest age group and became a negative correlation with increasing age. This pattern is consistent with that predicted by quantitative studies for protective immune responses and has already been demonstrated for antibody responses associated with protection to re/infection in human hookworm [63] and schistosome infections [64]. Importantly levels of adult worm specific IgG are correlated with proportions of mDCs. This effect was more significant in the oldest age group, in which protective immunity becomes effective.
Our data confirm some of the major findings of a previous study of DC populations in human schistosomiasis conducted by Everts et al [32]. Proportions of mDCs were lower in an age group 14–45 years (mean 22.6 years) which is comparable to the population studied by Everts et al (17–39 years). Since cellular immune responses might depend on many different parameters such as transmission dynamics or genetic background of the investigated study population the consistency in DC proportions in schistosome-exposed human between our study and that by Everts et al, by itself is an important finding. A major difference between the two studies is that we did not observe differences in pDC percentages in regards to infection status, although we had high variation in pDC levels in the oldest age group. In deed a slight tendency of lower pDC proportions (expressed as percentage of PBMC) was observed in this age group. Differences in co-infections may also account for the differences between our observations and those of Everts et al [32] in Gabon. While the population in Zimbabwe had no co-infection with other helminths, Plasmodium or HIV, that in Gabon had co-infections with microfilaria observed in 48% of the schistosome infected individuals which could affect the pDCs proportions.
Both our study and that of Everts et al [32] clearly indicate the difference in mDC populations between infected and un-infected adults. However, since infections in moderate to high transmission areas occur at relatively young age, we were interested in the dynamics of the DCs over time, which was not possible in Evert's study. We wanted to determine when if at all the difference between infected and un-infected people was established. Therefore, we analysed DC cell populations in younger age groups. Our study demonstrates that a lower frequency of mDCs in infected people can be already observed in 10–13 year olds. This is accordance with changes in other immunological parameters as clearly shown for schistosome-specific IgG, IgE and IgM. The importance of DC in TH2 induction has been recently shown [23]. By altering cytokine responses DC could finally be involved in isotype switching and development of protective immunity. This is supported by our finding of the correlation between schisotosome-specific IgG and mDCs.
Changes with age have been also observed for serum cytokines in other study populations with moderate/high schistosome transmission [10], [11], [64]. For instance, high levels of parasite-specific or serum levels of IL-10 have been reported already at this age in different populations [11], [65] and might contribute to such changes. For example, DCs treated with the regulatory cytokine IL-10 inhibit T cell activation [66]–[69]. Earlier studies showed that T cell hypo-responsiveness can be already observed at early ages [70], [71]. The role of mDCs in induction of this impairment remains to be further investigated.
Other reasons for reduced relative numbers of mDCs in infected individuals in older age groups remain to be determined, e.g. enhanced apoptosis, reduced bone marrow output or increased migration to inflamed tissues or lymphoid organs were proposed by Everts et al [32]. Indeed enhanced migration in the context of TH2 priming DCs has been reported [25], [72]. This was induced by eosinophil-derived neurotoxin; a molecule elevated in serum of S. haematobium infected people [73]. Everts et al reported a more general impairment of DCs affecting both TH1 and TH2 cytokines as well as IL-10 in older individuals (17–39 years). The mechanisms involved are still unclear since despite this impairment of DC function, older individuals carrying schistosome infection can show enhanced immunopathology [74], [75] indicating an effective immune response.
In contrast, our data show that in the younger individuals (5–9 years) mDCs proportions were higher in peripheral blood of infected individuals in comparison to un-infected. Potential reasons for this include enhanced bone marrow output or reduced migration of DCs in the context of blood-born antigens [76]. Infection levels were highest in the youngest age group of our study population and young people might have more blood migrating schistosome stages affecting proportions of circulating mDCs. In contrast enhanced immune response including enhanced migration of mDCs might lead to increased killing of infiltrating schistosomula in older age groups.
Dendritic cell activation/maturation is characterised by up-regulation of cell-surface markers such as CD86 and MHC class II molecules. As previously reported by Everts et al [32], we observed no changes in expression of CD86 in infected people. However HLA-DR levels differed. In the youngest age group (5–9 years), where we found higher percentages of mDCs, expression levels of HLA-DR were lower in infected people. Reduced HLA-DR expression might be an indication of less matured and possibly fewer migrating cells, which subsequently could result in a less pronounced induction of T cell response in an early phase of exposure to schistosomes accompanied by a more regulatory environment due to Tregs and parasite-specific IL-10 [11], [12], [65]. Analysis of Tregs from another study cohort from the same population revealed a comparable pattern as recently published by our group [12]. In addition a direct comparison of mDCs and Tregs would need a larger sample size, since statistical power is lost by comparing the two parameters in the presence of other potentially confounding variables such as host age.
In contrast to the study by Everts et al, we did not observe a significant difference of HLA-DR levels in the older age groups. Again more significant changes of HLA-DR in these age groups might be due to cumulative effect of co-infections, not present in our study population.
Inhibition of up-regulation of activation markers such as CD86 has been suggested to be involved in induction of tolerogenic DCs [66]. However we did not see changes in the activation marker CD86. There is need to analyse other markers, such as B7-H1, B7-H2, lg-like transcripts 3 and 4 which might play a role in the function of tolerogenic DCs [77]–[81], particularly in younger age groups.
Functional studies, will be very informative and will clarify some of these areas, but were beyond the scope of this study.
In summary this study clearly shows there is an age-related pattern in the proportions mDCs in people exposed to S. haematobium infection as is typical for many already investigated humoral and cellular responses. Furthermore, in the case of mDCs, this pattern differs between schistosome infected and un-infected people and also follows the age profile with infection of protective responses, which is supported by the correlation between schistosome-specific IgG and proportions of mDCs.
The mechanism and consequences (such as alterations in the potential to induce different T cell phenotypes) behind the association between mDCs and infection remain to be determined. Understanding the nature and dynamics of immunological parameters in natural helminth infections and, in particular, changes of DCs involved in polarising the immune responses has an important impact on the development of vaccines against schistosomes.
|
10.1371/journal.pcbi.1005942 | What drives the perceptual change resulting from speech motor adaptation? Evaluation of hypotheses in a Bayesian modeling framework | Shifts in perceptual boundaries resulting from speech motor learning induced by perturbations of the auditory feedback were taken as evidence for the involvement of motor functions in auditory speech perception. Beyond this general statement, the precise mechanisms underlying this involvement are not yet fully understood. In this paper we propose a quantitative evaluation of some hypotheses concerning the motor and auditory updates that could result from motor learning, in the context of various assumptions about the roles of the auditory and somatosensory pathways in speech perception. This analysis was made possible thanks to the use of a Bayesian model that implements these hypotheses by expressing the relationships between speech production and speech perception in a joint probability distribution. The evaluation focuses on how the hypotheses can (1) predict the location of perceptual boundary shifts once the perturbation has been removed, (2) account for the magnitude of the compensation in presence of the perturbation, and (3) describe the correlation between these two behavioral characteristics. Experimental findings about changes in speech perception following adaptation to auditory feedback perturbations serve as reference. Simulations suggest that they are compatible with a framework in which motor adaptation updates both the auditory-motor internal model and the auditory characterization of the perturbed phoneme, and where perception involves both auditory and somatosensory pathways.
| Experimental evidence suggest that motor learning influences categories in speech perception. These observations are consistent with studies of arm motor control showing that motor learning alters the perception of the arm location in the space, and that these perceptual changes are associated with increased connectivity between regions of the motor cortex. Still, the interpretation of experimental findings is severely handicapped by a lack of precise hypotheses about underlying mechanisms. We reanalyze the results of the most advanced experimental studies of this kind in speech, in light of a systematic and computational evaluation of hypotheses concerning motor and auditory updates that could result from motor learning. To do so, we mathematically translate these hypotheses into a unified Bayesian model that integrates for the first time speech production and speech perception in a coherent architecture. We show that experimental findings are best accounted for when motor learning is assumed to generate updates of the auditory-motor internal model and the auditory characterization of phonemes, and when perception is assumed to involve both auditory and somatosensory pathways. This strongly reinforces the view that auditory and motor knowledge intervene in speech perception, and suggests likely mechanisms for motor learning in speech production.
| The fact that perception has an influence on motor learning is known and has been the focus of a large number of studies. The converse, i.e. that motor learning would influence perception, seems more intriguing and unclear. For speech, shifts in perceptual boundaries have been shown to result from motor learning induced by perturbations of the auditory feedback [1, 2] or perturbations of the articulatory gestures [3]. In the context of the well-known historical debates about the primitives (auditory/articulatory/motor) of speech perception [4–8], these findings could be interpreted as evidence in support of theories assuming the involvement of speech production processes in speech perception. However, an influence of speech motor learning on perceptual categorization of speech sounds does not necessarily imply an involvement of brain motor areas in speech perception. Indeed, the unusual auditory signals experienced during the adaptation process may by themselves be responsible for the observed perceptual shift.
From this observation, and building up on Shiller et al.’s experiment [2], Lametti et al. [1] specifically attempted to disentangle the respective influence of motor functions and altered sensory inputs on the perceptual boundary shifts. To do so, they developed an experimental protocol designed to assess separately the learning effects induced by changes in auditory feedback, on the one hand, and those arising from changes in motor control, on the other hand. They concluded that the origin of the perceptual change is indeed motor rather than sensory.
Lametti et al.’s study is very rich and relies on a solid experimental methodology. However we argue that their reasoning, because it is only qualitative, is incomplete, and does not enable to fully understand the nature of the mechanisms underlying the link observed after motor learning between changes in motor functions and perceptual changes.
In the present work we propose to dig into these questions using a previously defined Bayesian model [9]. This model was previously used to study the relative roles of auditory and proprioceptive representations in speech gesture planning; here we adapt this model to identify, implement and compare different hypotheses concerning motor adaptation. We analyze the consequences of these different hypotheses on perception and production mechanisms and suggest additional tentative interpretations of the experimental findings reported by Lametti et al. [1]. This constitutes, in our view, an important step to better relate experimental data to theories of speech production and speech perception, and further enlighten the possible role of motor processes in speech perception. Importantly, the Bayesian model we use enables to translate classical and transversal questions about motor control, perception, learning and adaptation into computations and predictions. Such a model is a methodological tool to tackle these issues widely in speech production and speech perception, as well as in arm motor control [10, 11].
The body of this paper is divided into four sections. The remaining of this section gives an overview of the main experimental paradigms and facts reported by Lametti et al. [1]. We then present our modeling framework to deal with these experimental findings; this is presented in Section “Model”. The interpretation of the results of simulations are presented in Section “Results”, and discussed in Section “Discussion”.
The influence of speech motor learning on speech perception was first reported by Shiller et al. [2] (this study is called “S-09” henceforth). Motor learning was implemented by perturbing the auditory feedback of subjects when they were producing the fricative /s/: it consisted in shifting down the first spectral moment of /s/ in such a way that it sounded more like /∫/. They observed that subjects adapted their articulation after training in order to compensate, partially, for the perturbation, and the perceptual test after adaptation revealed a shift of the perceptual boundary between /s/ and /∫/ toward /∫/ (more sounds were perceived like /s/).
Five years later, Lametti et al. [1] published a new study (referred to as “L-14” henceforth) aiming at clarifying whether the observed perceptual change was related to “the change to motor function that occurs during learning, [to the] perceptual learning related to the altered sensory inputs, [or to] some combination of the two”(p 10339). To this end they proposed an original experimental design supposed to disentangle the effects of sensory vs. motor processes on perceptual categorization. While in S-09 a perturbation of the fricative /s/ was introduced in only one direction (toward the fricative /∫/), in L-14 the vowel /ɛ/ was perturbed in two directions. For one group of subjects, the perturbation was applied toward the vowel /a/ by increasing the frequency of the first formant F1 (left panel in Fig 1). For the other group it was applied toward the vowel /i/ by decreasing F1 (right panel in Fig 1).
To make the reasoning in L-14 clear, let us analyze the case of the perturbation toward /a/ (see Fig 1, left panel). The shift of the auditory percept along the /ɛ-a/ continuum generated a compensatory movement of the tongue frontwards, which corresponds in the absence of perturbation to an auditory percept along the /ɛ-i/ continuum. Since compensation is never complete, it results with altered auditory feedback in /ɛ/ sounds that remain partly perturbed and belong to the /ɛ-a/ region, while speaker’s gestures and their corresponding somatosensory information actually belong to the /ɛ-i/ region. This is the clever method used by the authors to attempt to disentangle auditory and motor interpretations of the perturbation effects. Indeed, in their reasoning, measuring the shift of the perceptual boundary between /ɛ/ and /a/ provides a measure of the effect of the altered sensory inputs on perceptual categories, while measuring the shift of the perceptual boundary between /ɛ/ and /i/ provides a measure of the effects of the changed articulation, i.e. of the motor function, on perceptual categories. A symmetric reasoning applies for the perturbation toward /i/ (Fig 1, right panel).
Concerning motor learning, consistent with S-09 and other auditory perturbation studies in speech, motor compensation was observed and its magnitude was on average below 40% of the amplitude of perturbation. Concerning perception, a significant boundary shift was also observed in L-14. Consistent with observations reported in S-09, the resulting perceptual shifts were in the same direction as the perturbation. However, contrary to S-09, no significant shift was observed in L-14 in the region of the altered auditory inputs (i.e. the /ɛ-a/ continuum for a perturbation towards /a/); the significant shift was found in the region corresponding to the altered articulation (i.e. the /ɛ-i/ continuum for a perturbation toward /a/, see Fig 1, left panel). A control group in which subjects produced the same sequence of sounds without alteration of the auditory feedback did not show any perceptual boundary shift.
The authors concluded that their findings are “consistent with the idea that changes to central motor commands associated with speech learning are the source of changes observed in the perceptual classification of speech sounds” [1, p 10340]. Notice that if it is true that the origin of the observed perceptual shift is due to motor functions, greater changes in motor functions should induce greater changes in perception, inducing after learning positive correlations between the amount of compensation and the amplitude of the resulting perceptual shift. Intriguingly, an absence of significant correlation was reported in L-14.
Our aim is to exploit a previously defined computational framework [12–14] modelling the interactions of perception and production in speech communication, and to apply it to model and better understand the experimental data of L-14. In our modeling approach our prime concern is to extract the deeper meaning of the experimental observations and to specify a limited number of facts that best characterize them. The following summary presents the main experimental facts on which we will focus in our modeling work.
This section introduces our model, which is an instance of the Bayesian algorithmic modeling framework [15], that is, the application of Bayesian Programming [16] to Marr’s algorithmic level of cognitive modeling [17]. With this framework, we have previously developed a series of models, under the COSMO moniker, to study speech perception and speech production in different contexts, such as speech communication and the emergence of phonological systems [13], speech perception in adverse conditions [12, 14], sensorimotor learning [18] and the emergence of speech idiosyncrasies [19]. Variants have also been applied, in speech production, to token-to-token variability [20], the incorporation of multiple constraints in speech planning [21] and the modeling of multisensory (acoustic and somatosensory) speech targets [9]. It is this last variant that we adapt here to our current study.
In the Bayesian algorithmic modeling approach, an overarching feature is that perception and production processes are not directly modeled. Instead, we build an undirected model of speech-relevant knowledge using probability distributions. Then, from this model, we compute distributions using Bayesian inference to simulate perception and production tasks. Perception and production processes, therefore, if they involve the same knowledge, become related. Let us consider the case of speech: in our approach, we commonly assume that the description of acoustic targets in speech planning is the same piece of knowledge as would be used in a purely auditory decoder in speech perception. This distinction between the knowledge stored in the model and its use to generate processes makes our framework ideal for the study of the links between production and perception mechanisms, such as those addressed in this work.
The model includes selected aspects of speech production and speech perception that are described in Section “Selected aspects for modeling”. Their implementation in the model is explained in Sections “Model definition” and “Formulation of speech production and perception questions”. The strategy used to simulate the experimental paradigm of L-14 is detailed in Section “Implementation of the experimental paradigm: Normal vs. adapted conditions”. Finally, the simulation results and their analysis are presented in Section “Results”.
Our aim is to study the interaction between speech production and speech perception processes in light of the experimental results provided in L-14. The first step in such a modeling approach consists in reducing the complexity of the experimental world into a core set of simplified components likely to capture its essential ingredients. This simplification phase should result in constraining and focusing both model implementation and results interpretation. We have selected a reduced number of aspects in speech production and speech perception that we consider to be crucial and sufficiently representative for the investigation of the interaction between motor learning and perception of isolated phonemes—here, isolated vowels /i/, /ɛ/ and /a/.
The structure of the model consists in implementing a chain of probabilistic dependencies between phonological, motor and sensory variables. Variables and their dependencies are illustrated in Fig 2, and we now describe the most salient aspects of the model (a more complete mathematical description is provided in Supporting information S1 Text.
In the previous section we proposed a computational definition of the joint probability distribution of the model. This definition was based on particular assumptions concerning relations between variables. The Bayesian formalism allows to simulate speech production and perception by defining and computing probability distributions of interest, that we call “questions”.
Our aim is to simulate and compare the outcome of the production and perception tests in L-14, prior to the auditory perturbation and after the training phase, i.e. when perturbation is removed and adaptation has been reached. These tests are naturally implemented in the model as the outcome of the production and perception questions defined in the previous section.
Adaptation is implemented as the update of a part of the knowledge included in the model. This knowledge is represented by the four relations defined in Section “Parametric forms”: the two sensory-motor internal models, P(AM | M) and P(SM | M), and the two sensory characterizations of phonemes, P(AΦ | Φ) and P(SΦ | Φ). In this context, normal and adapted conditions are implemented by different values of the parameters characterizing these relations. Values of parameters in normal condition are arbitrary initial values. This is why we chose them to be as simple as possible. They are specified in Section “Normal condition: Initial values of parameters”.
Two fundamental questions remain to be answered in order to specify how adaptation will affect these initial values: (1) which of the four relations is changed during adaptation, and (2) how? The first question actually rephrases in computational terms the question raised in L-14 (p 10339), and quoted in its original formulation in Section “Influence of motor learning upon speech perception: overview of experimental facts”, extending it to behavioral changes in both production and perception: “So what produces the [behavioral changes] during motor learning? Is it the change to [parameters of the sensory-motor internal models], that occurs during learning? Is it changes to [parameters of the sensory characterizations of phonemes], related to the altered sensory inputs? Or is it some combination of the two?”.
In the following sections, we address these two questions in two steps. In Section “Adaptation hypotheses” we partially answer the first question by motivating the selection of a subset of possible changes induced by adaptation. In Section “Results” we further answer these questions by evaluating the outcome of different implementations of the selected changes and by comparing them with the experimental facts summarized in Section “Summary of experimental results we aim at modeling”.
The primary goal of this section is to evaluate which of the 7 adaptation hypotheses account for the experimental facts reported in L-14. To do so, we proceed sequentially: we first focus on perception and evaluate results corresponding to the two categorization questions Q Per A and Q Per F. For the hypotheses that are compatible with the perceptual boundary shift observed in L-14, the associated compensation in production is evaluated, and again only the hypotheses that are compatible with the results of L-14 are kept. Finally, in a third step, we further evaluate the selected adaptation hypotheses with respect to the corresponding correlations between the amount of compensation in production and the magnitude of perceptual boundary shift.
Let us now evaluate the effect of the three previous adaptation hypotheses, H Ad M, H Ad Φ and H Ad M Φ with respect to the production question Q Prod F.
These three combined hypotheses are equivalent in terms of the qualitative effects predicted with respect to changes in production and perception; they all account for incomplete compensation and for the asymmetric perceptual boundary shift in the direction of perturbation. However, the magnitudes of the perceptual boundary shift and of the motor command shift associated with the compensation differ across the three hypotheses. Experimental studies display large differences across subjects in their capacity to compensate for a perturbation of the auditory feedback [56–58]. Moreover, in L-14 and S-09 subjects differ in the amount of perceptual boundary shift induced by adaptation to the perturbation. If, as suggested in L-14, the perceptual change is mainly due to a change in motor functions, one would expect that subjects who compensate more would exhibit a greater perceptual boundary shift. However, no significant correlation between these two phenomena was found in L-14.
In the present section we focus on this question. First, we identify possible origins for the reported differences concerning the amount of compensation and perceptual shift among subjects. Then, we implement these origins under each of the three combined hypotheses and evaluate their predictions in terms of the correlations between compensation magnitudes and amount of perceptual boundary shift.
Using our model, implemented in the Bayesian programming framework, we have been able to implement and test different hypotheses concerning speech motor adaptation to perturbed auditory feedback. In this framework, processes are not directly modeled but are derived from a common set of knowledge, which is represented by means of a joint probability distribution. Hence, in this approach, perception and production processes become naturally related since changes to the underlying knowledge may impact them together. Note that this framework is not restricted to speech, but may be of interest in other areas where production and perception processes have been shown to interact (for instance in the arm motor control literature, see Haith et al. [62] and Ito et al. [63] for alternative approaches, see also Gilet et al. [36] in the context of joint modeling of perception and production of isolated cursive letters).
We have applied this framework to study the perceptual changes that result from motor learning in adaptation to an auditory perturbation in speech. To do so, we have proposed a number of hypotheses about the changes to the common underlying knowledge that may result from motor learning and we have investigated how these changes may give rise to the observed changes in perception and production. This approach has allowed us to identify different possible origins that all may contribute to these changes, supporting but also specifying the interpretation proposed by Lametti et al. [1].
Our experimental simulations provide a number of major results: (1) the induced perceptual shift may actually be compatible with either an auditory or a combined auditory and somatosensory characterization of perceptual targets; (2) the incomplete motor response to auditory perturbations may be due to a mixture of components, related to the combined specification of the phonemic targets for speech production in auditory and somatosensory terms; (3) the asymmetry in perceptual compensation observed in L-14 is also compatible with both theoretical frameworks in speech perception, but actually appears to be sensitive to fine tuning of the experimental parameters in the simulations; (4) patterns of correlations between perceptual and motor responses may be driven by various factors that shed a crucial light on final interpretations of the experimental data.
Of course, these simulations quantitatively depend on a number of modeling choices introduced in Section “Selected aspects for modeling”, that are aimed at making simulations tractable and easy to analyze and interpret. This basically includes: (1) the assumption that sensory and motor spaces are one-dimensional, (2) the assumption that sensory-motor mappings are linear (Eqs (2–5, 15 and 16), and (3) the specific tuning of parameters considered in the update hypotheses for adaptation. Still, it is important to stress that the four major results summarized previously have an intrinsic validity, which makes them largely independent of the specific modeling choices. This is due to two major reasons. Firstly, the modeling framework introduced in this work has actually been developed over the years completely independently of the experimental data discussed here. This framework is essentially conceived as a general architecture for formalizing classical assumptions about perceptuo-motor relationships in speech communication [12–14].
Secondly, the four major results appear as general, and likely to be obtained whatever the specific choices in the model. Indeed, the first, second and fourth of these results express direct consequences of the model architecture, in which multisensory fusion (between auditory and somatosensory representations) in speech production and possibly in speech perception naturally result in trading relationships leading to (1) perceptual adaptation in response to the motor adaptation (2) incomplete response to perturbation and (3) various types of correlation patterns between motor and perceptual adaptation. The case of the third result (asymmetry in perceptual compensation) is quite interesting in this respect. Indeed, it is, contrary to the others, largely ad hoc and related to the specific modeling choices (i.e., the precise relation between mean and variance and parameters of the local update of the internal model, see Supporting information S3 and S4 Text). This makes it fragile and probably not very robust experimentally. But this fragility can also be construed as a prediction: it means that asymmetries should vary from one study to the other, and that this observation is probably not as reliable as what was expected by the authors of L-14 (see for instance a recent study by Schuerman et al. [64] where no significant boundary shift was obtained).
Interestingly, the symmetric vs. asymmetric nature of the perceptuo-motor adaptation process should also largely depend on the nature of the motor-to-sensory internal model, and it is quite well-known that the motor-to-sensory relationship is indeed highly nonlinear, and likely to vary greatly depending on the involved region of the motor or sensory space. This could well explain the difference between the study by Lametti et al. [1] on vowels, that shows a lack of perceptual shift in the region of the auditory space related to what subjects heard in presence of the perturbation, and the study by Shiller et al. [2] in which a perceptual shift in the corresponding regions with fricatives was observed.
Finally, with respect to the one-dimensional assumption, including additional dimensions in sensory and motor spaces may certainly bring interesting behaviors, such as trading relations between dimensions in compensation. However, the /i ɛ a/ continuum considered in L-14 can be basically seen as one-dimensional both in the articulatory space in which the location of the highest point of the tongue is controlled along the high/front—low/back dimension thanks to strong correlations between jaw opening and tongue position [65], —and in the acoustic space with correlated variations between F1 and F2 respectively increasing and decreasing from /i/ to /a/ [66]. Therefore, such additional effects would likely bring only a modulatory change to the magnitude of the resulting shifts in production and perception, without changing the general patterns of results in our simulation.
Therefore, we consider that the simulation results presented here have intrinsic validity. As a consequence, it is of interest to discuss them as some new evidence that can be confronted to important questions related to perceptuo-motor adaptation as discussed in the literature. This is what we will do now, around two points that are the nature of perceptual representations and the origins of incomplete compensation, before introducing some predictions and proposals for new experiments in the field.
The first stage of our simulations (Section “Evaluation with respect to perception”) both supports and challenges the interpretation by Lametti et al. [1], whereby their data would provide evidence for the role of motor knowledge in speech perception. On the one hand, hypothesis Q Per F ⊕ H Ad M, involving only an update of motor functions, is compatible with their interpretation and in fact also specifies it. Indeed, under this hypothesis a local compensation for the perturbation is required to generate a pattern of perceptual adaptation fitting the asymmetry reported in L-14. On the other hand, in the context of hypothesis Q Per A ⊕ H Ad M Φ, involving both a local update of the auditory-motor internal model and a modification of the auditory characterization of the perturbed phoneme, a pure auditory theory of speech perception (Q Per A) also provides a pattern of perceptual shifts compatible with their data, even including asymmetries that were considered as key in their reasoning against auditory theories. In this case, changes in the auditory characterization of a phoneme, involving a coordinated shift of the center of its characterization and a reduction of its variance, are required to explain their results.
It is important to note that it is not unrealistic to assume that motor learning can induce such coordinated changes. Indeed, the shift in location may be explained by a mechanism aligning the auditory characterization of a vowel with its actual realization in presence of the auditory feedback perturbation. The reduction of variance could be attributed to the well-known selective adaptation phenomenon, as suggested by Kleinschmidt et al. [67]: the repeated exposure to the same sound tends to make listeners more sensitive to variations of this sound. Note that, in S-09, selective adaptation was mentioned in order to explain the small perceptual boundary shift observed in their control group after the repeated exposure to the unaltered fricative /s/.
Therefore, at this stage, both an audio-motor and a pure auditory theory may be compatible with the data in L-14. However, the analysis, based on correlations between the amplitude of the perceptual shift and the magnitude of the compensation, indicates that none of the two previous interpretations is compatible with the observations described in L-14. Only hypothesis Q Per F ⊕ H Ad M Φ, assuming the fusion of sensory pathways in speech perception and adaptation involving the combined updates of the auditory-motor internal model and the auditory characterization of the perturbed phoneme, was compatible with the absence of significant correlation reported in L-14.
In summary, our results support and clarify the initial interpretation of Lametti et al. [1]. By exploiting perceptuo-motor correlations, our results support the claim that both sensory and motor processes intervene in the observed perceptual shift. This result certainly speaks in favor of perceptuo-motor theories of speech perception, though further work should be done in order to better assess the relative contributions of each of these two sets of processes [14].
Interestingly, in our model, all possible explanations of the link between motor learning and perceptual boundary shift are associated with incomplete compensation for the perturbation, even if the magnitude of the local update of the auditory-motor internal model fully matches the amplitude of the auditory perturbation. This is an important prediction of our model, since incomplete compensations have been systematically observed in all experiments involving a perturbation of the auditory feedback during speech production.
Three mechanisms can indeed be at the origin of incomplete compensation. Firstly, if motor learning induces only an update of the auditory-motor internal model in the context of a bi-modal speech production process, incomplete compensation comes from the interaction between the somatosensory and the auditory specifications of vowels. Secondly, if motor learning also induces a shift and a reduction of variance of the auditory specification of the perturbed phoneme, this provides an additional counter-influence to compensation and the magnitude of the change of the auditory characterization contributes to incomplete compensation. Thirdly, in all cases, if motor learning induces an update of the auditory-motor internal model, the magnitude of this update influences the extent of the compensation: the smaller the update, the more incomplete the compensation.
All these potential explanations of incomplete compensation for perturbations of the auditory feedback have been previously suggested in the literature. In particular, Katseff et al. [68], among other hypotheses, compared the respective influences on the compensation magnitude of a possible interaction between the auditory and the somatosensory feedback versus of a possible shift of the auditory region characterizing the pronounced phoneme. They concluded that behavioral data about compensation for auditory perturbation published in the literature (including those in S-09) are more compatible with an interaction between the two sensory feedbacks.
According to them, in the case of the data in S-09, if the perceptual boundary shift is due to a shift of the auditory characterization of the perturbed phoneme, this latter shift should have the same small amplitude as the former one. Such a small shift of the auditory characterization of the phoneme could not explain the large magnitude of the reduction in compensation.
Our results allow us to qualify their conclusion. Indeed, we have shown that when the shift of the auditory characterization is associated with a reduction of its variance, the magnitude of this shift can be much larger than the magnitude of the perceptual boundary shift. In this case, the shift of the auditory characterization of the perturbed phoneme would perfectly account for the amplitude of the compensation.
At this stage, we have at our disposal a modeling framework to account for the links between production and perception processes. However, the present work focuses on adaptation, by comparing states before and after learning. Investigating the dynamic process occurring during adaptation could provide interesting further insights into the phenomena associated with adaptation. More specifically, the manner with which compensation strategies integrate sensory feedback would inform about the way the sensory-motor characteristics of speech production are updated during the learning phase. For instance, the completeness of compensation appears to be dependent on the amplitude of the perturbation: greater amplitudes of perturbation induce greater sensory errors which appear to result in smaller percentage of total compensation compared to smaller sensory errors. This result seems to be a general property of sensorimotor learning: indeed it has been reported for speech [53, 68], for eye and arm movements [69, 70] and even for bird song [71]. Still, the mechanisms responsible for this decrease in relative adaptation in the case of increasing sensory errors remain unclear.
Our model, in its current state, does not address this question, since it deals only with the consequences of parameters updates, and not with how these updates happen during the learning phase. However, the three possible origins of incomplete compensation (discussed in Section “Three suggested origins for incomplete compensation”) actually suggest three possible mechanisms whereby different magnitudes of sensory error would result in different degrees of compensation completeness. First, at the level of the sensory motor mappings, larger sensory errors may drive slower update in order to avoid a faulty reorganization of the learned mapping in the case of totally unexpected and inappropriate sensory signals (see for instance the work of [72] for a modeling approach in line with this idea). Second, at the level of the relative weighting of sensory pathways, the magnitude of sensory errors could disadvantage the pathway with larger errors, assuming that large unexpected errors would arise from inaccurate sensors, which would then be considered unreliable. Finally, at the level of the sensory characterization of the target, larger sensory perturbations may drive larger shifts of the intended target, resulting in smaller amounts of compensation compared to baseline. Each of these hypotheses deserves more careful analysis in light of the existing experimental data: for example, the third hypothesis appears unlikely, since, after the removal of the perturbation, subjects usually return close to the original baseline. Still, these three hypotheses definitely deserve further experimental focus.
Interestingly, our model gives different predictions for these three hypotheses. For instance, if larger sensory errors disadvantage the weighting of one of the sensory pathways, the model would predict that subjects would begin to compensate more for perturbations in the other sensory modality. Such sensory preferences have been reported previously in speech production [59]; however, to our knowledge, no study has explored the possibility that these preferences may be experimentally modulated by providing larger perturbations to one of the sensory modalities. On the other hand, if sensory errors only influence the update of the sensory-motor mapping or the shift of the sensory characterization of the target, the model would predict no influence of the amount of compensation to perturbations on the other sensory modality. Furthermore, evaluating the influence of the amplitude of perturbation with respect to the resulting perceptual shift could also allow distinguishing between these last two hypotheses. Indeed, if larger sensory errors decrease the update of the sensory-motor mapping, the model would predict a decrease in the amount of perceptual shift, whereas the contrary would happen if larger sensory errors drive greater shifts in the sensory characterization of the target.
Furthermore, as we suggested above, the present model is not limited to the study of auditory perturbations, and investigating the consequences of somatosensory perturbations would allow further evaluation of its pertinence. Indeed, another interesting prediction of the model is that, if adaptation to a somatosensory perturbation updates the somatosensory-motor mapping, it would also induce a boundary shift in the auditory categorization of the perturbed phoneme (but in an opposite direction to perturbation, contrary to the case of auditory perturbations). Such perceptual change following adaptation to a somatosensory perturbation has been actually reported in speech by Nasir and Ostry [3]. Future development of the model would be needed to account for their results, since Nasir and Ostry’s paradigm uses a perturbation of the jaw along the horizontal direction, making thus possible a perturbation of the somatosensory feedback without inducing changes in the auditory domain.
More generally, the present model provides a powerful framework for testing hypotheses on the relative roles of auditory and somatosensory representations and processes in perceptual and motor responses to perturbations. Indeed, any means likely to modulate one or the other input (e.g., by exploiting inter-individual variability—or by decreasing the salience of one modality relative to the other, by various techniques such as masking or inhibition of a given channel) should modify the amount of response to perturbations, and thus generate specific quantitative predictions to be compared with new experimental data (e.g., [73]).
Finally, it could be interesting to relate our computational framework with putative neuroanatomical networks suggested by neurocognitive data from the literature. As a matter of fact, a number of studies have explored the neuroanatomy of circuits in charge of monitoring responses to auditory or somatosensory perturbations in speech production (e.g., [74–81]). Even though this is out of the focus of the present study, we have already undertaken studies suggesting possible neuroanatomical correlates of the generic COSMO model [82], which is compatible with the current computational model. A future step in this direction is to adapt the generic architecture to the specific processes associated to perturbation compensation. This would be necessary for better addressing the dynamic adaptation processes mentioned previously in this section.
In order to better understand the mechanisms underlying the observations reported by Lametti et al. [1], we have elaborated a simplified Bayesian model of speech production and speech perception in which phonemes are characterized both in somatosensory and auditory terms. Speech production is assumed to be guided by both sensory characterizations (hypothesis Q Prod F). Two hypotheses concerning speech perception processes were evaluated: (1) speech perception relies only on the auditory pathway (hypothesis Q Per A), or (2) speech perception relies on the fusion of both auditory and somatosensory pathways (hypothesis Q Per F). We have also considered different hypotheses on the possible consequences of motor adaptation: (1) an update of the auditory-motor internal model, (2) an update of the auditory characterization of the perturbed phoneme, and (3) an update of its somatosensory characterization. Taken separately or in combination, these three update hypotheses lead to seven possible adaptation hypotheses. Combined with the two perception hypotheses Q Per A and Q Per F, these adaptation hypotheses lead to different possible scenarios for explaining the observations of the study of Lametti et al. [1].
In the context of our Bayesian model, we have compared the predictions of these possible scenarios with the experimental observations reported by Lametti et al. [1]. Considering results in perception and production, our simulations indicate that three combined perception-adaptation hypotheses can reproduce the characteristics of the perceptual boundary shift observed in L-14: (1) speech perception relies both on the somatosensory and auditory pathways, and motor adaptation induces only a local update of the auditory-motor internal model (Q Per F ⊕ H Ad M); (2) speech perception relies only on the auditory pathway and motor adaptation induces both a local update of the auditory-motor internal model and the combined shift and size reduction of the auditory characterization of the perturbed phoneme (Q Per A ⊕ H Ad M Φ), (3) speech perception relies both on the somatosensory and auditory pathways and motor adaptation induces both a local update of the auditory-motor internal model and the combined shift and size reduction of the perturbed phoneme (Q Per F ⊕ H Ad M Φ).
From that basis, these three selected hypotheses were further evaluated with respect to the predicted correlation between compensation in production and perceptual shift. Our results indicate that only the third hypothesis (Q Per F ⊕ H Ad M Φ) is able to account for the absence of correlation reported by Lametti et al. [1].
Altogether, this computational approach strengthens and specifies the interpretation by Lametti et al. [1] of their experimental data in favor of perceptuo-motor links in speech perception. Our model provides novel insights into the mechanisms influencing speech perception and production after adaptation to perturbations of the auditory feedback. Future work should focus on the dynamics of adaptation as well as on the relation between the degree of adaptation and the amount of perceptual changes.
|
10.1371/journal.pntd.0006749 | Community-based prevalence of typhoid fever, typhus, brucellosis and malaria among symptomatic individuals in Afar Region, Ethiopia | In sub-Saharan Africa, where there is the scarcity of proper diagnostic tools, febrile illness related symptoms are often misdiagnosed as malaria. Information on causative agents of febrile illness related symptoms among pastoral communities in Ethiopia have rarely been described.
In this a community based cross-sectional survey, we assessed the prevalence of typhoid fever, typhus, brucellosis and malaria among individuals with a set of given symptoms in Amibara district, Afar Region, Ethiopia. Blood samples were collected from 650 study participants, and examined by Widal and Weilfelix direct card agglutination test (DCAT) as well as test tube based titration test for Salmonella enterica serotype Typhi (S. Typhi) and Rickettsia infections. Rose Bengal Plate Test (RBPT) and Complement Fixation Test (CFT) were used to screen Brucella infection. Thin and thick blood smears were used to diagnosis malaria.
Out of 630 sera screened by DCAT, 83 (13.2%) were reactive to H and/or O antigens for S. Typhi infection. Among these, 46 (55.4%) were reactive by the titration test at the cut off value ≥ 1:80. The combined sero-prevalence for S. Typhi by the two tests was 7.3% (46/630). The seroprevalence for Rickettsia infection was 26.2% (165/630) by DCAT and 53.3% (88/165) by the titration test at the cut off value ≥ 1:80. The combined sero-prevalence for Rickettsia infection by the two tests was 14.0% (88/630). The sero-prevalence for Brucella infection was 12.7% (80/630) by RBPT, of which 28/80 (35%) were positive by CFT. The combined sero-prevalence for Brucella infection by the two tests was 4.4% (28/630). Out 650 suspected individuals for malaria, 16 (2.5%) were found positive for P. falciparum infection.
In this study, typhoid fever, typhus, brucellosis and malaria were observed among symptomatic individuals. The study also highlighted that brucellosis cases can be misdiagnosed as malaria or other disease based solely on clinical diagnosis. Therefore, efforts are needed to improve disease awareness and laboratory services for the diagnosis of brucellosis and other zoonotic diseases to identify other causes of febrile illness in this pastoral setting.
| Many diseases such as typhoid fever, typhus, brucellosis and malaria show common symptoms such as fever, headache, joint pain and back pain. Hence, in countries where there is a problem of appropriate laboratory based diagnostic tools, health workers cannot properly diagnose these diseases and provide appropriate treatment. A community- based studies of the causative agents of the above mentioned illness would provide important information for health workers about some of the common causative agents in that particular area. In this study, we assessed the prevalence of typhoid fever, typhus, brucellosis and malaria among individuals who were complaining illnesses such fever, headache, joint pain and back pain in the pastoral of the Amibara district, Afar Region, Ethiopia. Among 650 individuals who were complaining symptoms, 46 (7.3%), 88 (14.0%), 28 (4.4%) and 16 (2.5%) were diagnosed for typhoid fever, typhus, brucellosis and malaria in that order. However, for the majority of the participants (75.4%), the cause of their illness remained unknown, and further investigations on the causative agents of febrile illness related symptoms is important in the present study area.
| Sub-Saharan Africa is plagued by a myriad of infectious diseases posing significant public health and economic challenges. In addition, the often non-specific clinical signs of these diseases and the scarcity of proper diagnostic tools are the major challenges for health professionals in properly diagnosing and treating adequately patients [1, 2]. Studies showed that symptoms such as fever, headache, joint pain and back pain are often misdiagnosed as malaria, especially until the introduction of rapid diagnostics for malaria though these symptoms are not only specific to malaria [3–5].
Many studies have shown that diseases such as typhoid fever, rickettsioses, brucellosis, Q fever, and leptospirosis are the leading causes for febrile illness with symptoms such as fever, headache, joint pain and back pain [6–9]. For instance, typhoid fever due to S. Typhi has been reported as the leading cause of over 21 million febrile cases and over 200,000 deaths each year in many low- and middle-income countries [7–9].
Brucellosis has been considered as an important zoonotic disease worldwide and is responsible for big economic losses as it causes abortion in livestock [10, 11]. It also causes a considerable human morbidity and spontaneous abortion among pregnant women in endemic areas [12–15].
Although many malarious countries including Ethiopia are scaling up malaria intervention programs towards elimination, the disease remains one of the worst health problems with an estimated 216 million cases and 445 000 deaths globally in 2016, while most of the cases and deaths occurred in African [16]. Hence, in low- and middle-income countries where there is a shortage of effective routine diagnostic tools to identify a wide range of infectious diseases that manifest similar symptoms and where there is also a low awareness among community members and health professionals about the common causative agents of such illnesses, a community-based approach epidemiological survey would help health professionals to improve clinical diagnosis and provide appropriate treatment [1, 2].
In Ethiopia, there have been few health facilities based studies to determine the prevalence of typhoid fever, typhus, malaria and brucellosis among individuals presented with febrile illness related symptoms [17–19]. Communities based epidemiological data on the causative agents of common febrile illness related symptoms is generally lacking in the pastoralists areas due to the remoteness of sites and pastoralists way of life. Moreover, in the present study area, health professionals had no clear information on the magnitude of brucellosis, and its clinical based diagnosis might not be even considered. Hence, we assessed the prevalence of typhoid fever, typhus, brucellosis and malaria in individuals who were complaining of symptoms such as fever, headache, joint pain and back pain in the pastoral community of the Amibara district, Afar Region, Ethiopia.
The study was conducted in the pastoral community of Amibara district in the Afar Region of Ethiopia, around 260 km from Addis Ababa. The majority of the study population are pastoralists, depending on livestock for their livelihoods, while some started to practice agro-pastoralism and growing crops along river Awash. The study area and the population has been previously described in detail [20, 21].
A community based cross-sectional survey was carried out between September and December 2016 to determine the prevalence of typhoid fever, typhus, brucellosis and malaria among individuals who were complaining of a range of symptoms. The result of previous health facilities based sero-prevalence of brucellosis (34%) in other pastoral community areas of Ethiopia was used to estimate the sample size [17] for the study on brucellosis. Using this information, a sample size of 380 individuals (95% confidence level, 5% degree of accuracy and 10% compensation for refusal of blood sample) was initially considered. According to the information from Melka Worer health center 50% of patients with symptoms like fever, headache and joint pain often diagnosed as positive for typhoid, typhus or malaria. Based on this supplementary information, sample size was increased to 422 individuals (95% confidence level, 5% degree of accuracy and 10% compensation for refusal of blood sample). However, during the survey period all eligible individuals who came for diagnosis were considered and the sample size was increased to 650.
In this study, six accessible pastoral kebeles of the district were conveniently selected, and a house-to-house survey of all households in the selected kebeles was conducted by community health workers under the supervision of the research team. The heads of the households (husband or wife) or individuals over 18 years were asked if there was any family member (age ≥ 2 years) who manifested symptoms such as fever, fatigue, headache, joint pain, and back pain for short or long periods of time. The individuals were asked to come to the nearest health post, and they were interviewed in their local language (Afar language) using a structured questionnaire that captured common signs/symptoms they felt, onset of illness, treatment sought, and information on socio-demographic characteristics of the individuals. Body temperature was also recorded using a digital thermometer. All individuals equal or older than 2 years, who reported the above symptoms, came for examination and willing to provide blood, and gave informed consent and/or assents were included in the study.
Three ml venous blood sample was collected into plain vacutainer test tube and transported to Melka Werer Health Center. Serum was separated and tested for typhoid fever, typhus and malaria on the same days. The remaining serum was stored at -20°C until transported to the laboratory of Aklilu Lemma Institute of Pathobiology, Addis Ababa University and tested for brucellosis and also tested for typhoid and typhus by test tube based titration method. Anemic individuals and pregnant women were included in the malaria study and provided only finger prick blood sample.
Widal and Weilfelix direct card agglutination tests (DCAT) were used for the serological screening of S. Typhi and Rickettsia infections, respectively following the manufacturer’s instructions (Rapid Labs Ltd, Hall Farm Business Centre, UK), and as previously described [22, 23]. A test tube based titration test was performed for all samples that were found to be reactive by the DCAT and for other 25 randomly selected samples which were found non-reactive as previously described [24]. Rose Bengal Plate Test (RBPT) was used to screen for Brucella infection as previously described [25]. All sera which tested positive by the RBPT and other randomly selected 68 negative samples were further tested using Complement Fixation Test (CFT) following the guidelines of OIE 2008 [26]. The guideline of Ministry of Health (MOH) was followed for the diagnosis of malaria and identification of Plasmodium species at the Health center [27].
Data was entered into EpiData 3.1 and analyzed with Stata/SE 11.0. Descriptive analysis was used to summarize the data in the form of frequencies and percentages of variables. Pearson chi-square test was used to evaluate the statistically significant difference in the level of prevalence of typhoid fever, typhus, brucellosis and malaria between male and female study participants and according to the reported clinical features. Bivariable and multivariable logistic regression analyses were performed to explore associations of socio-demographic characteristics of the study participants with increased odds of having higher prevalence of typhoid fever, typhus, brucellosis and malaria. P-value below 5% was considered as indicator of statistical significance.
This study received ethical clearance from the Institutional Review Board of Aklilu Lemma Institute of Pathobiology, Addis Ababa University (ALIPB/IRB/005/2015/16)). Permission was obtained from Amibara Health Office. Participants’ information sheet which contains the objective of the study, inclusion/exclusion criteria, the required data and methods of data collection as well as informed consent document were prepared in Amharic the national language of the country. Then, the elements of participants’ information sheet initially were orally translated to the local language and described to community leaders and to each of the study participant or parent in case of children under 18 years by trained local health personnel. Informed written consent was obtained from illiterate participants and/or assent in children aged between 12 and 18 years by signing with their finger. Blood sample was collected under aseptic condition by experienced laboratory technicians. Study participants who were found positive for the investigated diseases were treated accordingly as per physician recommendation.
A total of 657 individuals who were complaining various symptoms such as headache, joint pain fever and back pain appeared for clinical examination. However, seven individuals were not volunteers to consent to provide blood sample and they were excluded. Out of the 650 study participants, 630 provided venous blood and 20 provided finger prick blood sample due to anemia and/or pregnancy. The participants’ age ranged from 2 to 80 years with mean age 34.25 ±17.38 years. The majority were illiterate (75.2%) and pastoralists (78.8%) (Table 1).
Table 2 shows the clinical signs, duration of the illness and treatment history as reported by the study participants. Headache (74.8%), joint pain (74.8%), and general malaise (24.9%) were the frequently reported symptoms. The duration of the illness reported by the participants ranged between 2 and 7300 days, with a median duration of 257 days. A total of 159 individuals (24.5%) reported that they sought treatment at various health facilities for their or their family member current illness. Among them, 68/159 (42.8%) were examined and treated for malaria, typhus or typhoid fever, while others (91 participants) reported that they and their families received a treatment though they did not get adequate information for which disease they were treated. The remaining 491(75.5%) individuals did not seek treatment because of various reasons like distance from health facility, the intermittent nature of the illness or lack of money. History of abortion was reported by 116/325 (35.7%) women and the majority (70.7%) of them didn’t know the cause of the abortion.
Out of 630 sera screened by the DCAT, 83 (13.2%) were reactive for S.Typhi infection either against flagella (H) antigen (18/83, 21.7%) or against somatic (O) antigen (41/83, 49.4%) or against both H and O antigens (24/83, 28.9%).
Among the reactive sera to O antigen, 17(19.54%), 8(9.20%), 1(1.2%) and 1(1.2%) were reactive at the titration of 1:80, 1:160, 1:320 and 1:640, respectively. Among the reactive sera to H antigen, 15 (22.73%), 9 (13.64%) and 1(1.52%) were reactive at the titration of 1:80, 1:160 and 1:320, respectively. Thus, the overall sero-prevalence of current infection with S. Typhi as indicated either by H and/or O antigen was considered as 7.3% (46/630) at the cut off value ≥ 1:80. The cases were more common among females than among males (17.2% vs 6.9%, X2 = 14.06, P<0.001) as detected by DCAT and by the titration test (9.4% vs. 4.1%, X2 = 6.35, P = 0.012). All the randomly selected 25 samples which were non-reactive by DCAT were also found non- reactive by the titration test.
In multivariable regression analyses, being female (AOR = 2.21, 95%CI: 1.01–4.83, P = 0.047) and duration of illness above a month (AOR = 2.70, 95%CI: 1.02–7.18, P = 0.046) were found to be associated with a high sero-positivity for S. Typhi infection (Table 3).
Of the 630 sera screened for Rickettsia infection by DCAT, 165 (26.2%) were reactive. Out of these sera, 41(21.8%), 33(17.6%), 9(4.8%) and 5(2.7%) were reactive at the titration of 1:80, 1:160, 1:320 and 1:640, respectively. Hence, 88 (53.3%) samples were reactive by the titration test at the cut off value ≥ 1:80. The combined sero-prevalence for Rickettsia infection by the two tests was 14.0% (88/630). The sero- prevalence was frequent among females compared to males (32.9% vs 15.8%, X2 = 22.74, P<0.001) by DCAT, as well as by titration test (18.5% vs. 6.9%, X2 = 16.83, P<0.001). It was also higher among those individuals who reported headache compared to who did not (28.3% vs 20.0%, X2 = 4.18, P = 0.041) by DCAT and (16.5% vs. 6.5%, X2 = 9.81, P = 0.002) by titration test. All the 25 samples which were non-reactive by DCAT were also non-reactive by titration test.
In multivariable regression analyses, being female (AOR = 3.10, 95%CI: 1.67–5.77, P <0.001) and reporting headache (AOR = 2.80, 95%CI: 1.26–6.22, P = 0.011) were significantly associated with sero-positivity for Rickettsia infection (Table 4).
The sero-prevalence for Brucella infection among the study participants was 12.7% (80/630) by RBPT and 35% (28/80) by CFT. The combined sero-prevalence for Brucella infection by the two tests was 4.4% (28/630). The sero-prevalence for Brucella infection was relatively high in the age group between 2–14 and 15–24 (Table 5). The sero-prevalence was also relatively high among individuals who reported drinking raw milk from aborted animals (13.0% vs. 6.9%) by RBPT and (20.6% vs 6.7%) by CFT. Among the study participants, 569 (90.7%) reported drinking raw milk from aborted animals, 566 (90.3%) touched aborted fetus/discharges from aborted animals without protection and 562 (90.1%) responded that they had no clear information about a disease that causes abortion in their animals.
In the univariable logistic regression analysis; being children (COR = 3.43,95%CI:1.40–8.40, P = 0.007) was found to be associated with high seropositivity for Brucella infection. On the other hand, age 45 and above was found to be associated with a low risk for Brucella infection (COR = 0.22, 95%CI: 0.06–0.75, P = 0.015). In multivariable logistic regression analysis, agropastoralism by occupation was associated with a high risk (AOR = 9.51, 95%CI: 2.30–39.34, P = 0.002) for Brucella infection. None of the 68 samples which were negative by RBPT was found positive by CFT.
The sero-prevalence of Brucella infection was not significantly associated with clinical symptoms reported by the study participants (Table 6).
Of the 650 suspected individuals for malaria, 16 (2.5%) were found positive for P. falciparum malaria infection microscopically, and P. falciparum was the only species detected.
P. falciparum malaria cases were more common among males than among females (4.4% vs 1.3%, X2 = 6.14, p = 0.013). The case was also high in the age group between 2–14 years (8.8%, X2 = 25.13, p < 0.001) and among individuals with body temperature ≥ 37.5°C (18.8% vs 1.6%, X2 = 35.80, p < 0.001). It was also high among individuals felt the illness for a week or less (4.4%, X2 = 6.59, p = 0.037). Multivariable regression analysis showed that being a male (AOR = 4.47, 95% CI:1.24–16.14, P = 0.022) and having fever ≥ 37.5 °C (AOR = 9.17, 95%CI: 1.96–42.84, P = 0.005) were independently associated with increased odds of having P. falciparum malaria infection (Table 7).
Among the total 650 study participants who were tested for S. Typhi, Rickettsial, Brucella and/or Plasmodium infections, 344 (52.9%) were found to be positive for one or more of the infectious agents by the screening tests (Widal and Weilfelix direct card agglutination, Rose Bengal Plate Test and blood films). However, only 24.6% (160/650) were found to be positive for one or more of the infectious agents by the confirmatory tests (titration test for S. Typhi and Rickettsia infections, and Complement Fixation Test for Brucella infection).
We investigated the prevalence of typhoid fever, typhus, brucellosis and malaria among individuals reported signs of fever, headache, joint pain and back pain in Amibara district, Afar Region, Ethiopia, through a community-based cross-sectional study. A quarter of the individuals (24.6%) were sero-positive for S.Typhi, Rickettsia, Brucella infection by confirmatory tests, and/or positive for P. falciparum infection by microscopy. This result is in line with previous health facility based studies in other parts of Ethiopia on causes of febrile illnesses [18, 19], and the highest disease prevalence was found for typhus (14.0%) followed by typhoid fever (7.3%).
However, the sero-prevalence of Rickettsia infection in this study was lower than the one reported from other parts of Ethiopia [19, 28], but higher than the findings of the studies by Tadesse and Tadesse [18] and Birhane et al. [29]. The variation might be linked to the type of environment in the study area. Studies also showed that the occurrence of typhus in Ethiopia is linked to poor hygienic/crowded living condition, where it can cause high mortality rates [30]. The sero-prevalence of Rickettsia infection was high among individuals reported headache. Headache has been shown to be one of the main clinical symptoms of endemic typhus [31]. The study also showed that the seroprevalence for Rickettsia infection was more common among females than among males. This high sero-prevalence for Rickettsia infection among females could be due to the large number of female study participants involved in this study. A previous retrospective sero-prevalence study of typhus among prisoners in Ethiopia also showed a higher seroprevalence among males than among females which could be due to a higher proportion of male study participants (86%) compared to that of female study participants (14%) [28]. However, further well designed community-based study is needed to investigate the reason including differences in treatment-seeking behaviour among adult females and males in the present pastoral area.
Nevertheless, our study showed that typhus is one of the major public health concerns in the study area. Hence, emphasis should be given to appropriate diagnosis/treatment and prevention of Rickettsia infection like through increasing community’s awareness in the present study area.
The second most common disease found was typhoid fever (7.3%). Our result is comparable with other health facility based studies in different parts of Ethiopia [18, 24]. The observed sero-prevalence is lower than the results of health facility based studies in Ziway area [19] and Northwest Ethiopia [29]. Various factors could explain these differences: seasons, environmental hygiene, geographical location and the nature of the study population [32]. The study also showed that the seroprevalence for S.Typhi infection was slightly high among females which is similar to the results of other previous study in Ethiopia [18]. In the present study, the proportion (61%) of female participants was higher than the proportion (39%) of male participants, and this might contribute to the observed high sero prevalence for S.Typhi infection among females. However, a previous health facility-based study in Ethiopia revealed a slightly higher seroprevalence for S.Typhi infection among female study participants (22.5%) compared to that of males (16.7%) despite a higher proportion of male participants (60%) compared to that of females (40%) [29]. In present study area, pastoralists share stagnant and open natural water sources with their livestock, which increases the risk for getting S.Typhi infection and favors the spread of the disease. Moreover, the high seroprevalence for S.Typhi infection among females could be associated with the daily living habits of females like frequency of exposure to contaminated water during fetching water from river and wells or washing clothes as most of these activities are usually performed by females.
In addition, health facilities in the study area had limited laboratory facilities to accurately diagnosis and treat the disease. Thus, increasing access to safe water, strengthening health facility/system for the diagnosis of typhoid fever and treatment as well as increasing community awareness are very important in order to reduce the mortality and morbidity due to this disease.
Several studies have shown the occurrence of brucellosis in livestock in different parts of Ethiopia [33–35]. However, there was no health facilities/community based information on the status of brucellosis in humans in the present study area. The overall sero-prevalence of brucellosis in the study participants was 4.4% by RBPT and CFT. The result is higher than the finding of health facilities based previous studies in individuals with febrile illness in other part of Ethiopia [19,36], but lower than the results of health facility based study from Borena area, South Ethiopia and Metema area, north Ethiopia [17]. Another recent study from Jimma area (south Ethiopia) also revealed a low sero-prevalence of Brucella infection as detected by RBPT (2.1%) and CFT (0.0%) [37].
Brucellosis has an overlap of clinical symptoms with many other febrile diseases, and can be misdiagnosed with malaria or other diseases due to lack of awareness of medical staff and lack of diagnostic capabilities in the present study area. In this study, 4.4% of symptomatic study participants who were found sero-positive for Brucella infection, would not have been diagnosed for brucellosis if the physician would have based the diagnosis solely on clinical signs. Therefore, efforts need to be made to improving laboratory services for the diagnosis of brucellosis in the present study area.
In the present study, significant difference was not found in the prevalence of brucellosis between males and females, which is similar with other studies done in Ethiopia [36], Tanzania [38] and Kenya [39]. However, a hospital based study in Uganda showed a higher sero-prevalence of Brucella infection among males than in females [40]. This can be explained by the fact that among pastoralists, both women and men are equally exposed to risk factors for Brucella infection. Unlike results from Uganda [40] and Bangladesh [41], where elders were more affected by brucellosis, our study showed relatively high prevalence of Brucella infection in the younger age group and children. Children can be exposed to Brucella infection by regularly drinking raw milk, contaminated soil with the bacteria and having regular close contact with livestock, particularly goat and sheep [35].
This study, was undertaken during a high malaria season though it showed a low prevalence of undiagnosed and untreated malaria (2.5%) among the study participants compared to results of previous community-based prevalence study of malaria among non-febrile individuals in Gondar town, North Ethiopia [42], and in the pastoral community of the Bena-Tsemay district, South Ethiopia [43]. Although this low prevalence of malaria might be due to the result of the prevention and control measures employed by the Ministry of Health to eliminate malaria in the country [44], the present observed prevalence of undiagnosed and untreated malaria cases should not be considered insignificant since these undiagnosed and untreated individuals would contribute to the transmission of the disease among the community. Moreover, in the present study area, P. falciparum which causes the most severe form of malaria is the widely distributed species as previously reported [45]. Hence, strengthening community based malaria case detection in this area, for example through community health extension workers is very important in order to achieve the plan for malaria elimination.
In this study, among other clinical features, body temperature ≥ 37.5 °C was found to be strong indicator for infection with falciparum malaria. Previous studies also suggested that increased body temperature could be helpful in diagnosing and treating children with febrile illness [46, 47].
In Ethiopia, serological tests (Widal and Weil-Felix) using the DCAT are commonly used to diagnose typhoid fever and typhus. A number of subsequent studies indicated that these tests are extremely valuable in the absence of adequate laboratory facilities and culture methods like in resources limited countries [48–51].
On the other hand, several previous studies have shown a high seroprevalence of S.Typhi infection using serological based screened tests, but revealed absence or very low prevalence using blood culture or fecal samples which are considered as gold standard tests for the diagnosis of a current infection with S.Typhi [24, 29, 52, 53]. Bacteriological isolation is the gold standard for the diagnosis of current infection with Brucella. RBPT was found to be simple and useful for the screening for Brucella infection in health institutions where bacterial culturing facilities are not available [25], though it is not a useful test to distinguish between acute and chronic Brucella infection [54]. Whereas, CFT can be used as a confirmatory test.
In this study, we have used Widal, Weil-Felix, and RBPT/CFT tests to report the seroprevalence for S.Typhi, Rickettsia and Brucella infections, respectively despite the fact that these serological based tests are not convincing tests for the diagnosis of current infection because of short comings such as false positivity due to previous exposure or false negativity in an endemic setting [24], and this could be one of the major limitations of the present study.
The study participants were recruited based on the clinical signs/symptoms reported by the study participants that may not necessarily have been caused by an infectious agents that cause acute or chronic illness and this might result in a selection bias. The primary objective of this study was to identify the prevalence of brucellosis, malaria, typhoid and typhus among symptomatic individuals with febrile illness related symptoms. However, in an endemic area, these diseases could be prevalent among asymptomatic individuals. Hence, the findings of this study cannot be generalized to the entire population in the study area.
In pastoralists like in the present study area who are living in close contact with their animals on a daily basis, many neglected zoonotic diseases such as campylobacteriosis, Q fever, and leptospirosis can cause a significant health problems both in humans and animals [55]. In the present study, among 650 individuals who complained various illnesses, only 160 (24.6%) were diagnosed for one or more of the above mentioned diseases, and in the majority (75.4%) of the symptomatic individuals the cause of their illness remained unknown, because diagnosis of other related diseases was not considered due to lack of diagnostic tools/reagents and this could also be considered as one of the limitations of this study.
In this study, typhoid fever, typhus, brucellosis and malaria were observed among symptomatic individuals. The study also highlighted that brucellosis cases can be misdiagnosed as malaria or other disease based solely on clinical diagnosis. Therefore, efforts are needed to improve disease awareness and laboratory services for the diagnosis of brucellosis, which should be considered in the routine differential clinical diagnosis of febrile illness in the study area. Only a quarter of the study participants (24.6%) were diagnosed for one or more of the above mentioned diseases. In the majority (75.4%) of the symptomatic individuals, the cause of their illness remained unknown. In addition, a high prevalence of unexplained abortions in women (35.7%) was observed. Hence, further community based studies on other zoonotic diseases like leptospirosis and Q fever are warranted to identify other causes of febrile illness in this pastoral setting.
|
10.1371/journal.pbio.1002103 | A Voltage-Gated Calcium Channel Regulates Lysosomal Fusion with Endosomes and Autophagosomes and Is Required for Neuronal Homeostasis | Autophagy helps deliver sequestered intracellular cargo to lysosomes for proteolytic degradation and thereby maintains cellular homeostasis by preventing accumulation of toxic substances in cells. In a forward mosaic screen in Drosophila designed to identify genes required for neuronal function and maintenance, we identified multiple cacophony (cac) mutant alleles. They exhibit an age-dependent accumulation of autophagic vacuoles (AVs) in photoreceptor terminals and eventually a degeneration of the terminals and surrounding glia. cac encodes an α1 subunit of a Drosophila voltage-gated calcium channel (VGCC) that is required for synaptic vesicle fusion with the plasma membrane and neurotransmitter release. Here, we show that cac mutant photoreceptor terminals accumulate AV-lysosomal fusion intermediates, suggesting that Cac is necessary for the fusion of AVs with lysosomes, a poorly defined process. Loss of another subunit of the VGCC, α2δ or straightjacket (stj), causes phenotypes very similar to those caused by the loss of cac, indicating that the VGCC is required for AV-lysosomal fusion. The role of VGCC in AV-lysosomal fusion is evolutionarily conserved, as the loss of the mouse homologues, Cacna1a and Cacna2d2, also leads to autophagic defects in mice. Moreover, we find that CACNA1A is localized to the lysosomes and that loss of lysosomal Cacna1a in cerebellar cultured neurons leads to a failure of lysosomes to fuse with endosomes and autophagosomes. Finally, we show that the lysosomal CACNA1A but not the plasma-membrane resident CACNA1A is required for lysosomal fusion. In summary, we present a model in which the VGCC plays a role in autophagy by regulating the fusion of AVs with lysosomes through its calcium channel activity and hence functions in maintaining neuronal homeostasis.
| Autophagy is a cellular process used by cells to prevent the accumulation of toxic substances. It delivers misfolded proteins and damaged organelles by fusing autophagosomes—organelles formed by a double membrane that surrounds the “debris” to be eliminated—with lysosomes. How this fusion process is regulated during autophagy, however, remains to be established. Here, we analyze this process in flies and mice, and find that loss of different subunits of a specific type of Voltage Gated Calcium Channel (VGCC) leads to defects in lysosomal fusion with autophagosomes in neurons. It was already known that VGCCs control calcium entry at synaptic terminals to promote the fusion of synaptic vesicles with the plasma membrane, and that mutations in the subunits of VGCCs in humans cause neurological diseases. Our data indicate that defects in autophagy and lysosomal fusion are independent of defects in synaptic vesicle fusion and neurotransmitter release, and we show that a specific VGCC is present on lysosomal membranes where it is required for lysosomal fusion with endosomes and autophagosomes. These observations suggest that the fusion events required in autophagy rely on mechanisms similar to those that trigger the fusion of synaptic vesicles with the presynaptic membrane.
| Autophagy is an evolutionarily conserved, lysosome-mediated degradation process required to maintain cellular homeostasis [1,2]. In eukaryotic cells, autophagy is a ubiquitous process that is important for several physiological processes. It occurs at a basal level in most cells to remove damaged organelles and is required for the turnover of long-lived proteins and other cellular macromolecules. Cellular quality control through autophagy is particularly relevant in long-lived neurons, as evidenced by autophagic malfunction in many human neurological disorders, including Alzheimer’s disease, Parkinson’s disease, Huntington’s disease, and amyotrophic lateral sclerosis (ALS) [3]. In both flies and mice, loss of autophagy-related genes leads to progressive neurodegeneration. It is still an open question whether neurons have their own tailored mechanism to regulate autophagy.
Autophagy is characterized by the formation of an isolation membrane that further elongates to form the double membrane autophagosome, which then fuses with the late endosomes and lysosomes [2]. Soluble N-ethylmaleimide-sensitive factor activating protein receptor (SNARE) proteins have been shown to be required for the fusion of autophagosomes with lysosomes. In yeast, the fusion of autophagosomes with vacuoles, the counterparts of lysosomes, involves the SNARE proteins Vti1 (Q04338.3), Ykt6 (CAA82040.1), Vam3 (CAA99304.1), and Vam7 (CAA96928.1) [4–7], but the latter two have no obvious homologues in metazoan cells. In Drosophila, the SNARE complex required for the fusion of autophagosomes with late endosomes and lysosomes consists of Syntaxin 17 (Syx17) (AGB94109.1), ubiSNAP (SNAP-29) (AAF47071.1), and Vamp7 (AHN56053.1) [8]. The requirement of these SNARE proteins for this fusion step is evolutionarily conserved as Vamp7, and Syntaxin 17 also play similar roles in mammalian cells [9]. Recent studies have shown that two pore channel (TPC), a lysosomal sodium channel, depolarizes lysosome membranes and promotes lysosome fusion upon PI(3,5)P2 stimulation or translocation of mammalian target of rapamycin (mTOR) away from the lysosome [10,11]. It is not established how the change in lysosome membrane potential coordinates the SNARE mediated fusion events.
In an unbiased genetic screen designed to isolate mutations that cause neurodegenerative phenotypes, we isolated many mutant alleles of cacophony (cac) (ID: 32158) that encode an α1 subunit of a Drosophila voltage-gated calcium channel (VGCC). VGCCs consist of multiple subunits, including the conducting pore forming subunit α1, and the accessory subunits α2δ, β, and γ [12]. The α1 subunit contains four internal repeats, each consisting of six transmembrane segments (S1–S6). The loop between transmembrane segments S5 and S6 of each repeat contains conserved domains for short segments 1 and 2 (ss1 and ss2). The calcium ion selectivity of the conducting pore is conferred by a conserved glutamate residue in the ss2 loop of each of the four internal repeats in the α1 subunits [13]. The α2δ subunit of VGCC consists of two disulfide-linked subunits, α2 and δ, derived from posttranslational cleavage of a single gene product [14,15]. In flies, a gene named straitjacket (stj) (ID: 36526) encodes the α2δ subunit which mediates the proper localization of Cac (P91645.3) at synapses [16]. Loss of cac is embryonic lethal in Drosophila and causes an almost complete loss of synaptic transmission [17,18]. stj mutants also exhibit a severe reduction in neurotransmitter release [16].
Mutations in human Cacna1a (ID: 773) and Cacna2d2 (ID: 9254), the orthologs of cac and stj respectively, lead to severe neurological diseases, including episodic ataxia 2, familial hemiplegic migraine 1 (FHM1), absence epilepsy, progressive ataxia, and the polyglutamine disorder spinocerebellar ataxia 6 (SCA6) [19,20]. Mutations in two subunits of Cav2.1 in mice, CACNA1A (AAW56205.1) and CACNA2D2 (Q6PHS9.1), also exhibit ataxia, epilepsy and neurodegeneration [21]. Aside from these spontaneous mutations, knock-in models of FHM1 and SCA6 have also been generated in mice. However, impairments in synaptic transmission do not underlie the mutant phenotypes observed in CACNA1A null mutant mice, and the molecular mechanisms underlying these diseases are still unclear [22]. Indeed, Jun et al. showed that in CACNA1A null mutant mice, excitatory synaptic transmission is largely unaffected because the N- and R-type VGCCs provide the calcium influx needed for synaptic vesicle (SV) fusion. However, these mice exhibit severe neurological deficits, implying that the P/Q-type VGCC plays other important roles than in synaptic transmission [22].
Here we show that, mutant alleles of cac and stj exhibit age-dependent autophagic defects in photoreceptor cells. We find that the role of the VGCC complex in neuronal autophagy is evolutionarily conserved as the loss of the mouse homologues, Cacna1a (ID: 12286) and Cacna2d2 (ID: 56808), also result in autophagic defects in mice cerebella. We provide compelling evidence that the VGCC functions at the lysosomal fusion steps and that its role in autophagy is independent of its role in synaptic transmission. We further demonstrate that the α1 VGCC subunit CACNA1A is present on lysosomes, where it serves as a calcium channel required for lysosomal fusion with endosomes and autophagosomes. We propose that the VGCC in neurons regulates lysosomal fusion through its calcium channel activity on lysosomes.
In order to identify essential genes on the X chromosome that are involved in neuronal homeostasis, we performed a forward genetic screen using ethyl methanesulfonate (EMS). We used the FLP/FRT system to induce homozygous mutant clones in the photoreceptor neurons of the otherwise heterozygous flies and performed electroretinograms (ERGs) on 3- and 33-days-old flies. Flies were exposed to a 1 s light pulse, and the electrophysiological responses were recorded. The amplitude of depolarization reflects photoreceptor activity and the on-off transients reflect pre- and post-synaptic connections. One of the complementation groups corresponds to XE06. The ERGs of these mutants exhibited a reduction of “on” transients in young and old animals (Fig. 1A and B), indicating a loss of synaptic transmission [23]. We mapped the mutations to cac using deficiency and duplication mapping (Fig. 1C). We then performed Sanger sequencing and identified seven different alleles (Fig. 1D): two early nonsense mutations, four missense mutations, and one splicing donor mutant. We selected two alleles for further characterization: cacJ has an early nonsense mutation which is embryonic lethal and cacF has a missense mutation that destroys a key glutamate residue in the calcium ion selectivity loop and is third instar larval lethal (Fig. 1D and E). The lethality associated with both alleles is rescued with a transgene (Fig. 1E).
To examine whether the cac mutants have degenerative phenotypes in the eyes, we performed Transmission Electron Microscopy (TEM) on the mosaic eyes with most photoreceptor cells homozygous for cac mutant. As shown in Fig. 2A and S1 Fig, TEM of cacJ and cacF photoreceptor terminals at day 3 show aberrantly expanded terminals that are more densely filled with SVs when compared to controls (CTL). As the flies age (day 27), the terminals expand further, and the cartridge structure in the lamina is lost (Fig. 2A and S1 Fig). In addition, the number of capitate projections (CPs) (Fig. 2C) and active zones (AZs) (Fig. 2D) decrease dramatically, whereas the number of mitochondria per terminal is increased (Fig. 2E). We also observed a significant accumulation of AVs in aged photoreceptor terminals, showing a progressive worsening of the phenotype (Fig. 2A, B and S1 Fig). The intermediate AVs, especially fusion-primed AVs (blue arrow in Fig. 2B) are greatly increased in aged mutants (Fig. 2F). This suggests that although cac mutant photoreceptor terminals form autophagosomes, there is a defect in autophagosomal maturation and fusion. The genomic fragment containing cac rescues both the morphology defects of photoreceptor terminals in the lamina and the accumulation of AVs in the photoreceptor terminals (S1 Fig).
Poly-ubiquitinated proteins are delivered to autophagosomes and degraded by lysosomes through AV-lysosomal fusion [24]. Therefore, they could serve as a marker to examine the autophagy flux. Indeed, impairment of autophagy has been shown to cause an increase in poly-ubiquitinated proteins in aged Atg7 (ID: 37141) mutant flies [25]. To distinguish whether the accumulation of AVs in cac mutant flies is due to the blockage of autophagy at the late steps or the enhancement of autophagy induction, we measured the autophagic flux by monitoring poly-ubiquitinated proteins in cac mutant flies. We made mosaic flies with cac depleted in the eye, stained the eye-brain complexes of the aged cac flies with an anti-poly-Ubiquitin antibody, and compared the phenotype to the aged Atg7 flies. As shown in Fig. 2G, cac mutant brains accumulate poly-ubiquitinated proteins similar to Atg7 flies, supporting a defect in autophagy. To confirm that autophagic flux is reduced, we also examined protein levels of p62, one of the selective substrates for autophagy, using western blotting (Fig. 2H), in aged fly brains. We find an almost 2-fold increase in p62 in aged mutants, suggesting a reduction in autophagic flux (Fig. 2H and I). All these results, together with the TEM data, indicate that cac is required for autophagy.
Cac regulates autophagy either through its VGCC channel activity or other activities independent of the calcium channel functions. To distinguish between these possibilities, we tested whether loss of other VGCC subunits leads to similar autophagy defects in flies. We examined the flies with a mutation for the α2δ subunit of VGCC encoded by the stj gene [16]. Loss of stj or cac causes very similar autophagic defects, a great accumulation of AVs, suggesting that the VGCC complex, not just Cac, is required for proper autophagy (Fig. 3A–C). We then examined several other mutants involved in lysosomal fusion and function. Loss of the Vacuolar H+ ATPase 100 kD subunit 1 (Vha100-1) (Q8IML5.1), a protein required for endosomal acidification [26], also results in accumulation of AVs in the photoreceptor terminals (Fig. 3D–F). Since cac, stj, and Vha100–1 (ID: 43442)are all required for neurotransmitter release, and mutations in these genes result in SV accumulation in the photoreceptor terminals, it is possible that the autophagy defects we observed in these mutant terminals were a secondary effect of SV accumulation. However, loss of neuronal synaptobrevin (n-Syb) (ID: 38196), encoding a key regulator for neurotransmitter release, results in accumulation of SVs in the terminals [27] but does not lead to the autophagy phenotype (Fig. 3G–I). It indicates that loss of neurotransmitter release in cac and stj mutants is not causing autophagy defects per se. Moreover, loss of Vamp7 [28], a SNARE required for autophagosomal maturation and lysosomal fusion and Fab1 (O96838.2), a kinase required for autophagosomal-lysosomal fusion [29] cause similar AV accumulation phenotypes as cac and stj flies (Fig. 3J–O). These data confirm that Cac and Stj play a role in the AV-lysosomal fusion step of autophagy.
To determine if the role of Cac and Stj in autophagy is conserved in vertebrates, we obtained leaner (Cacna1atg-la) [30,31] and ducky (Cacna2d2du-2J) mice [32,33] that carry mutations in orthologs of cac and stj respectively. Leaner mice have a splicing mutation in the Cacna1a locus [31], whereas ducky mice have a 2 bp deletion within the exon 9 of Cacna2d2 [32]. These mice are deficient in neuronal autophagy and show striking similarities to the mice in which Atg5 (ID: 11793) or Atg7 (ID: 74244) is lost in neurons. The four mouse models exhibit motor defects, ataxia, reduced body size and weight, and smaller cerebella than the wild type (WT) littermates (S1 Table) [34–37]. In addition, these defects occur at similar ages in all these four mutant mice. Cacna1atg-la mice have gradually narrowed granule cell layer and display a purkinje cell (PC) loss starting at day 30 that worsens gradually until there are almost no PCs visible in the anterior lobe at 3 months of age (Fig. 4A and S2 Fig). The degeneration of the Cacna1atg-la mice cerebella resembles that of the Atg5 and Atg7 neuron specific knockout mice which show extensive PC loss by 2 months of age [34–37]. Moreover, all four mouse models exhibit swollen PC axons in the granule cell layers of the cerebella (S1 Table, Fig. 4B–E and S3 Fig).
To examine whether Cacna1atg-la and Cacna2d2du-2J mice indeed suffer from autophagy defects, we performed ultrastructural studies on their cerebella using TEM. We observed many similarities between cerebella from these mutant mice and mice with neuronal autophagy defects. The swollen axons of both Cacna1atg-la and Cacna2d2du-2J mice contain numerous abnormal-looking mitochondria, expanded membranes and endoplasmic reticulum (ER), expanded Golgi cisternal stacks and increased number of autophagosomes, multivesicular bodies (MVBs), and various cytoplasmic vesicles, a hallmark of lysosomal malfunction (S1 Table, Fig. 4C, E and S3 Fig). All of these phenotypes except for autophagosome accumulation have also been documented in mice with neuronal knockouts of Atg7 [37], suggesting that Cacna1atg-la and Cacna2d2du-2J mutant mice are also defective in autophagy. The difference of the autophagosome accumulation between the Atg7 mice and the VGCC mutant mice probably is due to the fact that Atg7 is required for autophagosome formation [38] whereas VGCC functions at the later fusion steps. We then probed the cerebellar lysates of both mutant mice with antibodies against autophagosomal protein LC3 and autophagy substrate p62 (Q64337.1) and observed a small but significant increase in the levels of both the LC3-II form of LC3 and p62 proteins, indicating that the autophagic flux is reduced in these mutant mice (Fig. 4F and G). Immunohistochemistry of Cacna1atg-la Cerebella section also displayed increased levels of LC3 and p62 in the PC soma (S4 Fig), supporting a reduction in autophagic flux.
The reduction in autophagy flux and the accumulation of MVBs and autophagosomes in mutant mice suggest that lysosomal fusion or degradation is compromised. It has been reported that CACNA1A not only localizes to the plasma membrane, but is also present in the neuronal cytosol [39]. To assess if CACNA1A affects lysosomal function by residing on lysosomes, we stained primary cultured cerebellar neurons with two different commercially available CACNA1A antibodies (Millipore and Abcam) and confirmed the specificity of the CACNA1A antibodies using peptide competition assays (S5 Fig). We observe a co-localization of CACNA1A with the lysosomal marker LAMP1 (P11438.2) both in WT and Cacna1atg-la cerebellar neurons, showing that CACNA1A localizes to lysosomes (Fig. 5A, C and S6 Fig). We also stained the neurons with an early endosome marker (EEA1, Q8BL66.2) or ER marker (Calreticulin, P14211.1) in combination with the CACNA1A antibody and detect no obvious colocalization of CACNA1A with these two markers (S7 Fig and S8 Fig). To confirm lysosomal distribution of CACNA1A, we enlarged the lysosomes by pre-treating the primary neurons with Vacuolin-1 [40] and examined the distribution of CACNA1A with LAMP1 and CACNA1A antibodies. In both WT and Cacna1atg-la primary cultured neurons, CACNA1A staining is present as punctae on the membrane of the enlarged lysosomes (Fig. 5B).
To provide independent biochemical evidence, we dissected cerebella from WT mice, and purified and separated lysosomes by iodixanol gradient and observe that full length CACNA1A is enriched in the lysosomal fractions (Fig. 5D). To exclude the possibility of plasma membrane contamination, we also probed the fractions with Annexin V antibody, and no signal was detected in the lysosomal fractions (Fig. 5D). In addition, we also extracted lysosomes from whole brains of WT mice using a subcellular fractionation protocol [41] and find that CACNA1A protein is present in the lysosomal fraction, which was verified with LAMP1 antibody (S9 Fig). The fainter CACNA1A band in the lysosomal lysate as compared to total brain lysate may be due to a much higher level of the protein on plasma membranes than lysosomes.
To assess whether lysosomal function is affected in the Cacna1a mutant cells, we isolated primary neurons from WT and Cacna1a mutant mice cerebella and stained the cells with LysoTracker Red DND-99. WT neurons show big and bright LysoTracker positive vesicles, and LysoTracker staining is almost completely abolished when the neurons are pre-treated with a lysosomal acidification inhibitor bafilomycin A1. As shown in Fig. 6A–D, Cacna1a mutant cells also display severely reduced LysoTracker staining, suggesting that lysosomes are impaired in Cacna1a mutant neurons.
In order to determine if Cacna1a deficient neurons exhibit defects in autophagosomal-lysosomal fusion, we co-stained primary cultured WT and Cacna1a mutant neurons with antibodies against a lysosomal marker (LAMP1) and an autophagosomal marker (LC3). In WT neurons, LC3 and LAMP1 co-localize, whereas little co-localization is observed in Cacna1a mutant cells (Fig. 6E and F), indicating that the fusion between AVs and lysosomes is affected in Cacna1a mutant neurons.
Lysosomes not only fuse with autophagosomes to degrade and recycle intracellular materials but also fuse with late endosomes to degrade and recycle membrane proteins and extracellular material [42]. To determine if lysosomal fusion with other organelles is compromised in Cacna1a mutant neurons, we labeled late endosomes and lysosomes with DQ-BSA. This fluorogenic proteolysis probe permits tracking of endocytic compartments after fluid-phase endocytosis [43]. As shown in Fig. 7A and C, LAMP1 co-localizes extensively with DQ-BSA in WT neurons, but not in Cacna1a mutant neurons. Indeed, numerous green and red punctae are observed in the mutant neurons, whereas most punctae are labeled yellow in WT neurons. These data show that lysosomes fail to fuse with late endosomes in Cacna1a cells.
Given that the primary role of a VGCC is its calcium channel activity and that a mis-sense mutation in the calcium ion selectivity pore in the fly homolog of Cacna1a causes lysosomal fusion defects similar to the loss of the gene, CACNA1A likely regulates lysosomal fusion through its calcium channel activity. Even though CACNA1A is present on lysosomes, the CACNA1A localized at the plasma membrane of synaptic terminals may be required for the calcium influx needed for lysosomal fusion in the cytosol. To rule out that CACNA1A activity on the plasma membrane is required for lysosomal fusion, we applied a P/Q-type calcium channel blocker ω-agatoxin TK at a saturating concentration (1 μM) [44,45] to primary cultured neurons and analyzed endosomal-lysosomal fusion with DQ-BSA. To first test whether ω-agatoxin TK could efficiently block the P/Q-type VGCC on the cell surface, we depolarized the neurons with high potassium chloride solution (90 mM) to activate VGCC, and recorded the calcium influx using a calcium indicator, Fluo 4-AM in the presence or absence of ω-agatoxin TK. ω-agatoxin TK greatly reduces calcium influx induced by the depolarization, indicating that the toxin is blocking depolarization-induced calcium entry via VGCCs (S10 Fig). However, we detected no obvious endo-lysosomal fusion defects in the WT neurons treated with the toxin (Fig. 7C and D). Since ω-agatoxin TK is not cell permeable, our data imply that the cell surface CACNA1A is not essential for lysosomal fusion. We then applied a cell-permeable calcium channel blocker Bepridil (10 µM) to the primary cultured neurons and analyzed endo-lysosomal fusion with DQ-BSA. We detected a significant reduction in colocalization between DQ-BSA and LAMP1 (Fig. 7C and D). A partial block of lysosomal fusion observed here may be due to insufficient block of the intracellular VGCC under current conditions. Taken together, our data suggested that the intracellular CACNA1A but not the cell surface CACNA1A is required for lysosomal fusion.
Here, we show that CACNA1A is present on lysosomes and that it is required for endo-lysosomal fusion and autophagy (Fig. 8). Our work suggests that the VGCC regulates fusion of lysosomes with endosomes and AVs through its calcium channel activity on lysosomes. In the absence of VGCC subunits, as the neurons age and undergo basal autophagy, they accumulate AVs that are unable to fuse with lysosomes. Neurons then accumulate other damaged cellular organelles and misfolded proteins. Eventually this initiates a process of degeneration that mostly affects synapses and synaptic glial cells in fly eyes (Fig. 2A and S1 Fig). The latter phenotype is unlike most other neurodegenerative mutations, which cause a degeneration of rhabdomeres and cell body of the photoreceptors [46,47].
VGCCs have so far been implicated in the fusion of synaptic vesicles and dense core secretory vesicles with the plasma membrane as resident proteins of the plasma membrane [48]. In synapses, a sodium channel gates sodium upon an action potential that depolarizes the plasma membrane and promotes the activation of VGCC. The opening of the VGCC allows influx of extracellular calcium into the cytosol, which stimulates the assembly of the SNARE complex at the fusion site and subsequently promotes the fusion between neurotransmitter loaded vesicles and plasma membrane [49]. Although calcium and SNAREs have also been shown to be required for different steps of autophagy, a VGCC has not been implicated in autophagy previously. Our work suggests that the fusion events required in autophagy are not fundamentally different from those observed at presynaptic terminals for neurotransmitter release.
We show that CACNA1A is present on the lysosomal membrane in neurons. The cytosolic C terminal region of this protein contains a conserved YxxΦ motif across the different orthologs that is known to be required for lysosome targeting of some other lysosomal proteins (S1 Text) [50,51]. All these conserved motifs are remained in Cacna1atg-la. In addition, the CACNA2D2 protein is heavily glycosylated, similar to other lysosomal membrane proteins [15,52], which is known to protect lysosomal membrane proteins from proteolysis [53]. Thus it is likely that these characteristics might be responsible for the lysosomal localization of VGCC subunits, although this idea needs to be further investigated. It is not clear whether the delivery of VGCCs to lysosomes occurs by an indirect route via the plasma membrane or by a direct intracellular trafficking pathway. However, the route must ensure the correct topology of the channel on the lysosomal membrane so as to allow calcium to flow from the lysosome to the cytosol. The lysosomal ionic compositions are similar to the extracellular environment [10] and lysosomes are known to have high calcium content [54]. Studies have shown that the resting calcium concentration inside the lysosomal lumen of human macrophages is between 0.4–0.6 mM [55], a range that is much higher than the approximately 100 nM calcium concentration in the cytosol [56]. The 0.5 mM range of extracellular calcium concentration is sufficient to promote a robust calcium influx during synaptic transmission [57]. Hence, the luminal lysosomal calcium concentration is sufficient to induce calcium efflux from lysosomes through the VGCC to facilitate SNARE-mediated fusion of lysosomes with late endosomes and autophagosomes.
In addition to the requirement of high levels of resting luminal calcium concentration in the lysosomes, there should be a need for a depolarization of the lysosomal membrane to activate the VGCC. Recently, Cang et al. demonstrated that lysosomes are electrically excitable and contain a voltage-activated sodium channel NaV (formed by TPC1, Q9EQJ0.1) [58]. However, the role of this voltage depolarization in lysomes is unknown. Our findings imply that such a NaV-mediated depolarization may be able to activate the VGCCs to trigger Ca2+ efflux from lysosomes in neurons [10,11]. The activity of the sodium channels on the lysosome is triggered by PI(3,5)P2 stimulation or loss of lysosomal mTOR activity, and both are closely associated with autophagy and lysosomal fusion events [10,11,58]. Thus, the TPC proteins are well poised to act as the trigger to activate lysosomal VGCC and facilitate lysosomal fusion with autophagic or late endosomal organelles.
Another interesting question to consider is how the calcium efflux from the lysosomes through the VGCC triggers lysosomal fusion events. At the presynaptic plasma membrane, calcium regulates the fusion machinery through its binding to the C2 domains of synaptotagmin. As a calcium sensor, synaptotagmin interacts with SNAREs and phospholipids to facilitate fusion pore formation upon calcium entry [59]. Synaptotagmin VII (Syt7, Q9R0N7.1) is a calcium sensor that is present on lysosomes and has been shown to be required for lysosomal exocytosis during membrane repair [60]. It would therefore be interesting to test whether Syt7 or other calcium sensors participate in lysosomal fusion events.
In flies, cac is required in neurons because the neuronal specific expression of a transgene of cac can rescue the lethality and phenotypes associated with cac null mutant flies [18]. Mammalian P/Q-type VGCC subunits CACNA1A and CACNA2D2 are also expressed in neuronal tissues [31,32], and mutations in subunits in humans and mice mostly affect neurons. Hence, the requirement of the P/Q-type VGCC for lysosomal fusion might be specific to the neuronal system. Autophagy is a rather ubiquitous process that exists in all tissues. It is still an open question how lysosomal fusion is regulated in cells without P/Q-type VGCCs. Other calcium channels, such as transient receptor potential channels (TRPCs), may play a similar role in non-neuronal cells. Indeed, the TRPC homolog in yeast Yvc1 resides in the vacuole, a lysosome like organelle that can release Ca2+ in response to voltage changes. By contrast, the sole VGCC homolog in yeast localizes on the plasma membrane and is regulated by the stimuli that typically activate TRPC in animal cells [61]. The interchangeable regulation modes of TRPC and VGCC during evolution suggest they may play similar roles in certain conditions.
Our data also suggest a mechanism underlying the role of lysosomal dysfunction in the mouse model of human SCA6 [62,63]. SCA6 is a late-onset neurodegenerative disease caused by a polyglutamine tract expansion at the C terminal of CACNA1A. Unno et al. observed lysosomal involvement based on the accelerated neurodegeneration in SCA6 mice that were also lacking a key lysosomal cysteine protease in the cerebellum, Cathepsin B (P10605.2) [62]. However, they failed to detect autophagic defects in the SCA6 mouse model. This could be due to the low basal autophagy level in cerebella as evidenced by the low levels of LC3II in the cerebellar lysates (Fig. 4F and G) [64]. This may also explain why the AV accumulation in Cacna1atg-la and Cacna2d2du-2J cerebella is not as dramatic as that in fly synapses (Fig. 4C, E and S3 Fig).
In summary, our work has uncovered an unexpected role of VGCC in AV- lysosomal fusion in neurons in helping to maintain cellular homeostasis and provides a new angle to our understanding of the pathology of Cacna1a- and Cacna2d2-related diseases in humans.
The experimental procedures using animals were in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals and approved by Zhejiang University Institutional Animal Care and Use Committee.
cac mutants were isolated from an ey-FLP EMS screen as described previously [46]. The duplication and deficiency mapping were performed as described [65]. The genotypes of the fly strains generated in the paper are as following: Figs 1, 2 and S1 Fig CTL: y w, iso FRT19A / P{w+} cl(1)FRT19A; eyFLP. cacJ: y w, cacJ FRT19A/ P{w+} cl(1) FRT19A; eyFLP. and cacF: y w, cacF FRT19A/ P{w+} cl(1) FRT19A; eyFLP. S2 Fig rescue: y w, cacJ FRT19A/Y; Dp(1;3)DC131. Fig. 2G: Atg7 d4. Fig. 3A and B: y w, eyFLP; stj1 FRT42D/ P{w+} cl(1) FRT42D [16]. Fig. 3D and E: y w, eyFLP; V100–12 FRT82B/ P{w+} cl(1) FRT82B [66]. Fig. 3G and H: y w, eyFLP; FRT sybΔF33B FRT42D/ P{w+} cl(1) FRT42D [67]. Fig. 3J and K: y w, eyFLP; P(EP)VAMP7G7738 FRT42D/ P{w+} cl(1) FRT42D. Fig. 3M and N: y w, eyFLP; fab121 FRT42D/ P{w+} cl(1) FRT42D [29].
For ERG recording, y w *cac (lethal) FRT19A/FM7c, Kr-Gal4, UAS-GFP flies were crossed to y w P{w+} cl(1) FRT19A/Dp(1;Y)y+; eyFLP to generate flies with mutant clones in the eyes, and ERGs were performed as previously described [16]. At least five flies of each genotype were used for quantification.
Dissected fly adult brains were fixed in PBS with 3.7% formaldehyde for 20 min, followed by washing with PBX (PBS + 0.4% Triton X-100) three times. The tissues were incubated with primary antibody overnight in 4°C followed by extensive washing and incubated with secondary antibody overnight at 4°C. After extensive washing, the samples were mounted in Vectashield (Vector Labs) followed by microscopy. Polyubiquitinylated conjugates antibody (FK1) was obtained from Enzo Life Sciences, 1:200 dilution was used. Elav antibody was obtained from Developmental Studies Hybridoma Bank and 1:100 dilution was used.
TEM was performed as described previously [68].
Heterozygous leaner mice with the control genotype, C57BL/6J: tgla/+ were originally obtained from The Jackson Laboratory in Bar Harbor, MA, United States. Male and female heterozygous leaner were mated to produce tgla/+ and homozygous tgla/tgla offspring.
Male and female heterozygous ducky mice (du2J/+) were originally obtained from The Jackson Laboratory and mated to each other to produce control (+/+) wild-type and homozygous mutant ducky (du2J/du2J) mice.
Rabbit polyclonal to LC3 antibody (Novus Biologicals, 1:200 dilution) and Mouse polyclonal to LC3 antibody (MBL, 1:50 dilution) were used for immunofluorescence studies and Rabbit pAb to LC3 (Novus Biologicals, 1:1,000 dilution) was used for immunoblotting. Two rabbit polyclonal to CACNA1A antibodies were purchased from Abcam (1:100 dilution) and Millipore (1:60 dilution) for immunofluorescence studies. Anti-CACNA1A antibody (Millipore, 1:1000 dilution) was used for immunoblotting. Anti-murine LAMP-1 (1D4B) mAb (1:1000 dilution for immunoblotting and 1:500 dilution for immunofluorescence studies) was purchased from Developmental Studies Hybridoma Bank. The mouse monoclonal antibody anti-p62 (1:500 dilution for immunohistochemistry and 1:1,000 dilution for immunoblotting) was from Abcam. Rabbit polyclonal antibody anti-Hsp60 (1:5,000 dilution) was from Epitomics. Rabbit mAb to tubulin (1:2,000 dilution) was from Cell Signaling, and rabbit pAb anti-Calbindin D-28K (1:500 dilution) was purchased from Millipore. Mouse anti-EEA1 mAb (1:1,000 dilution) was from MBL. Chicken pAb anti-Calreticulin (1:200 dilution) was from Abcam. DQ-BSA green and LysoTracker Red DND-99 were from Molecular Probes. Cytosine-β-D-arabinofuranoside were from Sigma. Bafilomycin A1 was from Tebu-Bio. ω-Agatonxin TK and Bepridil hydrochloride were purchased from Tocris Bioscience. Fluo 4-AM and Pluronic F-127 were from Dojindo Laboratories.
Four percent paraformaldehyde-fixed, paraffin-embedded sections in 5 μm thickness were deparaffinized with xylene and washed with distilled water. Tissue sections were boiled for 20 min in 10 mM citrate buffer (pH 7.4). After antigen retrievals, all sections were washed in distilled water, treated with 0.3% (vol/vol) hydrogen peroxide to quench endogenous peroxide, and then incubated with normal goat serum for 30 min. Sections were incubated for 2 h at room temperature with primary antibodies. The primary antibodies were serially detected with the appropriate biotinylated anti-rabbit IgG (Vector), avidin-biotinylated-peroxidase complex (Vector), and, finally, developed with diaminobenzidine (Vector). The sections were washed, counterstained with hematoxylin, dehydrated, and mounted.
Mice cerebella were homogenized in lysis buffer containing protease and phosphatase inhibitors. Protein concentration was determined using Bio-Rad protein assay reagent. Proteins were separated by SDS–PAGE, and transfer the protein onto a PVDF membrane. The membrane was blocked with 5% non-fat milk in TBST buffer and incubated with primary antibodies in 5% non-fat milk in TBST at room temperature for 1 h. Blots were incubated in goat anti-rabbit/mouse-HRP secondary antibody and diluted 1:2,500 in 5% non-fat milk/TBST for 1 h at room temperature. Blots were washed in TBST and then incubated with ECL reagent and exposed. Quantification of protein bands was done using the Image J software.
WT and Cacna1atg-la/ Cacna1atg-la primary cerebella neurons were derived from P0-P2 pups. Cerebella were dissected from pups and individually digested with trypsin. Single cell suspensions obtained were plated on a poly-D-lysine-coated surface in Dulbecco's Modified Eagle's Medium (DMEM) supplemented with 10% (vol/vol) Fetal bovine serum (FBS) and 10% F12 Nutrient Mixture. 12 h after plating, culture medium was half replaced by serum-free neurobasal medium supplemented with B27 (Gibco) and L-Glutamine (Gibco). Mixed cultures were maintained at 37°C and 5% CO2. After 3 days in vitro (div), 5µM cytosine-β-D-arabinofuranoside was added to restrict glial cell growth. The cultures were used for experiments at 7 div–14 div.
The mice were anesthetized with 10% chloral hydrate (0.12 ml/10 g) and perfused with 0.9% NaCl, followed by a 100 ml mix of 1% paraformaldehyde and 1% glutaraldehyde, made in PBS (pH 7.4). After perfusion, cerebella were dissected and stored in fresh fixative overnight at 4°C. 0.5 to 1 mm sagittal sections of each cerebellum were postfixed with 2% osmium tetroxide for 2–3 h, dehydrated through an ascending series of ethanol and embedded in epon812. Ultrathin sections were cut, mounted on uncoated copper grids, stained with 2% uranyl acetate and 1% lead citrate for 12 min each. All the samples were observed using a Hitachi HT7700 electron microscope.
Primary cultured cells were loaded with 2 µM Fluo-4 AM premixed with Pluronic F-127 in regular media for 30 min at 37ºC. Cells were washed in indicator-free media for three times and incubated for another 30 min to allow complete de-esterification of intracellular AM esters. Measurements were done at 37°C in Tyrode’s solution (NaCl 129 mM, KCl 5 mM, CaCl2 2 mM, MgCl2 1 mM, Glucose 35 mM, and HEPES 20 mM), and we added high potassium Tyrode’s solution (NaCl 5 mM, KCl 129 mM, CaCl2 2 mM, MgCl2 1 mM, Glucose 35 mM, and HEPES 20 mM) to a final concentration of 90 mM KCl for depolarization. Cell imaging was taken by high resolution living Cell system DeltaVision Elite.
Fresh cerebella were dissected from mice that were starved overnight and killed the next morning. Then lysosome isolation by subcellular fractionation from the mice cerebella was performed with a lysosome isolation kit (Sigma-Aldrich) according to the manufacturer's manual. After a discontinuous iodixanol gradient centrifugation using Optima MAX-XP Benchtop Ultracentrifuge (Beckman Coulter) with MLS-50 rotor at 150,000 × g and 4°C for 4 h, the sample was divided into ten fractions (0.5 ml each) for further biochemical analyses.
Lysosomes were enriched by centrifugation from a pool of three independent mouse brains in a discontinuous Nycodenz density gradient, as described in [41], with modifications. Briefly, homogenate was prepared in assay buffer (0.25 M sucrose, pH 7.2) and centrifuged in succession at 4,800 × g, 5 min, and 17,000 × g, 10 min. The sediment of the second centrifugation was washed at 17,000 × g, 10 min, resuspended 1:1 vol/vol in 84.5% nycodenz, and placed on the bottom of an Ultraclear (Beckman) tube. On top, a discontinuous gradient of Nycodenz was constructed (layers from bottom to top were: 32.8%, 26.3%, and 19.8% Nycodenz). Centrifugation was for 1 h in an SW 40 Ti rotor (Beckman) at 141,000 × g. Lysosomes were collected from the 26.3/19.8 interface, diluted in 5–10 volumes of assay buffer and centrifuged at 37,000 × g, 15 min. Pellet was resuspended in 500 μl of assay buffer.
Data were analyzed by two-tailed unpaired Student’s t test. A p-value of <0.05 was considered statistically significant.
|
10.1371/journal.pntd.0001054 | Rabies-Related Knowledge and Practices Among Persons At Risk of Bat Exposures in Thailand | Rabies is a fatal encephalitis caused by lyssaviruses. Evidence of lyssavirus circulation has recently emerged in Southeast Asian bats. A cross-sectional study was conducted in Thailand to assess rabies-related knowledge and practices among persons regularly exposed to bats and bat habitats. The objectives were to identify deficiencies in rabies awareness, describe the occurrence of bat exposures, and explore factors associated with transdermal bat exposures.
A survey was administered to a convenience sample of adult guano miners, bat hunters, game wardens, and residents/personnel at Buddhist temples where mass bat roosting occurs. The questionnaire elicited information on demographics, experience with bat exposures, and rabies knowledge. Participants were also asked to describe actions they would take in response to a bat bite as well as actions for a bite from a potentially rabid animal. Bivariate analysis was used to compare responses between groups and multivariable logistic regression was used to explore factors independently associated with being bitten or scratched by a bat.
Of 106 people interviewed, 11 (10%) identified bats as a potential source of rabies. A history of a bat bite or scratch was reported by 29 (27%), and 38 (36%) stated either that they would do nothing or that they did not know what they would do in response to a bat bite. Guano miners were less likely than other groups to indicate animal bites as a mechanism of rabies transmission (68% vs. 90%, p = 0.03) and were less likely to say they would respond appropriately to a bat bite or scratch (61% vs. 27%, p = 0.003). Guano mining, bat hunting, and being in a bat cave or roost area more than 5 times a year were associated with history of a bat bite or scratch.
These findings indicate the need for educational outreach to raise awareness of bat rabies, promote exposure prevention, and ensure appropriate health-seeking behaviors for bat-inflicted wounds, particularly among at-risk groups in Thailand.
| Rabies is a fatal encephalitis caused by lyssaviruses. Evidence of lyssavirus circulation has recently emerged in Southeast Asian bats. We surveyed persons regularly exposed to bats and bat habitats in Thailand to assess rabies‐related knowledge and practices. Targeted groups included guano miners, bat hunters, game wardens, and residents/personnel at Buddhist temples where mass bat roosting occurs. Of the 106 people interviewed, 11 (10%) identified bats as a source of rabies. History of a bat bite/scratch was reported by 29 (27%), and 38 (36%) expressed either that they would do nothing or that they did not know what they would do in response to a bat bite. Guano miners were less likely than other groups to indicate animal bites as a mechanism of transmission (68% vs. 90%, p=0.03) and were less likely to say they would respond appropriately to a bat bite or scratch (61% vs. 27%, p=0.003). These findings indicate a need for educational outreach in Thailand to raise awareness of bat rabies, promote exposure prevention, and ensure health‐seeking behaviors for bat‐inflicted wounds, particularly among at‐risk groups.
| Rabies is an exceptionally fatal encephalitis caused by Rhabdoviruses in the Lyssavirus genus. Transmission typically occurs when broken skin is contaminated with saliva from an infected mammal—usually in association with a bite but in rare instances by scratches. The most well-known and ubiquitous lyssavirus is the rabies virus (RABV), which circulates in New World bats and both Old and New World terrestrial mammals. The vast majority of human rabies cases worldwide are transmitted by dogs infected with RABV. A lesser known member of the genus is the Mokola virus, which has been isolated from a number of terrestrial mammals in Africa (most notably shrews) and has caused at least two human cases [1],[2]. Reservoirs for the remaining nine members of the Lyssavirus genus appear to be exclusively Old World bats [3].
Rabies is a major public health problem in Asia. Of the estimated 55,000 human cases that occur annually worldwide, more than half occur in Asian countries [4]. In recent decades, initiatives aimed at raising rabies awareness (e.g. the World Rabies Day campaign) and lowering human exposure risk through mass vaccination of leading reservoir species have been implemented globally, coinciding with the development of highly potent human rabies vaccines [4],[5],[6]. Notable trends have subsequently followed. Of all Asian countries, Thailand has experienced the steadiest decline in human rabies cases, with a near 10-fold decrease in reported cases during the last 20 years [7]. Much of this decline is attributable to the country's very extensive use of rabies vaccine in the treatment of persons bitten by dogs. In 2003, for instance, more than 400,000 people in Thailand were vaccinated against rabies following potential rabies exposures [8].
Historically in Southeast (SE) Asia, animal-based prevention efforts for rabies have almost exclusively been centered on dogs. Most reported human cases in the region are traced to these animals either through an exposure history or through molecular or antigenic subtyping of variants from rabid human patients [9],[10]. The canine-associated rabies variant has been the only one linked to terrestrial wildlife and domestic animals in Thailand [8], further evidence that dogs are the main lyssavirus reservoir in the region. To date, no human cases of rabies linked to lyssaviruses other than canine-associated RABV have been reported in Thailand or the rest of SE Asia. Like other dog-rabies endemic countries, the majority of human rabies victims in Thailand are children under the age of 15 years [11].
Within the last ten years, however, interest in bats and their role in lyssavirus transmission has increased in the region. The discovery of the Australian Bat Lyssavirus (ABLV) in Australian flying fox bats (Pteropus spp.) in the mid-1990's and the isolation of new bat lyssaviruses in the former Soviet Union [12],[13] were pivotal in turning scientific interest towards the study of potential Asian bat reservoirs. In the last 10 years, evidence of lyssavirus maintenance in SE Asian Chiropterans has emerged from surveillance in Cambodia, Thailand, Bangladesh, and the Philippines [14],[15],[16],[17]. Although lyssaviruses have not yet been isolated from these mammals, neutralizing antibodies associated with lyssaviruses have been detected in sera from both regional mega and microbats. These findings strongly suggest that Asian bats maintain lyssaviruses like their counterparts in Europe, Africa, Australia, and the Americas. Human deaths due to bat-borne rabies infection have been well documented in these continents [12], [18], [19], [20]. In particular, rabid vampire bats are a major cause of human mortality in South America's Amazon region [21].
Routine surveillance for bat rabies is lacking in Asia and as a consequence, understanding is limited regarding the extent of lyssavirus circulation among SE Asian bats and the impact on animal and public health. However, evidence thus far raises pressing questions about human health risks. The potential implications of bat rabies are particularly salient in SE Asia because human-bat interaction occurs routinely in many locales. Bat guano is regularly mined from caves for use as a fertilizer. Hunting of bats for sale and personal consumption occurs as well, despite laws to stop this practice. The presence of large numbers of bats at many Buddhist temples also promotes exposures, as these sites are focal points for commerce, tourism, and religious expression.
Because the severity of skin trauma inflicted by bats is usually minor and unlikely to prompt a medical visit on the basis of physical injury alone, public knowledge of appropriate health-seeking behaviors following a bat exposure is especially important in preventing cases of bat-borne rabies. Rabies postexposure prophylaxis (PEP) for bat bites and scratches, as recommended by both the World Health Organization (WHO) and the U.S. Advisory Committee on Immunization Practices (ACIP), includes thorough wound washing and the administration of rabies immune globulin (in non-immunized individuals) and rabies vaccine administered in a series of doses [4],[22]. When administered promptly and properly, rabies PEP is highly effective in preventing the disease. To date, however, no intervention has proven effectiveness in stopping the clinical course after symptom onset, a fact which further underscores the importance of early care following a possible lyssavirus exposure.
Little is known about the extent of bat-specific rabies awareness in SE Asia. To assess the knowledge and practices of individuals who are most at-risk of bat exposures in Thailand, we surveyed persons who regularly come in contact with bats or bat dwellings through occupational activities and other practices. We sought to elucidate gaps in knowledge that potentially have bearing on rabies prevention, describe the occurrence of bat-associated exposures in this population, and explore factors associated with bat exposures that are of potential consequence to lyssavirus transmission.
The desired sample size was 200 individuals. Surveyed individuals were a convenience sample of adults who collect guano from caves, engage in bat hunting, work or reside at temples that serve as sites for mass bat roosting, or work as game wardens responsible for monitoring and protecting bat caves. Engagement in at least one of these activities within the last 5 years and being 18 years or older were inclusion criteria for participation. Individuals were recruited from eight provinces as shown by Figure 1. Recruitment areas were selected based on proximity to bat caves and/or mass bat roosting sites and certain community characteristics known to promote bat-associated activities (e.g., an agrarian-economy that benefits from guano fertilizer use). Participants were classified based on the activity in which they most frequently engaged, and were primarily recruited from rural and semi-rural localities. In farming communities near bat caves, village leaders and other local contacts provided assistance in locating individuals known to engage in guano mining. Such individuals were also found through referral from existing participants. A similar method was employed to locate bat hunters/trappers in communities where fruit farms are known to attract flying fox bats. To recruit temple workers/residents and game wardens, permission from supervisory officials at Buddhist temples and national parks was obtained before individuals under their management were approached for participation. Recruits were not offered or given incentives for participating.
The study design and consent process was approved by the Institutional Review Board (IRB) at CDC (protocol# 5709). All participants were verbally informed of the study's purpose and assured that their responses would be kept anonymous, even if they engaged in illegal activities. Oral consent was obtained to ensure anonymity and accommodate illiterate participants, and was documented by the interviewer electronically via personal digital assistants (PDA) prior to administering the survey. This method of obtaining informed consent was approved by CDC's IRB.
A 41-item structured questionnaire was developed in English and translated and reviewed by native Thai speakers employed by the office of the U.S. CDC, International Emerging Infections Program in Bangkok. The questionnaire was designed to be administered in Thai via face-to-face interviews, with responses entered in PDAs using GeoAge FAST software. Not all questions provided data used in this study. The questionnaire was developed based on socio-ecological reasoning about gaps in rabies knowledge that potentially translate into failed prevention on the individual level.
Data were collected on demographics; primary bat-associated activity and years of experience; history of rabies vaccination; and type and frequency of bat exposures such as cave entry, direct contact with bats, bites and scratches from bats, and bat consumption. Individuals who reported receiving rabies vaccination were asked to indicate whether it was in direct response to an animal exposure (i.e. PEP) or for pre-exposure immunization (PreP), which is a vaccination series most often administered to people who have a relatively high likelihood of rabies virus exposure due to occupational risks or other factors. Those who reported receiving PreP were asked to describe its administration and only those who indicated receiving a series of injections spaced over multiple days were counted as having PreP.
To assess rabies-related knowledge, participants were asked to rate their understanding of the disease as either “little or none”, “basic”, or “extensive”; explain how humans acquire the disease, and identify animal sources of the disease. Each knowledge question was evaluated independently, and the validity of a participant's self-reported knowledge level was not verified using other responses. Participants were also asked to describe the severity of rabies. Only responses that emphasized death or profound suffering with no suggestion that recovery was likely were considered evidence that the participant recognized rabies as being severe. Awareness of other diseases that humans can get from bats was also elicited.
To assess health-seeking practices following transdermal bat exposures, participants were asked about actions they would take if they were bitten or scratched by a bat. Responses to this open-ended question were compared to a similar question later asked about actions a person should take following a bite from a potentially rabid animal, based on the participant's own understanding of what constitutes a potentially rabid animal. Questions that were specifically asked about bats preceded all questions asked about rabies to minimize reporting bias, and whenever feasible, participants were asked open-ended questions to minimize the interviewer's influence on responses. Participants were also interviewed away from other people. Interviewers were instructed to not ask questions in a leading manner and to allow as much time as necessary for participants to answer.
Survey responses were transferred to a computer, exported into Microsoft Excel, and then imported into SAS version 9.2 for analysis. Data were summarized using descriptive statistics and comparisons by bat-associated activity group were made using Chi-square or Fisher's exact test. Multivariable logistic regression analysis was used to explore factors independently associated with being bitten or scratched by a bat. Variables related to the outcome at p-values≤0.25 were included in the model. Crude and adjusted odds ratios (OR) with 95% confidence intervals (CI) were calculated. Associations were statistically significant at p-values less than 0.05.
The study was conducted during August 3–18, 2009. A total of 106 people were interviewed. Interviews lasted an average of about 10 minutes. Demographic characteristics, history of rabies vaccination, and primary bat-associated activity of the participants are described in Table 1. All temple workers/residents and game wardens were involved in their activity at the time of interview; 71% of guano miners and 53% of bat hunters reported that their most recent engagement had occurred within the previous 12 months. Of all groups, guano miners had fewer years of schooling, with 89% educated at the primary level or less versus 55% of non-guano miners (p = 0.001). Temple workers/residents were more likely to have greater than 15 years of experience in their activity than other groups (59% vs. 26%, p = 0.001).
Thirty-one percent of participants reported a history of receiving either rabies PreP (7.5%) or PEP (23.5%) within their lifetimes. Of those who had received rabies PEP, 96% reported that they had been vaccinated in response to a dog exposure and 4% for a cat exposure. No participants reported receiving PEP for a bat exposure and none reported receiving both PreP and PEP. There were no statistically significant differences between activity groups with respect to rabies vaccination.
Table 2 describes participant's responses to rabies-related knowledge questions by activity group. A majority of participants (54%) reported having little or no knowledge of rabies. Proportionately more temple workers/residents reported basic or extensive knowledge than non-temple workers/residents (p = 0.03). Self-assessed rabies knowledge appeared to be lowest among guano miners, but not to a statistically significant degree (p = 0.06). Although most (85%) participants seemed to be aware that animal bites cause rabies, significant differences were observed between activity groups. Only 68% of guano miners indicated animal bites as a mechanism of transmission compared to 90% of non-guano miners (p = 0.03). When asked to identify which animals are sources of rabies, only 11 (10%) participants named bats. In contrast, dogs were named by 80 (76%), cats were named by 41 (39%), and other mammals (including rodents and large domestic animals) were named by 24 (23%). Fourteen participants (13%) were unable to name any animals as rabies sources. Differences between activity groups with respect to bat attribution were not statistically significant, and individuals who attributed rabies to dogs were no more likely to also attribute rabies in bats than those who did not attribute rabies to dogs (11% vs 8%, p = 1.0) (not shown in Table 2). When asked whether they were aware of any other diseases (besides rabies) that humans can get from bats, 18 (17%) answered yes; all were temple workers/residents (not shown in Table 2).
Table 3 shows how participants responded when asked “What actions would you take if you were bitten or scratched by a bat?” and “If someone has been bitten by an animal potentially infected with rabies what should that person do?” Twenty-eight (26%) participants expressed that they would seek medical care or rabies PEP for a bat bite or scratch, while a significantly higher proportion (95%) advocated these actions if the bite came from a potentially rabid animal (p = 0.0001). The proportion of participants who either said they would do nothing or that they didn't know what they would do if bitten or scratched by a bat was significantly higher than the proportion answering similarly when asked about an exposure to a potentially rabid animal (36% vs. 2%, p = 0.0001). Guano miners were more likely than non-guano miners to give this response for bat exposures (61% vs. 27%, p = 0.003). An incidental finding (not shown in the table) was that previous recipients of PreP or PEP were more likely than non-recipients to advocate a health-seeking behavior for bat bites and scratches but not to a stasticially significant extent (82% vs 56%; p = 0.15).
Table 4 describes bat-related exposures by the number and proportion of participants who reported experiencing it at least once in their lifetimes, along with those who reported experiencing it more than five times a year. A history of transdermal bat exposure (bite or scratch) was reported by 29 (27%) participants.
Table 5 shows factors independently associated with a bat bite or scratch history. Variables considered for inclusion in the multivariable model on the basis of biological plausibility included age, sex, years of experience, education, knowledge self-assessment, frequency in bat caves/roost areas, and bat-associated activity. All except the latter three variables were removed from the final model due to unadjusted p-values>0.25. No two variables were so strongly associated with one another or the outcome as to suggest the presence of colinearity. In the final model, self-assessed rabies knowledge of “little or none” was not significantly predictive of a bat bite or scratch history, although a strong association was observed prior to adjustment for other variables. Individuals who engaged in guano mining, bat hunting, or visiting a bat cave or roost area more than 5 times a year were more likely to report a history of bat bites or scratches.
In this survey among persons at risk for bat exposure in Thailand, we found that although general awareness of rabies transmission and severity were relatively high, awareness of bat rabies in particular was low, with only 10% of participants identifying bats as a potential source of rabies and 36% failing to say they would take any specific action if bitten or scratched by a bat. Bat exposures conducive to potential lyssavirus transmission were also common in this population and were reported by members of all four activity groups, supporting more than just a theoretical risk for these types of incidents. We found that guano miners reported the highest frequency of transdermal bat exposures, were the least knowledgeable about rabies, and were the least likely to say they would respond to bat exposures in a manner that would ensure rabies prevention. Based on these findings we conclude that of the groups we surveyed, bat rabies has the greatest potential impact on guano miners. The potential risk associated with guano mining is even more stark given that moribund bats (i.e., those most likely to be rabid) normally fall to the floor of caves, where they can readily come in contact with someone collecting bat droppings by hand. The effectiveness of any bat-borne rabies prevention strategy may hinge upon how well it diffuses into communities where guano mining regularly occurs.
Education at the community level is an important strategy in the prevention of human rabies [4]. Although the decreasing incidence of human rabies in Thailand points to the effectiveness of past and present rabies education efforts, our findings demonstrate a need to raise public awareness of the potential risk of rabies associated with bat exposures. Special attention should be placed on communities where bats or bat guano are commonly utilized, and if school-based, programs should include primary level students to ensure that they reach those who do not progress past this level of schooling. In addition to emphasizing the importance of exposure avoidance and countering attitudes that inappropriately lower risk perception towards bats, wound washing and healthcare utilization following bat bites and scratches are practices that should be promoted. Similarly, if the awareness we observed in the public is indicative of awareness in the medical community, outreach to healthcare professionals might also be needed to ensure that patients presenting with bat exposures are treated in accordance with WHO guidelines [4]. Studies aimed at assessing knowledge and practices in the Thai medical community should be explored to ensure that such outreach is well-informed. Education at temples and national parks is also recommended to ensure personnel at these sites know to avoid unnecessary bat contact and respond appropriately to bat-inflicted injury. A a strategy that integrates community outreach with law enforcement should be considered as well.
To date, there have been no reported cases of human rabies cases associated with bats in Thailand. One plausible explanation is that the prevalence of bat lyssaviruses in SE Asia is so low that humans are rarely if ever exposed to these pathogens. It is also possible that the prevailing assumption about dogs as the usual source of rabies leads patients and their family members to overlook relevant bat encounters when recounting animal exposures, resulting in misdiagnosis or misattribution. Additionally, the rate of human rabies vaccination in the population may be high enough to protect many people against bat rabies. In our study, we found that almost a third of all participants reported a history of rabies vaccination, mostly as a result of dog-associated PEP. This suggests that the percentage of individuals in our study population with at least some lyssavirus immunity is relatively high and may help account for why bats have yet to be linked to any human rabies cases in the country. However, immunity levels could change if PEP use becomes more conservative in the future. Currently, funds annually spent on the purchase of human rabies biologics by the Thai government are quite substantial [8], and this financial burden may be difficult to sustain indefinitely.
There are some limitations to our study that should be noted. First, it is unlikely that our relatively small convenience sample is representative of all persons engaged in bat-related activities in Thailand. Our findings may have also been subject to reporting bias, since guano miners and bat hunters may have been less willing than others to answer questions truthfully due to the illegal nature of their work. This potential bias may have led participants to understate their years of experience, which could explain why this variable was not found to be associated with a history of transdermal bat exposures. Estimated participation rates for these two groups were also much lower than the other two groups (participation rates were hard to definitively ascertain because participation was ultimately premised on self-identification). Additionally, we classified individuals based on their self-reported primary bat-associated activity; however, a few participants indicated involvement with other activities (e.g,. guano miners that also hunt bats) either presently or in the past. Having such a history was not accounted for this study, although it potentially could be associated with an increased lifetime risk of transdermal bat exposures. The desired sample size of 200 persons was somewhat arbitrarily determined given the lack of reliable estimates for the study population size. Failure to meet this number was largely due to the difficulty in finding willing participants who engaged in bat hunting and guano mining, and the limited availability of personnel and funds that could be used to extend the study period. As a consequence of our small sample size and low statistical power, truly significant associations may have gone undetected in this study. However, by recruiting from several provinces, we minimized the influence that geography might have imparted on the associations we observed. Another limitation is that the validity and reliability of the questionnaire may have been suboptimal because the survey instrument was not subject to very rigorous in-field testing.
Our findings have relevance to zoonotic diseases other than rabies. SE Asian bats have been linked to the encephalitis-causing Nipah virus and Hendra virus [23],[24],[25], and the corona virus associated with severe acute respiratory syndrome (SARS) [26], [27]. Less novel diseases associated with bats also include histoplasmosis, an invasive fungal respiratory disease linked to bat guano exposure [28]. Additionally, evidence suggests that bat ectoparasites may transmit pathogens such as bartonella and rickettsia [29],[30]. In this study, we found that exposures that could potentially facilitate transmission of these diseases appear to occur relatively frequently, with 36% of surveyed participants reporting that they experience direct contact with bats at least twice a year. Bat consumption—an activity that in and of itself may be low risk (assuming the bat is well cooked) but could be associated with increased disease risk through contact with bat carcasses—was reported by more than half the participants. Exposure to toxic or infectious aerosols is another potential hazard for this population as well, since almost all participants reported regularly being in bat caves and roosting areas. More epidemiological studies are needed to better assess the risks associated with bat-related exposures, particularly in regions of the world where outbreaks of severe zoonoses have occurred and questions remain regarding animal reservoirs for such diseases.
|
10.1371/journal.ppat.1000288 | Disruption of the Toxoplasma gondii Parasitophorous Vacuole by IFNγ-Inducible Immunity-Related GTPases (IRG Proteins) Triggers Necrotic Cell Death | Toxoplasma gondii is a natural intracellular protozoal pathogen of mice and other small mammals. After infection, the parasite replicates freely in many cell types (tachyzoite stage) before undergoing a phase transition and encysting in brain and muscle (bradyzoite stage). In the mouse, early immune resistance to the tachyzoite stage is mediated by the family of interferon-inducible immunity-related GTPases (IRG proteins), but little is known of the nature of this resistance. We reported earlier that IRG proteins accumulate on intracellular vacuoles containing the pathogen, and that the vacuolar membrane subsequently ruptures. In this report, live-cell imaging microscopy has been used to follow this process and its consequences in real time. We show that the rupture of the vacuole is inevitably followed by death of the intracellular parasite, shown by its permeability to cytosolic protein markers. Death of the parasite is followed by the death of the infected cell. The death of the cell has features of pyronecrosis, including membrane permeabilisation and release of the inflammatory protein, HMGB1, but caspase-1 cleavage is not detected. This sequence of events occurs on a large scale only following infection of IFNγ-induced cells with an avirulent strain of T. gondii, and is reduced by expression of a dominant negative mutant IRG protein. Cells infected by virulent strains rarely undergo necrosis. We did not find autophagy to play any role in the key steps leading to the death of the parasite. We conclude that IRG proteins resist infection by avirulent T. gondii by a novel mechanism involving disruption of the vacuolar membrane, which in turn ultimately leads to the necrotic death of the infected cell.
| Toxoplasma gondii infects many warm-blooded animals, including approximately one quarter of the world's human population, residing life-long, usually asymptomatically, in cysts in the brain. If, however, the immune system is weakened for any reason, T. gondii can break out and cause life-threatening disease. Furthermore, early T. gondii infection can transfer to the fetus and cause damage. Yet, the human is not a natural host for this parasite. T. gondii reproduces sexually only in cats and the life cycle depends on carnivory by cats of infected intermediate hosts, normally small mammals like the mouse. If the mouse cannot slow down the infection by immune resistance, T. gondii can be lethal. We show that a family of intracellular mouse proteins called immunity-related GTPases (IRGs) can attack T. gondii early after it has infected a cell and kill at least some of the parasites. The infected cells die too, but we show that in dying they release a protein that can stimulate local immunity. For most strains of T. gondii, this early immune attack is probably enough to prevent the parasite from killing the mouse, and allows the parasite to establish a long-lasting infection in the mouse brain or muscle.
| The mouse is a natural intermediate host for Toxoplasma gondii, an apicomplexan parasite whose definitive host is the cat. Most T. gondii strains are not virulent for normal mice at low infective doses [1],[2]. Following the development of a strongly IFNγ-dependent primary immunity, rapidly replicating tachyzoites convert to the slowly-replicating bradyzoite stage and encyst in brain and muscle without causing severe symptoms, there to await completion of the infection cycle following ingestion by a cat at some later time [3]–[5]. This relatively benign course of infection is, however, drastically altered by disruption of genes encoding key components of the interferon-gamma (IFNγ)-response pathway [6]–[8]. In recent years it has become clear that a group of highly IFN-inducible GTPases, the IRG proteins (formerly p47 GTPases [9]) play an essential role in limiting the early tachyzoite replication stage [10]. Genomic disruption of individual members of this gene family causes normally avirulent T. gondii strains to behave as highly virulent pathogens, killing infected mice as early as 10 days after primary infection [11],[12].
The effects of IRG gene disruption on the whole animal are mirrored by the failure of IRG protein-mediated resistance processes occurring in individual T. gondii-infected cells. T. gondii tachyzoite replication in infected cells can be measured in tissue culture in a variety of cell types, including fibroblasts, macrophages and astrocytes, and pre-treatment of such cells with IFNγ causes potent inhibition [13]–[15]. In cells derived from mice with single disrupted IRG genes, IFNγ-mediated control of T. gondii replication is more or less reduced, though rarely completely eliminated [10], [16]–[19]. Marginal loss of control in cells deficient in Irga6 (IIGP1) contrasts with highly significant inhibition in IFNγ-treated wild-type cells expressing a dominant negative form of Irga6 [17], presumably as a result of the high level of interactivity recently documented between the IRG proteins [20]. Up till now, Irgm1 (LRG-47), Irgm3 (IGTP), Irgd (IRG-47), Irga6 and Irgb6 (TGTP) (for the new nomenclature of the p47 GTPases, see reference 50) have all been documented as participating in resistance to T. gondii either at the cellular or whole animal levels, or both.
Some years ago we reported the rapid accumulation of IRG proteins on the parasitophorous vacuole membrane (PVM) of T. gondii infecting IFNγ-treated mouse astrocytes and the subsequent appearance of vacuoles where the vacuolar membrane was apparently disrupted [17]. We documented local vesiculation and perforation of the IRG-coated PVM, and death of the included parasite as evidenced by penetration of Irga6 into the moribund parasite detected by immunoelectronmicroscopy. These findings have been confirmed in more recent studies both in macrophages [19],[21] and very recently again in astrocytes [16]. In the present report we have used live cell imaging to be able to put a time scale on these events. We are able to document the disruption of IRG protein-loaded vacuoles within the first hours after infection. Approximately 20 minutes after disruption of the vacuole the parasite itself is dead, as documented by its permeability to cytosolic protein markers. About an hour after the death of the parasite the infected cell itself undergoes necrotic death, losing plasma membrane integrity and releasing the chromatin modelling protein, HMGB1, known to be a potent pro-inflammatory stimulus [22]. We show in two ways that cellular necrotic death depends on the disruption of the PVM: firstly, by direct observation, cellular necrosis occurred only after the disruption of at least one PVM and, secondly, both vacuolar disruption and cellular necrosis occur very rarely in fibroblasts infected with virulent strains of T. gondii. Our results thus connect the action of IRG proteins at the PVM directly to mouse resistance to T. gondii at both the cellular and whole animal levels.
We reported the accumulation of several IRG proteins at the PVM in T. gondii-infected, IFNγ-stimulated cells [17]. To monitor subsequent events at the PVM in live cell microscopy, Irga6 was tagged at the C-terminus with EGFP. A short linker sequence, ctag1, between the Irga6 coding region and the EGFP was also required to prevent aggregation of the tagged protein in vivo (see Materials and Methods). The correctly tagged Irga6 protein, Irga6-ctag1-EGFP, localises to the endoplasmic reticulum like wild-type Irga6 in IFNγ-induced cells ([17] and unpublished) and accumulates on the PVM of T. gondii-infected, IFNγ-induced cells. Irga6-ctag1-EGFP was transfected into IFNγ-induced mouse embryonic fibroblasts (MEFs) that were then infected with T. gondii ME49 strain. Vacuoles with Irga6-ctag1-EGFP accumulations were followed in live-cell microscopy. Vacuolar rupture was seen as a sudden breach at a single point in the normally smooth ring of Irga6-ctag1-EGFP surrounding the PV (Fig. 1 and Videos S1, S2, S5). In fixed cell microscopy we have previously shown that such breaches also correspond to breaches in the GRA7 signal at the PVM, confirming that they correspond to breaches in the parasitophorous vacuole membrane (see panel B of the fifth figure in [17]). Over a period of about 5 minutes the breach widened, usually continuing until the visible Irga6 was accumulated at one pole of the organism (see Fig. 1 and see also panel E of the second figure in [17]). It was frequently observed that the generally banana-shaped form of the PVM that follows the shape of the included parasite became rounded up shortly before rupture (Fig. 1a, c, white arrowheads). After rupture of the PVM the parasite reverted to its banana shape (compare the phase contrast images at 1 hour 36 min after addition of T. gondii (1:36, see Materials and Methods), before vacuolar disruption, and 1:42, after vacuolar disruption). The rounding up, the sudden development of the rupture, the rapid expansion of the breach and the reversion of the parasite to its banana shape after PVM rupture all suggest that the PVM is put under tension before it ruptures. The live cell observations reported here correspond generally to those very recently reported in astrocytes using Irgm3-EGFP (IGTP) instead of Irga6-ctag1-EGFP as a monitor of T. gondii vacuolar rupture [16].
In our earlier report we observed by immunoelectronmicroscopy that the normally partially cytosolic Irga6 was frequently found inside moribund T. gondii [17], suggesting that dying or dead intracellular T. gondii in IFNγ-induced cells become permeable to cytosolic proteins. To examine this further, we used modified GFP proteins, EGFP and Cherry, as markers for the cytosolic pool while observing individual T. gondii vacuoles by live cell imaging (Fig. 2 and Videos S3, S4, S7, S8). In Fig. 2A, vacuoles containing impermeable T. gondii are seen as dark forms excluding EGFP. During the observation period individual vacuoles suddenly fill with EGFP (Fig. 2A and Videos S3 and S4). The complete disappearance of the T. gondii “shadow” implied not only that the PVM became suddenly permeable to soluble fluorescent protein, but also that the parasite itself, still visible in the phase-contrast images, became permeable. Thus at this point the parasites are clearly dead. Monitoring the influx of a cytosolic protein marker, while efficient and sensitive, did not tell us when the vacuole ruptured relative to the moment of permeabilisation of the parasite. Is the PVM permeable to proteins, and the parasite dead, before the vacuolar membrane is visibly ruptured? Or does rupture of the PVM coincide with entry of the fluid phase marker, allowing for the possibility that the parasite is already dead when the PVM ruptures, or perhaps that permeabilisation of the parasite is essentially simultaneous with the rupture of the PVM? Finally, there was the possibility that PVM rupture significantly precedes permeabilisation of the parasite, perhaps implying that the rupture of the PVM is a precondition for the death of the parasite as monitored by breakdown of the permeability barrier. To resolve this issue we turned to two-color live cell imaging, loading IFNγ-induced cells by transfection simultaneously with Irga6-ctag1-EGFP and soluble Cherry before infection (Fig. 2B and Videos S5, S6, S7, and S8). With this approach we were able to show conclusively that the PVM ruptures significantly before the permeabilisation of the included parasite. In Fig. 2B the PVM disrupts at 1:39 and the T. gondii becomes permeable to Cherry at 2:06, thus 27 minutes after visible disruption of the PVM. In several observations with this double-labelling technique we could conclude that all T. gondii contained in ruptured vacuoles are dead within 20–40 minutes after rupture. Thus rupture of the PVM leads inexorably to the death of the included parasite. Naturally it is likely that the parasite is irretrievably committed to die before the membrane permeability barrier breaks down, but we have no convenient assay for earlier events.
A role of IRG proteins in this sequence of events is suggested but not formally demonstrated by these experiments. IRG proteins have however been shown to be required for successful IFNγ-dependent control of T. gondii replication, both by the relaxation of this restriction documented in cells with deleted Irgm1 or Irgm3 genes [10],[16],[18] and by similar loss of control following transfection of dominant negative Irga6 [17] and Irgb6 (unpublished) mutants. We now wished to show that permeabilisation and therefore death of T. gondii in IFNγ-induced cells was also dependent on IRG proteins. We therefore adapted the permeabilisation assay demonstrated in Fig. 2 to fixed cells in order to be able to obtain quantitatively significant data. MEFs were transfected with the pEGFP expression plasmid, infected with T. gondii, fixed at different times after infection then stained with antibodies against the T. gondii dense granule protein, GRA7, which is expressed in the PVM. Cells were examined by conventional fluorescence microscopy for disrupted vacuoles containing permeabilised T. gondii. Fig. 3A shows that permeabilised T. gondii were first found 30 minutes after infection, their numbers rising continuously to a plateau of about 20% of all vacuoles. In uninduced cells, a low percentage of permeabilised parasites was also seen after 2 hours, perhaps attributable to a low level of stimulation by Type I IFN released from primary MEFs as a result of the transfection procedure. This assay was then used to count permeabilised parasites in IFNγ-induced cells transfected with either wild-type Irgb6 or its dominant negative mutant, Irgb6-K69A, both FLAG-tagged at the C-terminus. Irgb6-K69A, like the homologous mutant of Irga6-K82A, can scarcely go to the PVM and forms aggregates in both uninduced and induced cells that trap the wild type IFNγ-induced protein in the cytoplasm and prevent its localisation on the PVM [17],[20]. The Irgb6-K69A protein may be a more efficient dominant negative than Irga6-K82A perhaps because more vacuoles normally load with Irgb6 than with Irga6 (unpublished results). In Fig. 3B (top row) an IFNγ-induced cell transfected with EGFP and wild-type Irgb6-FLAG contains 2 intracellular T. gondii. The vacuole to the left (white arrow) is already disrupted as judged by the polar distribution of Irgb6-FLAG (red) and the parasite has obviously already been permeated with EGFP. The vacuole and parasite to the right (white arrowhead) is still intact. In Fig. 3B (bottom row) an IFNγ-induced cell transfected with Irgb6-K69A-FLAG and EGFP contains one intracellular T. gondii intact within its vacuole. The PVM is very weakly labelled with Irgb6-K69A-FLAG, which is accumulated elsewhere in aggregates in the cytoplasm. Fig. 3C shows that transfection of the Irgb6-K69A mutant into IFNγ-induced cells causes a reduction in EGFP-positive T. gondii at both 2 h and 4 h after infection relative to IFNγ-induced cells transfected with the wild-type Irgb6 construct as a control.
These results support earlier evidence that resistance to T. gondii is reduced by dominant negative IRG proteins [17] and indicates that IRG proteins act early after infection.
We observed in live cell imaging that the fluid phase fluorescent protein markers suddenly and invariably disappeared from the cell shortly after the permeabilisation of the parasite. In Fig. 4 (see also Videos S9 and S10) is shown the complete sequence of events in an IFNγ-induced MEF expressing Irga6-ctag1-EGFP and mDsRed, infected with 3 T. gondii parasites, two with conspicuous Irga6-ctag1-EGFP positive PVMs. The upper ringed vacuole disrupts at 1:10 and the parasite becomes permeable between 1:25 and 1:30. The lower ringed vacuole disrupts at 1:40 and becomes permeable between 2:20 and 2:25. At 2:30 the cytosolic mDsRed is suddenly lost from the cell, one hour and 20 minutes after the disruption of the first vacuole. Parallel phase contrast images of the film sequences showed a drastic restructuring of the cytoplasm coinciding with the disappearance of the cytosolic marker at 2:30. This crisis clearly corresponded to the death of the cells. Thus the nucleus suddenly condensed and was separated from the cytoplasm by a clear phase-dense margin, and the cell ceased active movement. There was no sign of either nuclear or cytoplasmic blebbing. We never observed similar effects occurring in neighbouring uninfected cells, nor in infected cells in which neither PVM rupture nor parasite death had occurred. In repeated observations, loss of cell membrane integrity regularly occurred something over an hour after disruption of a PV. The death of the infected cell, like the permeabilisation and death of the parasite, seems also to be an inexorable consequence of a sequence of physiological events beginning with the rupture of the PVM.
The apparently reactive death of the parasite-infected, IFNγ-induced cells with sudden permeabilisation of the cell membrane was suggestive of necrotic rather than apoptotic death [23]. Consistent with this, when IFNγ-induced, EGFP-transfected cells were infected with T. gondii and incubated with soluble Alexa-555-labeled annexin V to detect phosphatidyl serine [24], the apoptotic marker was at no time detectable on the plasma membrane (Fig. 5A and Videos S11 and S12). After loss of EGFP, annexin V accumulated slowly on internal membranes. We next examined loss of cytochrome C from mitochondria as an early indication of apoptosis [25], using immunofluorescence on fixed IFNγ-induced cells transfected with EGFP as a fluid phase marker and infected with T. gondii (Fig. 5B). In no cell containing a permeabilised T. gondii, and therefore due to die within an hour or so, did we see release of cytochrome C into the cytoplasm (Fig. 5B panels c–f). When apoptosis was actively induced by treatment of the cells with TNFα and cycloheximide [25], cytoplasmic cytochrome C was present in many of the treated cells (Fig. 5B panels a, b). Furthermore, no cleavage of caspase-3 could be detected in western blots of lysates of IFNγ-induced, T. gondii-infected cells up to 8 hours after infection, nor cleavage of the caspase-3 substrate PARP, both of which signs of apoptosis [26] were easily detected in MEFs treated with cycloheximide and TNFα [25] (Fig. 5C). On the other hand, the pro-inflammatory chromatin binding protein, HMGB1 [22], was detected as early as 2 h after infection in the supernatant of the IFNγ-induced, T. gondii infected cells but not in the supernatants of the cycloheximide and TNFα-treated cells (Fig. 5C).
From these results we could conclude that disruption of the PVM and death of the included T. gondii initiates a non-apoptotic death in IFNγ-induced cells with some features of necrosis. By analogy with other pathogen-induced necrotic processes [27]–[29], we considered it likely that activation of caspase-1 by the inflammasome would be found [30]. However we were unable to demonstrate activation of caspase-1 or its substrate IL-1β in IFNγ-induced positive control MEFs stimulated by LPS and treatment with nigericin to induce necrosis [31]. We therefore turned to primary bone marrow derived macrophages (BMMs), which respond to IFNγ induction and T. gondii in the same way as MEFs, ending in cell death (unpublished results, and see below). Neither cleavage of caspase-1 nor processing of IL-1β were detected in IFNγ-induced BMMs infected with T. gondii for 6 hours compared with cells treated with LPS and nigericin as a positive control for inflammasome-mediated necrotic death [32] (Fig. 5D). This suggests that the IFNγ-dependent necrosis seen in T. gondii-infected cells may be related to the pyronecrosis reported in mouse macrophages infected with Shigella flexneri where caspase-1 cleavage was shown not to be required for necrosis [33].
The necrotic death of infected cells about an hour after the permeabilisation of the parasite explains why the peak percentage of permeabilised T. gondii cannot rise above about 20% (Fig. 3A). When the cell dies by necrosis it detaches and is lost during the processing of slides. Thus the only live cells that can be detected containing a permeabilised vacuole are those seen in the approximately 1 hour long window between parasite permeabilisation and cell death.
In view of the apparent inevitability of the post-vacuolar disruption-dependent necrotic death, it was surprising that this effect has not been reported before. To illustrate the scale of the phenomenon, we prepared a simple overview image of IFNγ-induced fibroblasts infected 8 hours previously with T. gondii ME49 strain at a MOI of 5 and washed free of floating T. gondii at 2 hours after infection. Propidium iodide was added immediately before microscopic examination of the living culture. Strikingly large numbers of heavily condensed, propidium iodide stained nuclei were seen, indicating dead cells, compared with infected control cells not induced with IFNγ (Fig S1).
Following recent evidence for the control of certain bacterial infections by engulfment of the microbes in autophagic membranes [34]–[36], Yap and colleagues employed electron microscopical evidence to implicate autophagy in the destruction of T. gondii in in vivo-activated mouse peritoneal macrophages [19]. We had earlier reported the presence of EGFP-LC3 vesicular structures in the vicinity of some disrupted T. gondii vacuoles in IFNγ-induced astrocytes, though no clear-cut co-localisation could be observed [17]. To re-examine this issue we again observed the behaviour of EGFP-LC3, both in fixed cell (Fig. 6A) and live-cell imaging (Fig. 6B and Videos S13 and S14) in MEFs infected with T. gondii ME49 strain. In the absence of IFNγ induction we occasionally saw some EGFP-LC3 associated with T. gondii vacuoles. Fig. 6A, top panel, shows a very rare cell with three T. gondii vacuoles with LC3 concentrated at the vacuoles (white arrowhead); a fourth vacuole in the same cell (white arrow) is not ringed with EGFP-LC3, and there is no sign of general formation of LC3-positive punctae. When the cells were previously induced with IFNγ, EGFP-LC3-positive structures could be found associated with occasional vacuoles, independently of whether Irga6 positive or negative (Fig. 6A, middle and lower panels). Equally, vacuoles with clearly disrupted PVMs (Fig. 6A, lower panel) were no more likely to be EGFP-LC3 positive than intact vacuoles (Fig. 6A, middle panel). EGFP-LC3-positive punctae were sometimes increased at later times after infection although at no time was there any obvious correlation between LC3 signals and Irga6-positive vacuoles. Indeed the majority of vacuoles in IFNγ-induced cells showed no association with LC3-positive structures.
In view of the heterogeneity of association of EGFP-LC3 with vacuoles in fixed-cell preparations we chose to examine the dynamic behaviour of the autophagic marker in live-cell imaging. Fig. 6B shows consecutive frames of an IFNγ-induced MEF infected with ME49 T. gondii. The EGFP-LC3 signal is uniformly distributed in the cell with some punctae, which show no striking association with the two visible T. gondii vacuoles. Between frames 2:40 and 3:20 (more accurately in the video between 3:00 and 3:05) after infection (Fig. 6B white arrow) one infecting T. gondii becomes permeable to the cytosolic EGFP-LC3 and 95 minutes later the LC3 marker is lost as cell membrane integrity breaks down and the cell undergoes necrotic changes visible in phase contrast (also see Videos S13 and S14). Thus it is clear that the whole series of events from disruption of the vacuole through the death of the T. gondii to the necrotic death of the cell can occur without the enclosure of the vacuole in any LC3-positive membranes.
The final elimination of intracellular pathogens through macroautophagy depends on fusion of the pathogen-containing autophagic vacuole with lysosomes followed by degradation of the pathogen [35]. Consistent with the live-cell imaging showing no necessary participation of EGFP-LC3 in the destruction of the pathogen and the initiation of necrosis, no evidence was seen for fusion of T. gondii-containing vacuoles with lysosomes, assayed by association with LAMP1, in IFNγ-induced MEFs at any time-point after infection, as we reported earlier [17] and in contrast to the observations of Ling et al [19] (Fig S2). In summary, we found no evidence that macroautophagy plays any necessary role in the IRG-dependent destruction of T. gondii in IFNγ-induced MEFs.
We have shown elsewhere by 3H-uracil incorporation assays that the replication of the virulent type I T. gondii strain, RH, is only slightly inhibited by induction of MEFs with IFNγ [37]. We also showed in fixed preparations that IRG proteins failed to accumulate normally on the PVM of infecting RH parasites: fewer vacuoles accumulated each of the IRG proteins tested and the failure of Irgb6 to accumulate on RH vacuoles was especially striking (9% of RH vacuoles positive for Irgb6 compared with 70% of ME49 vacuoles [37]). It was therefore of interest to examine the fate of virulent T. gondii in IFNγ-induced cells. The frequency of parasites detected as permeabilised in the fixed cell system exploited above was too low to be estimated above background levels. In view of the temporal correlation between vacuolar loading, vacuolar disruption, T. gondii permeabilisation and cellular death by necrosis documented above in cells infected with the ME49 avirulent T. gondii, it was of interest to establish whether cells infected with a virulent strain of T. gondii also undergo necrotic death. The large scale of necrotic death of IFNγ-induced, T. gondii infected cells (Fig S1) can be documented quantitatively at the population level by a standard cell viability assay (Materials and Methods). Fig. 7A shows the results for ME49-infected MEFs (panel a) and BMMs (panel b). Both cell types showed an IFNγ concentration dependent loss of viability with increasing multiplicity of infection. Thus the IFNγ- and T. gondii-dependent cell death programme is not distinctive for fibroblasts but, as noted above, also occurs in macrophages. In striking contrast, scarcely any loss of viability was seen in IFNγ-induced MEFs infected with the virulent type I strain RH or its transgenic descendant, RH-YFP, even at the highest MOI (Fig. 7A, panels c, d). However, by direct microscopical observation of IFNγ-induced cells infected with RH strain, very occasional necrotic cells containing apparently moribund T. gondii were detected (unpublished). Thus the necrotic process can be initiated by virulent T. gondii, but at a far lower frequency. It will be important to establish whether the rare necrotic cells are included in the rare subset of RH-infected cells that load normally with IRG proteins.
This result showed that virulent T. gondii are scarcely subject to the IRG-mediated resistance programme. Thus PVMs containing virulent T. gondii are generally not disrupted, virulent T. gondii are not permeabilised in the cytosol, and the host cells do not undergo necrotic death. In the light of these results, and from the evidence that IRG protein loading of the virulent vacuole is significantly reduced [37], it seemed likely that virulent T. gondii prevent the IRG-mediated resistance programme from initiating. To support this conjecture, we examined the behaviour of cells doubly infected with avirulent (ME49) and virulent (RH-YFP) organisms. In the cell viability assay (Fig. 7B) it was clear that doubly-infected MEFs were just as vulnerable to IFNγ- and T. gondii-dependent cell death as MEFs infected only with the avirulent strain, showing that the resistance of the virulent strain depends on failure of an earlier event than the necrotic process itself. In co-infected, IFNγ-induced MEFs transfected with Cherry as a fluid phase marker and assayed at 4 hours after infection (Fig. 7C), the presence of the virulent RH-YFP had no impact on the permeabilisation of parasites, which reached the usual figure of over 20% (see Fig. 3A). Thus the resistance of virulent T. gondii to IRG proteins is not due to the secretion of a soluble factor that renders the cell incompetent to resist any infecting T. gondii, and the vacuolar destruction mechanism is also intact in cells infected with virulent T. gondii. We conclude from these results that it is the disruption of the vacuoles that distinguishes avirulent from virulent T. gondii strains. Since dominant negative Irga6 and Irgb6 prevent the accumulation of IRG proteins on the PVM [17],[20] and also inhibit vacuole disruption and permeabilisation of the parasite (Fig. 3C), our experiments suggest that the essential difference between virulent type I and avirulent type II T. gondii lies in the ability of the virulent parasite to prevent the massive accumulation of IRG proteins on the PVM [37]. The same line of argument leads us also to the conclusion that the critical function of the IRG proteins is probably already fulfilled with the disruption of the PVM.
In the present study we have documented a succession of events, beginning shortly after infection of an IFN-stimulated mouse cell by Toxoplasma gondii, and ending with the necrotic death of the infected cell. We can characterise this series, as follows: (1) the accumulation of IRG proteins on the parasitophorous vacuole, which begins on some vacuoles as early as 2 minutes after entry of the parasite and typically reaches a maximum between 30 minutes and one hour later (unpublished data), (2) rupture of the IRG-loaded PVM, a process which occurs suddenly and is completed in a few minutes, (3) the permeabilisation of the T. gondii plasma membrane documented by entry of fluorescent cytosolic protein markers, which occurs as a sudden event between 20 and 40 minutes after PVM disruption, (4) the necrotic death of the cell, documented by sudden loss of fluorescent cytosolic protein markers and release of the chromatin modelling protein, HMGB1, from the necrotic cell. Integrating our results we can describe an approximate time course for the overall process (Fig. 8 and Table S1). Major variables in this chronology are: the time to initiate PVM loading with IRG proteins, which may be almost immediately following infection to many hours or not at all, for different vacuoles even in the same cell ([17] and unpublished); the time between PVM loading with IRG proteins and vacuole disruption, which can also vary a great deal and may be dependent on the number of loaded vacuoles in the cell (unpublished observations); and lastly the time between permeabilisation of T. gondii in a disrupted vacuole and cell death, which has varied from 5 to 105 minutes over our experimental observations. The least variable time interval is that between vacuolar disruption and the permeabilisation of the included T. gondii, which, with a single exception at 45 mins has been consistently close to 20 mins.
By several criteria, the fulfilment of this programme is dependent on the activity of IRG proteins. Restriction of T. gondii in IFNγ-induced cells has been shown to be dependent on IRG proteins in several earlier studies [10],[17],[18] and two recent studies have specifically shown dependence of vacuolar disruption on the presence of Irgm3 [16],[19]. We show here (Fig. 3) that a dominant-negative mutant of Irgb6 inhibits the appearance of permeabilised T. gondii in the IFNγ-induced cell. We also show that cells infected by virulent T. gondii, which interfere with PVM loading by IRG proteins ([37] and unpublished), do not undergo necrotic death. As documented extensively elsewhere (Khaminets, unpublished) and confirmed in the present experiments (uncoated T. gondii in Fig. 4 and Videos S9 and S10) not all PVM accumulate IRG proteins in multiply infected cells. In our experiments such unloaded vacuoles have never been seen to disrupt. There is thus a clear implication that the accumulation of IRG proteins on a PVM is required for the disruption of that vacuole. Once the vacuolar membrane is disrupted, the rest of the programme moves inexorably forward to the permeabilisation of the parasite and the ultimate necrotic death of the cell. In this scenario, the disruption of the PVM is the critical step that is required for the completion of the programme. We have used permeability to cytosolic proteins as a criterion for death of the parasite because it is unlikely to be controversial. The rather tight linkage in time between vacuolar disruption and permeabilisation of the parasite may suggest that the parasite begins to die at the same time as PVM disruption.
Our experiments do not support the suggestion that the formation of autophagic membranes around disrupted vacuoles [19] plays a necessary role in the unfolding of the programme we describe. LC3-positive membranes undoubtedly form in a proportion of infected cells [17], although in contrast to studies of others based on macrophages [38],[39] we have seen no significant difference between IFNγ-treated and untreated cells (unpublished results). However the full vacuolar destruction and necrosis programme can apparently be implemented without any significant LC3-positive membrane formation (Fig. 6B and Videos S13 and S14). Indeed, the dying parasite exposed in a disrupted vacuole is fully accessible to a cytosolic protein marker, showing that the parasite is not at that stage isolated from the cytosol in an autophagic isolation membrane or membrane-bounded secondary lysosome. This latter conclusion in supported by our failure to find any co-localisation of LAMP1 with the dying parasite (Fig S2). Nevertheless, in an earlier study we observed that transformed fibroblasts from atg5-deficient mice were significantly less able to restrict T. gondii replication than transformed fibroblasts from control mice of the same strain [13]. In view of the lack of evidence for formation of autophagic membranes at the vacuoles, this result suggests that the effect of atg5 deficiency is not directly connected to the role of atg5 in autophagy. In fact, IFNγ-dependent accumulation of Irga6 on the PVM has recently been shown to be significantly reduced in atg5-deficient primary BMMs [21], a result we can confirm and have extended to other IRG proteins (Khaminets, unpublished). Since the results of the present paper argue that adequate accumulation of IRG proteins on the vacuole is a critical determinant of their function, it seems probable that the reduced control of T. gondii replication in atg5-deficient fibroblasts [13] is due to the defect in IRG protein accumulation on the PVM and not to any later effect that can be attributed to deficient autophagy.
The results of this study highlight many unknowns in the cell biology of the relationship between T. gondii and its mammalian hosts. We have already stressed that the early loading of IRG proteins onto the PVM is critical for the initiation of the resistance programme, and this is the step that appears to fail with virulent type I parasites [37]. Although loading of the PVM involves transition from the resting GDP-bound to the activated GTP-bound state of some IRG proteins [20],[40], little else is known about the means by which these proteins specifically access the target membrane, nor how this access is impeded by virulent type I T. gondii. Is it due to active interference with the IRG proteins, for instance by one of the rhoptry kinases which are known to function as virulence factors [41]–[43], or is the resistance to IRG proteins an intrinsic characteristic of the PV membrane itself, or of the parasite-derived components it contains, due, for example to polymorphism in the structure or presence of an unknown IRG receptor? The disruption of PVs carrying accumulations of active IRG proteins is a striking and now repeatedly confirmed finding [16],[17],[19],[21]. The apparent build-up of tension in the PVM shortly before disruption, seen as a tendency for the vacuole to become spherical, and the speed with which the local disruption spreads strongly suggests that the PVM surface area is being reduced, and this idea would be consistent with EM images showing apparent vesiculation of the PVM [16],[17],[19],[21]. The tendency to view IRG proteins as in some way related to dynamins [44],[45] further fosters the idea that IRG proteins participate in a process of active vesiculation of the PVM. This dynamic model will be hard to test if several IRG proteins participate cooperatively in causing membrane damage [20]. This problem has recently been exacerbated by evidence that members of the family of IFNγ-inducible 65 kDa guanylate binding proteins also target the T. gondii PVM and may participate in cell-autonomous resistance [46]. Is this a collaboration with IRG proteins, or a separate enterprise?
Once the PVM disrupts, the fate of the parasite is sealed. In about 20 minutes it becomes permeable to cytosolic proteins but it is unknown both what the cause of death is, and when the process is initiated. From the relatively close temporal correlation between vacuolar disruption and permeabilisation of the T. gondii it is plausible that the process is initiated at the time of vacuolar disruption. Why should the cytosol become an inimical environment when the PVM disrupts? Are cytoplasmic immune receptors triggered by material released from the disrupted vacuole? If so, what is the activating material, which are the receptors and what is the mechanism that leads to death of the T. gondii? We considered the possibility that reactive oxygen species could contribute to the killing of the parasite, but the entire necrotic programme played out without noticeable deviation from wild-type behaviour in T. gondii-infected IFNγ-induced cells derived from p47phox-deficient mice (unpublished results) confirming an earlier report that genomic disruption of p47phox does not prevent IFNγ mediated control of T. gondii in vivo or in vitro [47].
The death of the parasite is followed with the same inevitability by the death of the infected cell. This death is accompanied by no release of mitochondrial cytochrome C, no cleavage of caspase-3 or of its effectors, and no appearance of phosphatidyl serine on the outer leaflet of the plasma membrane, so it is clear that it does not reflect activation of an apoptotic cascade. On the other hand, the dramatic loss of plasma membrane integrity and release of HMBG1 speak strongly for necrosis. The absence of detectable caspase-1 cleavage or release of mature IL-1β recalls the cryopyrin-initiated cell death recently described as pyronecrosis [33] that follows Shigella flexneri infection of mouse macrophages, which proceeded unimpaired in infected BMMs from caspase-1 deficient mice. Normal resistance to T. gondii in vivo is also reported from caspase-1-deficient mice [48]. There is, however, the important difference that in that study caspase-1 was activated and IL-1β cleaved. Thus if the T. gondii-induced necrotic programme indeed proceeds via activation of cryopyrin (NLRP3), the pathway must diverge from the inflammasome before cleavage of caspase-1. Another potentially important difference between the present study and Shigella-induced necrosis is that T. gondii-induced necrosis with the properties we describe occurs only in IFNγ-induced cells. While our studies are directed to the role of IRG proteins in driving this process, it is not excluded that further IFNγ-induced components may participate in the terminal necrotic step and are responsible for its distinctive properties.
There is a considerable body of information showing that T. gondii-infected cells may become resistant to apoptotic stimuli, probably mediated via parasite-induced activation of NFκB and the transcriptional up-regulation of anti-apoptotic genes, as well as by inhibition of cytochrome C release and caspase-9 activation (reviewed in [49]). However several of these reports are studies on human cells which lack IFNγ-inducible IRG proteins [50] while those infecting mouse cells in vitro used the virulent RH strain [51]–[55], so the relevance to our own observations is not explicit. The necrotic death that we report with avirulent strains evidently overrides T. gondii-mediated control of apoptosis by superimposing a different lethal programme under the control of IFNγ. As noted above, however, it is unclear what role IFNγ plays in the necrotic programme. Is it simply the induction of IRG proteins leading to the disruption of the vacuole with release of an unknown pro-necrotic signal, or are there further IFNγ-induced components that play a crucial role in the subsequent necrotic events?
It is remarkable that the death of IFNγ-induced cells infected with avirulent T. gondii strains has not been noticed before since it is a large scale and relatively early event that, with reasonable MOIs, is immediately apparent after a cursory look down the microscope. There is, however, an isolated report of IFNγ-dependent cell death in a study on 3T3 fibroblasts infected with the close relative of T. gondii, Neospora caninum [56]. In this study the cell death observed was attributed to apoptosis, but this diagnosis was not confirmed experimentally. In the same study, in contrast, no cell death was detected in cells infected with T. gondii, encouraging the authors to suggest that the two related parasites stimulate different processes. In fact, however, they used virulent RH strain T. gondii, and as we show here and elsewhere [37], no necrotic death is associated with this strain because of its inhibition of IRG function.
The uracil incorporation assay for IFNγ-dependent inhibition of T. gondii replication is now 24 years old and has been used extensively in the T. gondii research field [14],[15]. It is an extremely useful and sensitive assay. However the necrotic death of many IFNγ-induced cells infected with T. gondii in the first 8 hours of the assay means that 3H-uracil added to the assay after 48 hours and measured at 72 hours to a certain extent records not only failure of the parasites to replicate, but also the fact that there are fewer live cells for the T. gondii to infect.
A further issue provoked by our results is their relevance to the in vivo situation. Although the process we describe leads to the death of individual ME49 T. gondii in infected, IFNγ-induced cells, additional parasites infecting the same cell seem to remain viable so long as the integrity of their vacuoles is maintained. In the Videos S9 and S10 a T. gondii enclosed in an intact vacuole that has accumulated no Irga6 remains visible until the last frame. As the host cell dies by necrosis following the earlier disruption of another vacuole this surviving parasite can be seen in the phase contrast image apparently to escape from the corpse of its host by active movement. The vision is fleeting, but a recent publication documented large-scale egress of live T. gondii from IFNγ-induced astrocytes [16] and egress has been reported before [57]–[59]. We did not see exit on such a scale in our experiments, but it may nevertheless be questioned whether the death of individual intracellular T. gondii documented here is the major adaptive purpose of the necrotic process. Melzer et al assert that the putatively released T. gondii in the astrocyte system are no longer capable of invasion, which would also obviously contribute to an adaptive advantage [16]. A further plausible and not exclusive hypothesis is that the local release of pro-inflammatory signals such as HMGB1 from necrotic cells, favouring local accumulation of cellular components of the innate and adaptive immune systems, contributes significantly to the resistance process in vivo [22].
An additional reason for seeing the process we describe as important for in vivo events in T. gondii infection is the correlation between the induction of necrotic death in cellular infection in vitro and the genetics of T. gondii virulence defined by lethality in mice. For type I strains, high lethality in vivo is associated with reduced IRG protein accumulation on the PVM (9% RH vacuoles are Irgb6-positive compared with 70% ME49 vacuoles [37]), no vacuole rupture and no subsequent necrotic death of infected cells in vitro. Several virulence factors have now been identified in rhoptry secretions [41]–[43] and it is plausible that one or more of these is dedicated to interference with the IRG resistance mechanism, presumably at the vacuolar loading step. The vacuolar loading step appears to be the Achilles heel of the IRG resistance mechanism since it was reported recently that failure of IFNγ-inducible resistance to Chlamydia muridarum in mice is also associated with failure of IRG proteins to accumulate normally on the chlamydial inclusions, compared with successful IRG loading and efficient IRG-mediated resistance to the closely-related C. trachomatis [60]. In this case, however, resistance to the IRG system was functionally dominant, since the presence of C. muridarum in the cell actively inhibited the accumulation of IRG protein on co-infecting C. trachomatis. The molecular basis of the attack of C. muridarum on the IRG system must therefore be different from the attack by virulent T. gondii, an unsurprising conclusion in view of the taxonomic disparity between these two pathogens. It was recently shown that several of the IFNγ-inducible p65 guanylate binding proteins also accumulate on the T. gondii PVM, and this accumulation too failed with a virulent strain [46]. Does this imply yet another defence mechanism? In the absence of new information it is difficult to see the defeat of innate resistance mechanisms in mice by virulent T. gondii as a success for either partner. The mouse is destined to die within 10 days [1],[2], hardly a triumph for the mouse, while death of the mouse before large-scale bradyzoite transition and encystment reduces the chance for T. gondii to infect any cat lucky enough to catch the mouse in the short time window before it dies. The adaptive significance of the virulent phenotype may be more relevant to the genotype of a different intermediate host, for example the rat, Rattus norvegicus, which is not vulnerable to virulent strains of T. gondii [61].
The IRG system is a complex resistance mechanism with multiple interacting components and many unexplained features. Its adaptive significance for pathogen resistance in mice is hard to judge at present since only C. muridarum and T. gondii of the organisms studied are natural mouse pathogens [62]. From data presently available, however, it seems likely that the IRG system is under selective pressure from polymorphic pathogens in the natural environment and may display significant variation as a consequence. This aspect of the IRG system has not yet been investigated systematically. What is, however, clear, is that IFNγ-inducible IRG genes are not present in humans [50], leaving resistance to pathogens such as T. gondii and Chlamydia to be organised by other means. At present the principal candidate mechanism in this role for T. gondii is the IFNγ-inducible catabolic enzyme, indoleamine dioxygenase, whose expression results in depletion of free cytosolic tryptophan [63]. It will, however, be of some interest to monitor the fate of T. gondii in IFNγ-induced human cells at the same level of resolution as has been employed here and elsewhere in mouse cells.
The Irga6-ctag1-EGFP construct was employed for live-imaging experiments following the demonstration (Sascha Martens, unpublished) that this doubly tagged construct showed normal localisation behaviour following transfection into IFNγ-induced cells. Irga6 carrying a single C-terminal EGFP tag formed aggregates in cells and probably activates spontaneously [20],[40]. The construct was generated by amplification of the Irga6ctag1 sequence from pGW1H-Irga6ctag1 [40] using Irga6ctag1 forward 5′-cccccccccgtcgaccaccatgggtcagctgttctcttcacctaag-3′ and reverse 5′-cccccccccgtcgacgtcacgatgcggccgctcgagtcggcctag-3′ primers and cloned into pEGFP-N3 (Clontech) by SalI digestion. The pmDsRed-N3 and pmCherry-N3 constructs were generated by amplification of mDsRed and mCherry respectively from pDsRed-Monomer-N In-Fusion (Clontech) or mCherry-pRsetB (generous gift from Dr.. Roger Y. Tsien, UCSD) using the following primers, and inserted into pEGFP-N3 following BamHI/NotI digestion: pmDsRed-N3 forward 5′- cccccccccggatccatggacaacaccgaggacgtcat-3′ and reverse 5′-cccccccccgcggccgcctactgggagccggagtggcgggc-3′, pCherry-N3 forward 5′-cccccccccggatccatggtgagcaagggcgaggagga-3′ and reverse 5′-cccccccccgcggccgcctacttgtacagctcgtccatgc-3′. EGFP-LC3, pGW1H-Irgb6-FLAG, pGW1H-Irgb6-K69A-FLAG constructs were generated as described [17],[20],[40].
C57BL/6 embryonic fibroblasts (MEFs) were prepared from mice at day 14 post coitum and cultured in DMEM (high glucose) (Invitrogen) supplemented with 10% FCS (Biochrom AG, Berlin, Germany), 2 mM L-glutamine, 1 mM sodium pyruvate, non-essential amino acids, 100 U/ml penicillin, 100 mg/ml streptomycin (all PAA, Pasching, Austria) as described [9]. C57BL/6 bone-marrow derived macrophages (BMMs) were prepared from young adult mouse bone marrow cells cultured in DMEM (high glucose) containing 10% L929 P2 cell-conditioned medium and supplemented with 10% FCS (Biochrom AG, Berlin, Germany), 2 mM L-glutamine, 1 mM sodium pyruvate, non-essential amino acids, 100 U/ml penicillin, 100 mg/ml streptomycin. Apoptosis was induced by adding TNFα (40 ng/ml, PeproTech, NJ, USA) and cycloheximide (10 µg/ml, Sigma-Aldrich) to the medium. Inflammasome activation is induced by incubating cells with LPS (1 µg/µl, Sigma-Aldrich) for 24 hours, and then treated with nigericin (20 µM, Sigma-Aldrich).
T. gondii tachyzoites from the type I strain RH-YFP or type II strain ME49 were maintained by serial passage in confluent monolayers of human foreskin fibroblasts (HS27, ATCC CRL-1634) as described [17]. RH-YFP parasites were propagated in the presence of Chloramphenicol (3.2 µg/ml, Sigma-Aldrich) to maintain the stably integrated YFP expression plasmid containing a chloramphenicol acetyltransferase selection marker [64]. Both T. gondii strains were a generous gift from Dr. Gaby Reichmann, Medical Microbiology, University of Düsseldorf.
Cells were transiently transfected using FuGENE6 (Roche) according to the manufacturer's instructions, and induced with mouse IFNγ (PeproTech, NJ, USA) 24 hours before infection with T. gondii tachyzoites at the MOI indicated in the data element.
All the live cell imaging experiments were performed in μ-slide I chambers (Ibidi, Munich)., which are exceptionally well-suited to this kind of experimentation. For live cell experiments, all procedures including transfection, medium modification (addition of IFNγ etc) and infection with T. gondii could be carried out while the cells were continuously incubated in an observation volume of 100 µL and cells could be maintained in excellent condition for at least 20 hours. Cells were incubated at 37°C in phenol-red-free DMEM supplemented with 10% FCS, 20 mM HEPES pH 7.4, 2 mM L-glutamine, 1 mM sodium pyruvate, 1× non-essential amino acids, 100 U/ml penicillin, and 100 µg/ml streptomycin. MEFs were transfected and stimulated with 200 U/ml IFNγ for 24 hours. After infection with T. gondii, the cells were observed with a Zeiss Axiovert 200 M motorized microscope fitted with a wrap-around temperature-controlled chamber, using an EC “Plan-Neofluar” 40×/1.30 Oil Ph3 objective (Zeiss). The time-lapse images were obtained and processed by Axiovision 4.6 software (Zeiss). Phosphatidylserine was detected by adding 1% (v/v) Alexa-555 labeled annexin V (Molecular Probes, Invitrogen) with 2.5 mM CaCl2 into the incubation medium.
The following serological reagents were used for immunofluorescence (IF) and western blot (WB): anti-Irga6 rabbit antiserum 165 (IF 1∶8000, [17]), anti-Irgb6 goat antiserum A20 (IF 1∶100, Santa Cruz Biotechnology), anti-GRA7 mouse monoclonal antibody (IF 1∶1000, gift from Dr. Gaby Reichmann, Medical Microbiology, University of Düsseldorf. [17]), anti-FLAG M2 mouse monoclonal antibody (IF 1∶4000, Sigma-Aldrich), anti-cytochrome C mouse monoclonal antibody (IF 1∶1000, BD PharMingen Clone: 6H2.B4), anti-LAMP1 rat monoclonal antibody 1D4B (IF 1∶1000, University of Iowa, USA), anti-PARP rabbit polyclonal antibody (WB 1∶1000, Cell Signaling Technology, MA, USA), anti-cleaved caspase-3 rabbit polyclonal antibody (WB 1∶1000, Cell signaling Technology, MA, USA), anti-HMGB-1 rabbit polyclonal antibody (WB 1∶250, Abcam), anti-Calnexin SPA-865 rabbit antiserum (WB 1∶10000, Stressgen), anti-IL-1β rabbit polyclonal antibody (WB 1∶2500, Abcam), anti-caspase-1 p10 M20 goat polyclonal antibody (WB 1∶200, Santa Cruz Biotechnology), goat anti-mouse Alexa 488 and 546, goat anti-rabbit Alexa 488 and 546, donkey anti-rat Alexa 488, donkey anti-goat Alexa 350, 488, 546 and 647, donkey anti-mouse Alexa 488, 555 and 647, donkey anti-rabbit Alexa 488, 555 and 647 (IF 1∶1000, Molecular Probes, Invitrogen), donkey anti-rabbit HRP (Amersham), donkey anti-goat HRP (Santa Cruz Biotechnology) and goat anti-mouse HRP (Pierce Biotechnology) (WB 1∶5000).
Immunofluorescent staining was performed on paraformaldehyde-fixed cells essentially as described earlier [17]. Images were taken with a Zeiss Axioplan II fluorescence microscope equipped with an AxioCam MRm camera (Zeiss) and processed with Axiovision 4.6 software (Zeiss). 4′,6-Diamidine-2′-phenylindole dihydrochloride (DAPI, Invitrogen) was used for nuclear counterstaining at a final concentration of 0.5 µg/ml. Intracellular parasites were identified by immunostaining for vacuolar localisation of the T. gondii protein dense granule proteins, GRA7 or by their distinctive appearance in phase contrast.
MEFs (7500 cells/well) or BMMs (2×104 cells/well) were seeded into 96-well plates and treated with IFNγ or under control conditions for 24 hours. The cells were then infected with T. gondii at the indicated MOI for 8 hours. Thereafter, viable cells were quantified by the CellTiter 96 AQueous non-radioactive cell proliferation assay (Promega) according to the manufacturer's instructions. The absorption of the bio-reduced form (formazan) of a substrate (MTS) generated by metabolically active cells during incubation at 37°C for 2–4 hours was measured in an ELISA reader (Molecular Devices) at 490 nm. The quantity of formazan product is proportional to the number of living cells in the culture.
|
10.1371/journal.ppat.1000827 | Serological Profiling of a Candida albicans Protein Microarray Reveals Permanent Host-Pathogen Interplay and Stage-Specific Responses during Candidemia | Candida albicans in the immunocompetent host is a benign member of the human microbiota. Though, when host physiology is disrupted, this commensal-host interaction can degenerate and lead to an opportunistic infection. Relatively little is known regarding the dynamics of C. albicans colonization and pathogenesis. We developed a C. albicans cell surface protein microarray to profile the immunoglobulin G response during commensal colonization and candidemia. The antibody response from the sera of patients with candidemia and our negative control groups indicate that the immunocompetent host exists in permanent host-pathogen interplay with commensal C. albicans. This report also identifies cell surface antigens that are specific to different phases (i.e. acute, early and mid convalescence) of candidemia. We identified a set of thirteen cell surface antigens capable of distinguishing acute candidemia from healthy individuals and uninfected hospital patients with commensal colonization. Interestingly, a large proportion of these cell surface antigens are involved in either oxidative stress or drug resistance. In addition, we identified 33 antigenic proteins that are enriched in convalescent sera of the candidemia patients. Intriguingly, we found within this subset an increase in antigens associated with heme-associated iron acquisition. These findings have important implications for the mechanisms of C. albicans colonization as well as the development of systemic infection.
| Candida albicans has both a benign and pathogenic association with the human host. Previous to this study, little was known in regard to how the host humoral system responds to the commensal colonization of C. albicans, as well as the development of hematogenously disseminated candidiasis. We show using a C. albicans cell surface protein microarray that the immunocompetent host exists in permanent host-pathogen interplay with commensal C. albicans, and undergoes stage-specific antibody responses as the yeast transitions from a benign microbe to an opportunistic fungal pathogen. Also identified were serological signatures specific for acute and convalescent stages of candidemia. Our findings provide new insight in the characterization of potential serodiagnostic antigens and vaccine candidates to the opportunistic pathogen C. albicans.
| The yeast Candida albicans exists in a dichotomist relationship with the human host. C. albicans is frequently found as a commensal organism on the human skin, gastrointestinal (GI) tract and the vulvovaginal tract [1]. Close to 60% of healthy individuals carry C. albicans as a commensal in the oral cavity. Colonic and rectal colonization is even higher, ranging from 45% to 75% among patient groups. Alterations in the host immunity, physiology, or normal microflora rather than the acquisition of novel or hypervirulent factors associated with C. albicans, are suggested to lead to the development of candidiasis [2]. Both neutrophils and mucosal integrity of the GI tract, are critical in preventing hematogenously disseminated candidiasis [3]. The development of candidemia can begin with the translocation of C. albicans into the bloodstream from initial commensal GI colonization or the shedding from developing biofilms on indwelling catheters [4],[5]. Fungal cells that evade the host immune system can spread to deep organ systems leading to hematogenously disseminated candidiasis, which has an estimated mortality rate of 40%, even with the use of antifungal drugs [2].
Information on in vivo gene expression would provide insight into how C. albicans interacts with host cells during the transition from commensal colonization to an opportunistic pathogen in the immunocompromised host. However, in vivo transcription profiling of C. albicans during commensal colonization or candidemia is technically challenging [6]. Instead, several genome-wide transcriptional analyses of C. albicans responses to host cells have been performed using ex vivo and in vivo infection models. These include phagocytosis of C. albicans cells by neutrophils [7] and macrophages [8], exposure to human blood, plasma, and blood cells [9],[10], as well as invasion of perfused pig liver and reconstituted human epithelium [11],[12]. Genes that are associated with morphological changes, metabolic adaptation, and oxidative stress are the major responses of C. albicans to host cells identified in these studies. The changes in gene expression identified in these in vitro model systems possibly reflect tissue- or stage-specific expression during an infection in patients. Profiling of antibody responses during infection in patients offers an alternative approach that can overcome technical challenges of in vivo transcription profiling. An antibody-based approach has been used to identify C. albicans gene expression during thrush in individuals with HIV [13].
Currently the isolation of C. albicans from blood cultures is the standard method for the diagnosis of candidemia. Nevertheless, blood cultures may only become positive late in infection, and in one study up to 50% of all autopsy-proven cases of candidemia were reported as negative in blood cultures [14]. Thus, the ability to rapidly and easily diagnose candidiasis is urgently needed. An alternative approach to microbiological confirmation of C. albicans infection is serological diagnosis. An immunoproteomic approach using two-dimensional electrophoresis followed by quantitative Western blotting and mass spectrometry has been used to profile serologic response to peptides from cell surface extracts in candidemia [15]–[17]. A significant proportion of antigens identified were glycolytic enzymes and heat shock proteins. An antigenic multiplex consisting of the peptides Bgl2, Eno1, Pgk1, Met6, Gap1, and Fba1 provides 87% sensitivity and 74% specificity when distinguishing patients with candidemia from uninfected hospital patients [17]. However, this approach has several limitations; only the most abundant and soluble proteins can be resolved on the immunoblot, there is a lack of reproducibility of cell wall preparations, and most importantly, there is the inability to account for various stage- and tissue-specific gene expressions from the cultured cells. These limitations can be addressed by using a protein microarray to profile antibody responses [18]–[21].
To investigate the establishment of the humoral immunity during commensal sensitization, as well as the adaptive immune response to candidemia, we have developed a C. albicans cell surface protein microarray. Our rationale in developing a cell surface protein microarray is that the cell surface of C. albicans is the immediate target of the human immune system when C. albicans cells enter the bloodstream. Cell surface proteins play important roles in host interaction, and many of them are known virulence factors. In addition, a recent study showed that there is a significant expansion of cell wall, secreted and transporter gene families in pathogenic Candida species in comparison to non-pathogenic yeasts [22]. In this study, profiling of serological response on the protein microarray with sera from candidemia patients, blood-culture negative hospital patients and healthy individuals lead to the identification of serological signature specific for acute and convalescent stages of candidemia. Intriguingly, large proportions of the identified antigens are involved in oxidative stress, drug resistance and iron acquisition. Furthermore, strong IgG response to many proteins known to be induced and/or required for C. albicans invasion of epithelial and endothelial cells is observed in both candidemia patients and non-candidemia controls, including all healthy individuals. Our findings provide new insights into commensal colonization and pathogenesis of C. albicans, as well as the characterization of potential serodiagnostic antigens and vaccine candidates.
Hospital patient sera were collected from Shands Hospital at the University of Florida (UF) (SH-UF) from January 2004 to December 2006. We collected sera from 21 patients with candidemia where the etiological agent was C. albicans. The median time from the date of positive culture to serum collection was two days. The study population was classified by age, gender, underlying disease, portal of entry, antifungal received, and outcome of stay (Table S1). A subset of the candidemia patients was followed through acute infection (days 0–14) to early convalescent (week 4) and mid convalescent (week 12) infection. We also used sera from 12 hospital patients and 50 healthy individuals who had no evidence of candidiasis as our negative control groups.
C. albicans cell surface proteins were chosen for the protein microarray because they interact directly with the host and thus are likely important for colonization and infection, as well as likely targets for the host immune system. Furthermore, many of their protein expression levels are regulated in response to extracelluar signals, such as stress, nutrients, host factors, or changes in environment. Known antigenic proteins are also included as controls (Bgl2, Eno1, Pgk1, Gap1, Cdc19, Tkl1, Hsp90, and members of the Hsp70 family) [15],[17]. The collection contains 451 His- and HA-tagged peptides (Table S2) that represent 363 different proteins, since ORFs >3,000 bps were cloned into two or more segments. All tagged proteins were confirmed individually by western blot and again on the protein microarray.
We have used the C. albicans cell surface microarray to evaluate the antibody profile of patients with candidemia against healthy individuals and uninfected hospital patients to determine relevant cell surface antigens that correlate with infection. Arrays were probed with a collection of sera consisting of different stages of candidemia: acute, early convalescent (approximately 4 weeks after onset of infection) and mid convalescent (approximately 12 weeks after onset of infection), as well as uninfected hospital patients and healthy individuals. Figure S1 shows a representative image of the microarray hybridized with the serum of an acute candidemia patient. All hybridizations in this study were done under the same conditions and dilutions with protein microarrays printed from the same batch. Their serological reactivity is shown as a heatmap where the antigens are sorted by increasing normalized global mean intensity, with bright green having the weakest intensity, red being the strongest, and black in between (Figure S2). An examination of the IgG response to the entire C. albicans cell surface protein microarray showed that the mean global signal intensity was similar among different groups (data not shown), although antigenic profiles are not identical between individuals.
We were interested in determining the most seroprevalent antibodies in the acute candidemia patients and how their humoral response compared against the negative control groups. Antigens to the most seroprevalent antibodies were defined as serodominant antigens and characterized as having mean antigen reactivity 2-fold greater than the in vitro transcription/translation reaction mixture containing no vector. The top-forty serodominant antigens in the candidemia patients consisted of many previously characterized antigenic peptides such as Bgl2 [17], Tkl1 [15], Hwp1 [13],[23], Eft2 [15], and Cdc24 [13] (Table 1). Also among the top-forty serodominant antigens were many previously identified virulence-associated and/or hyphal-regulated proteins (eg. Int1, Hwp1, Als1, Als3, Als5, Ece1, Hyr1, Cdc24, and Utr2) (Table 1) [24]–[32]. Interestingly, this serological response of acute candidemia patients was shared with both uninfected hospital patients and healthy individuals. The mean signal intensity to the top-forty serodominant antigens was 8,825 in acute candidemia patients, 8,837 in uninfected hospital patients, and 10,790 in healthy individuals. A two-way hierarchical cluster analysis of the top-forty serodominant antigens shows that the serum specimens of both the positive and negative candidemia groups were randomly dispersed throughout the hierarchical tree (Figure 1A). To further confirm that the top-forty serodominant antigenic signatures are shared among acute candidemia patients, the uninfected hospital patients and healthy individuals, principal component analysis (PCA) was used to generate a three-dimensional projection of the data (Figure 1B, 1C and 1D). The PCA shows that a large proportion of both the positive and negative acute candidemia sera are clustered together. These analyses suggest that IgG levels to the top-forty serodominant antigens are similar in both the negative control groups and acute candidemia sera. Since many of the top-forty antigens are either important for or induced during the invasion of epithelial or endothelial cells [11],[33], their expression in healthy people, inferred from the presence of their antibodies, indicates the existence of a permanent host-pathogen interplay in immunocompetent individuals.
To determine stage-specific biomarkers of acute candidemia, the normalized serological expression of acute candidemia patients were compared against the humoral reactivity of the uninfected hospital patients and healthy individuals. Serodiagnostic antigens were defined as having an IgG response significantly greater in acute candidemia patients (days 0–14) as compared to the negative control groups with Benjamini and Hochberg (BH) adjusted Cyber-T p-values <0.05. Thirteen antigens met this requirement (Table 2). Moreover, among the proteins identified as serodiagnostic markers, proteins involved in oxidative stress response appeared to be enriched over other functional categories. Sln1 and Nik1 are two out of three histidine kinases on the cell surface protein microarray and they are both identified as serodiagnostic antigens. Sln1 and Nik1 are sensors for the high-osmolarity glycerol (HOG) pathway, a mitogen-activated protein kinase cascade responsible for osmotic and oxidative stress adaptation in C. albicans [34],[35]. In addition, the expression levels of CDR4, RAS2, and ALS9 are up-regulated during oxidative stress [35]. Another functional group over-represented among the serodiagnostic antigens are transporters associated with drug resistance (Cdr1, Cdr4, and Yor1) [36].
The 13-serodiagnostic antigens were also evaluated with a two-way hierarchical cluster analysis on candidemia positive and negative sera. Interestingly, the sera clustered into two distinct groups based on their responses to the 13 antigens (Figure 2A). Cluster I contained 10 candidemia sera and only one uninfected hospital patient. Cluster II contained all 50 healthy individuals, 11 of the 12 hospital patients, and 8 acute candidemia sera (Figure 2A). To further confirm that the antigenic signatures identified during the acute phase of candidemia differed from the negative control groups, PCA was used to create a three-dimensional projection of the data (Figure 2B, 2C, and 2D). In agreement with the two-way hierarchical cluster analysis, two distinct groups were observed (Figure 2B and 2C). Also, the PCA of the negative control groups showed individuals are clustered together with the exception of one outlying uninfected hospital patient found clustered with the acute candidemia patients (Figure 2C and 2D). These data provide further support of the antigenic signature of patients during the acute phase of candidemia. Multiple linear regression models determined that the antigenic profiles of acute candidemia patients were not related to various risk factors (i.e. age, gender, course of treatment, coexisting disease, and recovery/fatality) (data not shown). However, this determination is limited by the small sample size of our study.
Multiple independent serodiagnostic antigens can dramatically improve the sensitivity and accuracy of serodiagnostic tests [37]. To establish a collection of antigens that could be used as a multiplex set to accurately distinguish candidemia cases from controls, we studied the discriminatory power of different sets of proteins using receiver operating characteristic (ROC) curves. First, ROC curves were generated for individual serodiagnostic antigens and the area under the ROC curves (AUC) for each antigen is listed in Table 2. The top-five cell surface proteins all have an AUC greater than 0.76, with CDR1 (3) (AUC 0.87, BH adjusted Cyber-T p-value <1.04e-7) giving the best single antigen discrimination (Table 2). The 13th antigen has an AUC of 0.630, which still exceeds the upper 95% confidence interval for random expectations for the AUC. To extend the analysis to combinations of antigens, we used kernel methods and support vector machines to build linear and nonlinear classifiers. As inputs to the classifier, we used the highest-ranking AUC antigens in combinations of 2, 5, 10, 11, 12 and 13 proteins and the results were validated with 10 runs of three-fold cross-validation (Figure 2E). Increasing the antigen number from 2 to 5, and 5 to 10 produced improvements in the classifier. But as the antigens increased to 13, a reduction in accuracy was observed. Using the ten most significant diagnostic antigens (in rank order: Cdr1 (3), Cfl91, Cdr4 (3), Als9 (2), Cdc19, Nik1 (2), Chs8 (2), Rta4, Sln1 (2), and Trk1 (2)), the classifier predicts 83% (95% CI, 76–89%) sensitivity, 72% (95% CI, 68–76%) specificity, and 74% (95% CI, 72–76%) accuracy in diagnosis of acute phase candidemia from the negative controls (healthy individuals and uninfected hospital patients) (Table 3).
We were next interested in identifying antigens that are significantly different between the early/mid convalescent candidemia patients (weeks 4 and 12 of the infection, respectively) and the negative control groups. The convalescent patient sera consisted of three patients whose serum was drawn under all three disease phases (acute phase, early and mid convalescent phases), 4 patients who had blood drawn at the acute and early convalescent phases, and 3 patients whose blood was drawn only at the early convalescent phase. Using BH adjusted Cyber-T p-values <0.05, we identified 33 antigens, 11 of which are from the 13 diagnostic antigens for the acute phase of infection (Table 4). Among the identified convalescent biomarkers were marked expansions in proteins involved in iron acquisition (Rbt5, Csa1, Flc1, and Cfl91) (Table 4). Cfl91 is a putative ferric reductase similar to Fre10, which is required for the release of iron from transferrin and the reduction to ferrous iron [38]. The protein Flc1 has been identified as having heme uptake activity [39] whereas, both Rbt5 and Csa1 have been implicated as receptors of hemoglobin whose function is to deliver the hemoglobin by endocystosis to the vacuole where iron is released by acidification [40],[41]. The remainders of the identified proteins have roles in cell wall biogenesis, membrane lipid organization, and drug resistance.
We next evaluated antibody response to the 33 antigens in the acute, convalescent candidemia patients and the negative control groups by two-way hierarchical cluster analysis. The individuals in Cluster II were the same as those identified previously with 13 serodiagnostic antigens (Figure 2A and 3A) with the addition of one convalescent candidemia patient whose only sera was drawn during week 4 of the infection. Individuals in Cluster I consisted of candidemia patients with the exception of the one uninfected hospital patient from Figure 2A. Three of the candidemia patients' acute and convalescent profiles were all found in Cluster I, whereas four candidemia patients' profiles converted from Cluster II to I during the convalescence phase of the disease. In addition, the remaining two-candidemia patients whose only blood draws were during week 4 also grouped in Cluster I (Figure 3A). This conversion of the antigenic profile from the negative control groups (Cluster II) to the antigenic profile consistent with candidemia (Cluster I), indicates an adaptive immune response to C. albicans that is different from commensal sensitization. Again, PCA was used to further confirm that the antigenic signatures identified during the convalescent phase of candidemia differed from the negative control groups (Figure 3B, 3C and 3D). ROC curves were generated to assess the ability to separate the control and convalescent candidemia. AUC was determined for each of the 33-serodiagnostic antigens and listed in Table 4 in decreasing order. The top-five ORFs all have an AUC greater than 0.94. We then used SVMs to build multiplex classifiers with 2, 5, and 10 antigens with the highest-ranking AUC from Table 4. The results were validated with 10 runs of three-fold cross-validation (Figure 3E). Increasing the antigen number from 2 to 5 maintained the diagnostic accuracy in the classifier and a reduction in accuracy occurred as the antigens increased to 10 due to over-fitting. The top-five serodiagnostic antigens are associated with xenobiotic-transporting activity (Cdr4 and Yor1) [36], phospholipid-transporting activity (Drs23), a putative ferric reductase (Cfl91), and a mucin-like cell wall protein (Ipf25023) (Table 4). Using the top-five antigens, the classifier predicts 93% (95% CI, 89–96%) sensitivity, 96% (95% CI, 95–96%) specificity, and 95% (95% CI, 94–96%) accuracy in the differentiation of early/mid convalescent phase candidemia from the negative controls (healthy individuals and uninfected hospital patients) (Table 3).
Having identified 33 antigens that are correlative with convalescent candidemia in comparison to the negative control groups, we next wanted to determine the temporal change in IgG response to these 33 antigens during the transition from acute infection (AI), to early convalescent (EC), and mid convalescent (MC). A two-way hierarchical cluster analyses was performed on differential IgG responses to the 33 antigens in 3 patients with AI, EC and MC sera, and 4 patients with only AI and EC sera (Figure S3). A one tailed t-test was carried out to look for differences where the EC antigen intensity is significantly greater than the AI antigen intensity, possibly indicating the selection of a protective antibody response. We observed a significant increase in the IgG response from AI to EC in the following antigens, which are ranked according to their p-values: Apc5 (2) (1.12E-03), Drs23 (3) (1.23E-03), Vps62 (1.57E-03), Rad50 (1.83E-03), Ssu1 (3.17E-03), Yor1 (3) (5.33E-03), Ipf885 (5.33E-03), Pga4 (5.88E-03), Cdr4 (3) (7.22E-03), Cfl91 (2) (0.0231), Cyr1 (2) (0.0274), Ipf25023 (2) (0.0330), Gsl2 (2) (0.0374), Chs1 (2) (0.0393), and Snq2 (3) (0.0486). The identified antigens could potentially be efficacious vaccine candidates due to the fact that the IgG response is being positively selected over the course of infection.
In this study, we have developed a C. albicans cell surface protein microarray and profiled host humoral responses during conmmensal colonization and during the progression of candidemia. Thirteen novel serodiagnostic antigens were identified for differentiating acute candidemia from commensal sensitization and 33 antigens were found to discriminate convalescent candidemia from non-candidemia controls. The sensitivity and specificity for the identification of acute candidemia determined by the top 10 antigens from the set of 13 serodiagnostic markers are comparable to that obtained using the method of 2D-PAGE and immunoblots [17]. When using the top 5 antigens from the set of 33, both sensitivity and specificity are dramatically improved for convalescent candidemia. Pitarch et al. reported that the anti-Bgl2p IgG antibody levels mainly define the proteomic signature for candidemia patients [17]. In this study, Bgl2 is on the list of 33 diagnostic antigens from convalescent sera. Although it is classified as a serodominant antigen by acute candidemia sera, the BH-adjusted p-value of Bgl2 (0.116) is just above cutoff (0.05) to be considered as diagnostic by our definition, and the mean anti-Bgl2 antibodies in acute candidemia is higher than the mean in non-candidemia controls. Bgl2 is a glycoprotein and the glycan moieties on other b-1,3-glucanosyltransferases seem to contribute to antigenicity. Since our Bgl2 is expressed in vitro without any glycosylation, its antigenicity is likely different from the Bgl2 produced by C. albicans used in the 2D-PAGE immunoblots. The previously identified immunogenic heat shock protein 90 (Hsp90) is also one of 33 biomarkers for convalescent candidemia identified from this study. Hsp90 has been shown to elicit a protective humoral response [42],[43] and its antibodies are known to associate with patients that recover from candidiasis. The use of protein microarray technology allowed us to identify new diagnostic antigens that were missed by previous studies. The use of 2-D PAGE to accurately identify and separate clinical markers of candidemia from commensal sensitization is limited by the range in protein abundance and various properties associated with peptides such as their mass, isoelectric point, hydrophobicity, and post-translational modification, as well as the semi-quantitative nature of a Western [18]. Using a C. albicans cell surface protein microarray helped us overcome many of the technical difficulties found with traditional proteomics, since the expression level of recombinant-derived proteins vary by only a single log and the use of fluorescent-labeled antibodies allows for greater linearity, precision, and sensitivity in the quantitative measurement of the humoral response to C. albicans. One of the most beneficial aspects in the use of the protein microarray assay is its ability to detect significant differences in the IgG response that under traditional immunoblot conditions would be below the detectable threshold. However, a potential limitation to our study is that the microarray is based on recombinant peptides. Because of the cell free nature of our in vitro translated peptides, potential epitopes may have been lost due to miss folding and a lack of glycosylation, both of which may affect the conformational structure of the native protein. On the other hand, the removal of posttranslational modifications, such as glycosylation, from the peptides may have revealed hidden peptide epitopes only seen during a strong host immune response. A large collection of peptide epitopes may increase the specificity in diagnosis of infection. In support of this, our study has identified many new clinical biomarkers that are associated with differing states of interactions with the host as well as the characterization of potential new targets for therapeutics and vaccine candidates. To our knowledge, this is the first study using a protein microarray to analyze the serological response to an organism that is capable of existing as both commensal flora and an opportunistic pathogen in the human population.
Commensal colonization of C. albicans is common in humans and attenuated host immunity is a perquisite for the transition from commensal colonization to infection. Historically, it was believed that C. albicans switched from a commensal to a pathogen using distinct pathogen-associated genetic programs when the host immune status was altered. An intriguing review challenges this notion, Hube postulates that C. albicans exists in a permanent host-pathogen interplay where overgrowth and invasion is only observed under immunocompromising conditions[44]. The review puts forth two-models of a permanent infection strategy: (1) constitutive gene expression where attenuated immunity induces little or no change in the pathogenic profile of C. albicans or (2) a variable transcriptional profile where C. albicans expression is dependent on the stage- and tissue-specific interactions with the host. Our study indicates the existence of permanent host-pathogen interplay with variable gene expression over the course of infection. The serological response to the entire C. albicans cell surface protein microarray detected considerable homogeneity as well as differences in the patterns of antigens recognized among patients and healthy individuals. The majority of healthy individuals and uninfected hospital patients have moderate to strong IgG responses to many C. albicans cell surface proteins that have long been associated with virulence or hyphal-regulation (a hallmark of virulence in itself). In agreement with our protein microarray data, Naglik et al. observed similar levels of IgG titers to the hyphal wall protein Hwp1 in patients with oral candidiasis and asymptomatic mucosal infections as well as healthy culture-negative controls [23]. These serodominant cross-reactive antigens include adhesins such as Als1, Als3, Als5, Hwp1 and Int1 and hyphal-regulated genes such as Als3, Hwp1, Ece1, Hyr1, and Cdc24. Both functional groups are known to be important for invasion and virulence [45]. Among the identified serodominant antigens are many previously characterized immunogenic peptides such as Bgl2 [17], Tkl1 [15], Hwp1 [13],[23], Eft2 [15], and Cdc24 [13]. Intriguingly, the average signal intensities to the top-forty serodominant antigens are higher in the healthy individuals than the uninfected hospital patients and acute candidiasis patients (10,380 vs. 8,837 and 8,825, respectively). It is interesting to speculate whether the healthy individuals' IgG response limits colonization and overgrowth since many of the serodominant antigens are against adhesins. In particular is the strong humoral response to the integrin-like protein, Int1, which may play dual roles in limiting both intestinal colonization of the cecum and systemic invasion of deep tissue organs [46],[47]. Another interesting serodominant antibody response is to the protein Ece1, which has been shown to promote adhesion and is important for GI colonization[48]. ECE1 transcription is highly expressed during GI colonization and invasion of host tissue [33],[48]. However, one can not discount that the high IgG titer of colonized individuals may be due to a previous superficial infections such candidal vaginitis [49],[50].
The microenvironmental conditions during commensal colonization of the host may also play a role in the induction of the IgG response to certain cell surface proteins. Previous studies have evaluated characteristics common to the GI and/or vulvovaginal tract such as blood, hypoxia, iron restriction and weak acid as modifiers of gene expression [9], [51]–[53]. Intriguingly, the expressions of these genes share common features to the identified serodominant antibodies. Interestingly, genes transcriptionally up-regulated in blood (Als1, Als3, Hwp1, Ece1, Hyr1, and Bgl2) were serodominant and cross-reactive with both positive and negative candidiasis individuals, as were genes up-regulated under hypoxic conditions (Als1, Als3, Hwp1, Rbt5, Utr2, and Tos1), iron restriction (Int1, Rbt5, and Fet35), and weak acid (Crp1, Fet35, and Ipf9655) (Table 1). Furthermore, some of the serodominant antigens (i.e. Als3, Ece1, Hwp1, and Rbt5) have been shown to be induced during the invasion of epithelial or endothelial cells [11],[33]. Therefore, the expression of the serodominant antigens in healthy individuals indicates the existence of permanent host-pathogen interplay during commensal colonization. In addition, the presence of serodominant IgGs in all 50 healthy individuals suggests that commensal colonization is much more prevalent than previously reported.
One of the most challenging tasks in characterizing serodiagnostic antigens from C. albicans is the identification of discriminating peptides that can differentiate between commensal colonization and candidemia with high sensitivity and specificity. By profiling antibody response from patients with varying stages of candidemia against healthy individuals and candidemia-negative hospital patients, we have identified 13 diagnostic antigens for acute phase of candidemia and 33 for the early/mid convalescent candidemia. The serologic signature in candidemia patients likely reflects an alteration in the level of those proteins due to a change either in transcription and/or protein stability. Stage- and tissue-specific gene expression during the course of systemic infection is expected as C. albicans cells transition through differing microenvironments of the host. Among the 13 diagnostic antigens for acute candidemia, three are associated with drug resistance (Cdr1, Cdr4, and Yor1) [36]. The exposure to antifungal drugs in patients undergoing acute candidemia may have acted as an additional environmental stress that stimulates the expression of these antifungal drug transporters [54]. Intriguingly, two out of the 13 biomarkers are the osmosensors Sln1 and Nik1 for the HOG pathway that is responsible for osmotic and oxidative stress adaptation in C. albicans [34],[35]. The host-pathogen interaction commonly associated with oxidative stress is typically seen during phagocytosis by neutrophils, the initiating immune response to C. albicans overgrowth and infection. Furthermore, a study of global transcriptional responses to oxidative stress observed an increase in the transcriptional expression of CDR4 (4.1-fold), RAS2 (2.5-fold) and ALS9 (1.5-fold) [35]. Taken together, our data indicates a strong correlation between the IgG response to oxidative stress-related cell surface proteins and the initial cell-mediated immune response during acute candidemia. In further agreement, previous studies have shown that oxidative stress functions are primarily induced when C. albicans is initially exposed to human blood or following phagocytosis by neutrophils and granulocytes [7],[9],[10],[55]. The 33 convalescent diagnostic antigens include proteins involved in iron acquisition, cell wall biogenesis, membrane lipid organization, and drug resistance. Of particular interest is the dramatic increase in antibodies to proteins for iron acquisition (Cfl91, ferric reductase; Rbt5 and Csa1, hemoglobin receptors; and Flc1, heme uptake). Iron is an essential nutrient for C. albicans. Circulating iron in serum is bound to transferrin and ferric reductases are required in the acquisition of iron from transferrin. Interestingly, Cfl91 is found as a biomarker for both acute and convalescent candidemia patients. Of particular interest is the increase antibody response to hemoglobin and heme-related proteins as these molecules are normally sequestered in erythrocytes [56]. The proteins Rbt5, Csa1 and Flc1 are required for iron acquisition from hemoglobin or heme [39],[40] and are diagnostic antigens only for convalescent candidemia. Thus, it is interesting to speculate whether free hemoglobin becomes a by-product of lysed erythrocytes after post-operative surgery or other invasive clinical procedures. Nevertheless, the data from this study should provide critical information for the development of diagnostic antigenic profiles for patients at risk for candidemia and for the assessment of progression of hematogenously disseminated candidiasis. Future studies will need to be done to determine whether serological differences exist between superficial and systemic infections, as well as commensal sensitization.
The development of the antigenic profiles over the course of candidiasis (acute infection, early convalescence, and mid convalescence) may also provide insight into a protective humoral response against C. albicans. Even though previous sensitization to commensal colonization does not limit mortality or even morbidity in patients, experimental studies have identified protective antibodies against hematogenously disseminated candidiasis, such as heat shock protein 90 (Hsp90) or β-mannan [57]–[60]. Future studies will need to address whether the serodiagnostic antigens identified in this study could provide protection from hematogenously disseminated candidiasis. Of particular interest are the convalescent serodiagnostic antigens where the EC antigen intensity is significantly greater than the AI antigen intensity, which may possibly indicate the selection of a protective antibody response.
Human sera from candidemia patients and hospitalized patients were collected from SH-UF under protocols approved and created by the UF Institutional Review Board. Sera from healthy individuals were obtained from volunteers at the General Clinical Research Center at the University of California, Irvine. Written, informed consent was obtained from participants.
Candidemia was defined as the recovery of C. albicans from blood cultures. Sera from candidemia patients and hospitalized patients (no clinical or microbiological evidence of candidemia) were collected from SH-UF as previously published [61]. Briefly, patients at SH-UF were identified on the day blood cultures were positive for C. albicans. The Infectious Diseases Consultation Service at SH-UF identified controls. Sera were collected and stored at −70°C in the repository at the UF Mycology Research Unit. For patients with candidemia, sera were obtained from the earliest possible date on or after the date that the first positive cultures were drawn. In all cases, this was within 7 days of the first positive culture (acute-phase sera). For ten patients with candidemia, sera were also recovered 4 to 12 weeks after the date on which the first positive cultures were drawn (convalescent-phase sera).
Cell surface proteins were selected from the Candida Genome Database (CGD) using keywords such as “cell surface”, “plasma membrane”, and “cell wall”. The CGD annotation of cell surface proteins is based on published experiments [32], [62]–[66], function-based prediction of cellular localization, and sequence prediction. Known antigenic proteins are also included as controls (Bgl2, Eno1, Pgk1, Gap1, Cdc19, Tkl1, Hsp90, and members of the Hsp70 family) [15],[17]. Coding regions of the genes were PCR amplified from the clinical isolate SC5314 of C. albicans with primers listed in Table S2, and cloned into a pXT7 expression vector with a HA-tag at the N-terminus and His-tag at the C-terminus by homologous recombination in E. coli as described [67]. Protein expression was carried out using an E. coli based cell-free in vitro transcription/translation system (RTS 100 E. coli HY kit, Roche). The protein microarray was made by printing the peptides onto nitrocellulose-coated FAST glass slides (Schleicher & Schuell) using the OmniGrid 100 (GeneMachines) in the UCI Microarray Facility. Each peptide was printed in duplicate and showed homogenous spot morphology as well as low background. Internal controls consisting of buffer alone and a reaction mixture with no DNA were also printed onto the FAST slides. After the addition of the plasma samples the microarray was incubated with a biotin-conjugated donkey anti-human IgG Fcγ fragment specific secondary antibody (Jackson Immunoresearch). The secondary antibody was then removed and the microarray was incubated with Streptavidin: SureLight ® P-3 (Columbia Biosciences). Details concerning microarray construction and controls, antibody profiling, data normalization, as well as the reproducibility and validity of the microarray are given in the Text S1.
All analysis was performed using the R statistical environment (http://www.r-project.org). It has been noted in the literature that data derived from microarray platforms is heteroskedatic [68]–[70]. This mean-variance dependence has been observed in the arrays presented in this manuscript [71],[72]. In order to stabilize the variance, the vsn method [73] implemented as part of the Bioconductor suite (www.bioconductor.org) was applied to the quantified array intensities. In addition to removing heteroskedacity, this procedure corrects for non-specific noise effects by finding maximum likelihood shifting and scaling parameters for each array such that the variances of a large number (default setting used: 85%) of the spots on the array are minimized. In other words, the method assumes that variance in binding for the vast majority of the proteins on the array are due to noise rather than true differential immunological response. In essence, 85% of the spots on the array are used as controls for sample-by-sample normalization. This calibration method has been shown to be effective on a number of platforms [74]–[76]. A simple ranking normalization where all of the proteins are ordered for each sample by binding intensity and assigning the integer rank was performed as well with similar results (results not shown). Finally, VSN normalized data is retransformed with the ‘sinh’ function to allow visualization and discussion at an approximate raw scale.
Diagnostic biomarkers between groups were determined using a Bayes regularized t-test adapted from Cyber-T for protein arrays [69],[77]. To account for multiple testing conditions, the Benjamini and Hochberg (BH) method was used to control the false discovery rate [78]. Statistical analyses were performed with R 2.0 (www.r-project.org) and STATA (version 10.0, StataCorp). Multiple antigen classifiers were constructed using linear and non-linear Support Vector Machines (SVMs) using the “e1071” R package. To prevent overfitting and show the generalization of the classification method, 10 repeats of three-fold cross-validation were performed. In this methodology, the data is split into 3 class-stratified subsets. For each subset, a classifier is trained using the remaining two-thirds of the data. The classifier is then evaluated on the one-third of the data not used for training. This process is repeated for each split and for 10 different splits, yielding 30 evaluation measures. The ROCR package was used to construct receiver-operating-characteristic curves and perform sensitivity and specificity analyses. Blast2Go (www.blast2go.org) was used for gene ontology annotation and enrichment analysis. To confirm that the identified antigens were accurate, their vectors were resequenced. The Tables S3 and S4 list the statistical data of acute and convalescent candidemia patients, respectively.
Detailed information for the genes/proteins from this study can be found at the Candida Genome Database http://www.candidagenome.org. The gene names and ORF numbers are listed here: INT1 (19.4257), CWH41 (19.4421), PGA13 (19.6420), RBT5 (19.5636), HWP1 (19.1321), SLK19 (19.6763), YPS7 (19.6481), ALS3 (19.1816), CHS2 (19.7298), EFT2 (19.5788), IPF9655 (19.3988), GNP3 (19.7565), PHR3 (19.5632), ECE1 (19.3374), BGL2 (19.4565), PAN1 (19.19.886), OSH2 (19.5095), CRP1 (19.4784), PRY1 (19.2787), PGA60 (19.5588), UTR2 (19.1671), HNM1 (19.2003), HYR1 (19.4975), WSC4 (19.7251), CDC24 (19.3174), HYR3 (19.575), DNF2 (19.932), MEP2 (19.5672), GCA1 (19.4899), CWH43 (19.3225), FRE10 (19.1415), ALS5 (19.5736), ALS1 (19.5741), SLN1 (19.3256), FCY21 (19.1357), TOS1 (19.1690), FET34 (19.4215), TKL1 (19.5112), CDR1 (19.6000), CFL91 (19.1844), CDR4 (19.5079), ALS9 (19.5742), CDC19 (19.3575), NIK1 (19.5181), CHS8 (19.5384), RTA4 (19.6595), TRK1 (19.600), YOR1 (19.1783), CSC25 (19.6926), RAS2 (19.5902), DRS23 (19.323), IPF25023 (19.2296), ALS6 (19.7414), VPS62 (19.1800), SNQ2 (19.5759), IPF885 (19.7214), CAG1 (19.4015), HNM4 (19.2946), APC5 (19.6861), HSP90 (19.6515), CSA1 (19.7114), GSL2 (19.3269), PGA4 (19.4035), FLC1 (19.2501), CHS1 (19.7298), IPF22247 (19.4940), YCK22 (19.2222), SSU1 (19.7313), RAD50 (19.1648), and CYR1 (19.5148).
|
10.1371/journal.pgen.1006898 | Notch-dependent epithelial fold determines boundary formation between developmental fields in the Drosophila antenna | Compartment boundary formation plays an important role in development by separating adjacent developmental fields. Drosophila imaginal discs have proven valuable for studying the mechanisms of boundary formation. We studied the boundary separating the proximal A1 segment and the distal segments, defined respectively by Lim1 and Dll expression in the eye-antenna disc. Sharp segregation of the Lim1 and Dll expression domains precedes activation of Notch at the Dll/Lim1 interface. By repressing bantam miRNA and elevating the actin regulator Enable, Notch signaling then induces actomyosin-dependent apical constriction and epithelial fold. Disruption of Notch signaling or the actomyosin network reduces apical constriction and epithelial fold, so that Dll and Lim1 cells become intermingled. Our results demonstrate a new mechanism of boundary formation by actomyosin-dependent tissue folding, which provides a physical barrier to prevent mixing of cells from adjacent developmental fields.
| During development, boundary formation between adjacent developmental fields is important to maintain the integrity of complex organs and tissues. We examined how boundaries become established between adjacent developmental fields—which are defined by expression of distinct selector genes and developmental fates—using the Drosophila eye-antennal disc as a model. We show that boundary formation is a progressive process. We focused our analysis on the antennal A1 fold that separates the A1 and A2-Ar segments, corresponding to the evolutionarily conserved segregation between coxopodite and telopodite segments of arthropod appendages. We describe a clear temporal and causal sequence of events from selector gene expression to establishment of a lineage-restricting boundary. We found that Notch activation at the boundary between adjacent fields of selector gene expression triggers actomyosin-mediated cell apical constriction, which induces the formation of an epithelial fold and prevents intermixing of cells from adjacent fields. Our findings describe a novel mechanism by which epithelial fold provides a physical barrier for cell segregation.
| During development, an organism is progressively divided into discrete fields that develop into different organs or parts of an organ. In many cases, the adjacent developmental fields develop distinct morphological, functional and molecular characteristics and are often divided by a sharp boundary that function to prevent lineage-related cells originating from one compartment from crossing into the adjacent compartment. Such lineage-restricting boundaries were first described in the fruitfly Drosophila wing and the milkweed bug Oncopeltus abdomen, using mitotic clones and cuticle markers to trace lineage distributions [1, 2]. The same phenomenon was then reported for other parts of the fly body and in vertebrates [3–10]. Nevertheless, not all boundaries have been analyzed for lineage restriction at single cell resolution.
Compartment boundaries generally coincide with the expression borders of the selector genes that determine the fates of developmental fields. For example, in the fly wing disc, the anterior-posterior (A/P) boundary correlates with the border of engrailed (en) expression in the posterior compartment, whereas the dorsal-ventral (D/V) boundary correlates with the border of apterous (ap) expression in the dorsal compartment. The expression domain of the selector genes does not begin as a sharply defined pattern (e.g. [11]), and usually evolves from a weak and fuzzy to a strong and sharply defined pattern through positive and negative regulation with other genes. Mutual repression between two selector genes, either direct or indirect, can force a cell at the expression border to express only one of the two selector genes. However, the cell-autonomous cell fate may result in a rough border of two cell types. A smooth and sharp alignment may require additional mechanisms to coordinate the cells at the expression border. Hence, the expression border and the lineage-restricting boundary are two phenomena characterized by different, though coinciding, processes. Therefore, the relationship between gene expression borders and lineage-restricting boundaries needs to be considered with respect to their temporal progression. We define ‘boundary’ as indicating lineage restriction, ‘compartment boundary’ to indicate absolute lineage restriction, ‘field boundary’ for incomplete lineage restriction, and ‘border’ to refer to expression domains.
Three types of mechanisms have been shown to play a role in boundary formation and maintenance. First, differential cell affinities modulated by cadherin interactions are responsible for various boundary formations [12–15]. Second, reduced cell proliferation found at the vertebrate somite and Drosophila D/V boundary can minimize movements resulting from mitosis [16–18]. However, whether reduced cell proliferation or bias in mitosis orientation is important for the maintenance of the boundary is unclear [11, 19, 20]. Third, mechanical forces provided by the intracellular cytoskeletal network can sharpen boundaries in both the vertebrate and invertebrate system [11, 19, 21–30]. For instance, actomyosin cables are responsible for cell partitioning in Drosophila A/P and D/V boundaries, as well as zebrafish rhombomeric boundaries [19, 21, 25, 26]. Actomyosin cables bind to adherens junctions to form belt-like supracellular structures [31, 32]. These cables are enriched for cells along the boundary, serving as physical barriers that restrict cells in adjacent compartments from mixing, with or without morphological changes [19, 25–27].
We used the larval eye-antenna disc (EAD) to explore the mechanism of boundary formation in the Drosophila head, with an emphasis on the boundaries in the proximal-distal (P/D) axis, i.e. the boundary between the antennal segments. The EAD is a sac-like tissue composed of monolayered epithelial cells covered by peripodial cells. It contributes to the majority of the adult head organs, including compound eyes, antennae, ocelli, maxillary palps and the head cuticle (Fig 1A). These organs abut each other, with smooth and clear boundaries. The antenna is further divided into six segments, A1-A5 and the most distal arista (Ar). Patterning of P/D antennal segments by critical transcription factors is achieved by hedgehog (hh)-dependent decapentaplegic (dpp, in dorsal) and wingless (wg, in ventral) inductions [33]. In the center and marginal antennal disc, which are destined to be the distal and proximal antennal segments, respectively, Distal-less (Dll) and homothorax (hth) are activated upon high and low levels of Dpp and Wg [33–36]. Cells that coexpress hth and Dll become the A2 to A4 segments [37]. The LIM-homeodomain protein Lim1, which is regulated by EGFR signaling, specifies the A1 and Ar segments [38–40].
Here, by examining the temporal sequence of Dll and Lim1 gene expressions, lineage restriction, and tissue morphogenesis, we report that the boundary separating the most proximal segment (Lim1-expressing, A1), from the more distal parts (Dll-expressing) of the antenna involves a Notch-dependent downregulation of bantam microRNA and de-repression of Enable (Ena). Strikingly, this pathway produces an epithelial fold that not only acts as a boundary to ensure cells stay within their respective fields, but also reinforces Notch signaling, thereby safeguarding boundary integrity. Thus, our results have uncovered a novel mechanism for the establishment of a field boundary that involves the formation of folded epithelial structures.
The EAD undergoes a series of progressive epithelial folds from the early third instar stage (e-L3, S1A Fig). EAD cells are cuboidal in the early second instar stage (e-L2). From the late second instar (l-L2) (S1B Fig), epithelial cells in the antennal and eye fields become columnar, but medial cells remain cuboidal and have a concave morphology in lateral view (S1B and S1B’ Fig). During e-L3, a ring fold (hereafter termed the ‘A1 fold’) is formed to separate the prospective A1 antennal segment from the distal A2-Ar antennal segments (S1C Fig). Also during e-L3, an E/C fold that separates the eye and head cuticle partially extends from the lateral to medial regions (S1C–S1C Fig), becoming complete by the late third instar (l-L3) (S1D–S1D” Fig). A fold that separates the most distal arista segment (Ar, termed the ‘Ar fold’ hereafter) and the other antennal segments forms during l-L3 (S1D Fig). In the l-L3 antennal disc, the A1 fold correlates with the border separating the Dll and Lim1 expression domains (Fig 1B). Dll is expressed in the A2-Ar segments, whereas Lim1 is specifically expressed in the A1 segment and the head cuticle. In the mid second instar (m-L2) EAD (Fig 1C, dashed line, 1F), before the A1 fold has been formed, Dll and Lim1 expressions are weak and partially overlap (co-expression), exhibiting a fuzzy border due to two to three rows of cells co-expressing Dll and Lim1. From l-L2 (Fig 1D and 1G) to e-L3 (Fig 1E and 1H), levels of Dll and Lim1 gradually increase and become sharply confined. At e-L3, the border between the Lim1 and Dll expression domains sharpens and the genes are rarely co-expressed (S2A–S2C Fig). The sharp cell-autonomous segregation of Dll and Lim1 expression begins before formation of the A1 fold, suggesting that the epithelial fold is not the cause of segregated the expression.
The distal A2-Ar segments specified by the Dll gene correspond to the evolutionarily conserved telopodite in arthropod appendages. Therefore, the A1 fold separates the proximal coxopodite from the distal telopodite. We hypothesize that the folded tissue architecture at the A1 fold may act as a lineage-restricting boundary between the proximal Lim1-dependent coxopodite and the distal Dll-dependent telopodite.
Next, we tested whether the A1 fold serves as a lineage-restricting boundary. The classical definition of a compartment boundary in Drosophila depends on cuticular markers (e.g. yellow (y) and multiple wing hair (mwh)) for wing, leg and antenna, or pigmentation (white, w) for compound eye. These markers can only be used on adult tissues. No single marker can be used for both eye and other head structures. We used Twin-Spot MARCM (TSM) to induce sister clones with different fluorescent proteins [41]. The fluorescent markers allowed analysis of clone distribution covering the entire head structure of both larval and adult stages (S3 Fig). Pairing of the sister clones allowed us to determine if a clone was indeed from a single origin. The TSM clones were induced at indicated time-points, and their distributions were analyzed in l-L3 discs (S3C–S3F Fig, and Fig 2). TSM clones in wing and antennal discs determined the timing of A/P and D/V boundary formation (S3C–S3F Fig). For example, in the wing disc, clones induced in L2 cross the D/V boundary (marked by Cut-expressing cells) but those induced at e-L3 do not, indicating that the D/V boundary is formed at e-L3 and not L2 (S3C–S3D Fig). These results are consistent with previous reports, and validate our TSM clonal analysis for the study of lineage restrictions.
We examined the distribution of the TSM clones relative to the A1 fold. All clones at the folds were examined in different focal planes to check whether they crossed or were restricted by the A1 fold. Even clones for which 1–2 cells crossed the A1 fold were counted as having crossed it. Therefore, our clonal analysis is defined by a very sharp border at single cell resolution. When clones were induced at m-L2, all except one of the TSM clones crossed the A1 fold (Fig 2A, red arrow; 2E; 2F, 2.94% restricted by boundary). The frequency of clones that were restricted by the epithelial fold increased when clones were induced at l-L2 (Fig 2B and 2C, blue arrow; 2E; 2F, 18.92% restricted by boundary), and they occurred at the A1 fold and the lateral part of the E/C fold (Fig 2E, blue triangle). Most of the clones induced at e-L3 were restricted by the A1 fold (Fig 2D, blue arrow; 2E; 2F, 85.71% restricted by boundary), and crossed the Ar fold (Fig 2E, red cross; the Ar fold forms in m-L3). The E/C boundary was established progressively, laterally to medially (S1C–S1D Fig, and Fig 2E) because, at e-L3, most lateral clones were restricted by this boundary (6/7), but the medial clones crossed it (2/2).
Our TSM clonal analysis showed that the A1 boundary is not an absolute lineage-restricting boundary. Even if we count clones with a single cell crossing as having been restricted by the A1 fold boundary, the frequency of e-L3 clones respecting the it is less than 100% (88.6% for TSM clones). In contrast, clones induced during the first instar (L1) absolutely respected the A/P boundary in wing disc at a single cell resolution (S3F Fig, marked by Patched, Ptc, 31/31). Since the A1 fold boundary does not fit the classical definition of a compartment boundary, we term it a ‘field boundary’ to differentiate it from a compartment boundary. In summary, lineage restriction at the A1 fold correlates temporally with the formation of the epithelial fold. This supports our hypothesis that the epithelial fold serves as a lineage-restricting boundary.
Since the epithelial fold strongly correlated temporally and spatially with the establishment of lineage restriction and the gene expression border, we investigated the process of epithelial fold in the EAD development. Previous studies have shown that apical actomyosin triggers apical constriction to initiate fold [42–44]. Spaghetti-squash (Sqh)—a non-muscle myosin regulatory light chain—is a key component of the actomyosin network [45]. Therefore we examined whether the EAD fold arises from apical constriction by live imaging ex vivo-cultured Sqh-GFP from l-L2 EAD (Fig 3A–3C) [46]. Based on their dynamics in the apical area, three groups of cells could be distinguished; namely, constant, fluctuating and decreasing cells (Fig 3B and 3C). Cells exhibiting a significant decrease in apical area coverage over the 5-hour period were located primarily along the A1 fold (Fig 3B and 3C). Fluctuating cells in the apical area were scattered close to the A1 fold (Fig 3B). Cells constant within the apical area were located further away from the A1 fold (Fig 3B and 3C). The extent of EAD apical area reduction is similar to that described for embryonic cells in mesoderm formation (Fig 3D) [42]. Cell height and volume before (l-L2) and after (e-L3) A1 fold formation were measured from fixed EAD for better Z resolution (Fig 3E, details in S1 Table). For cells in the A1 fold, the heights of the apical domains (defined by aPKC) of folded cells were similar to non-folded cells, but the apical volumes were significantly smaller (20% those of non-folded cells), likely due to constriction of the apical area (Fig 3C and 3D). For the basolateral domains (defined by FasIII) of cells in the A1 fold, height and volume were both lower (by 50%) than for non-folded cells, but the difference were not as drastic as for apical volumes and dimensions (Fig 3C–3E). Cells surrounding the A1 fold (i.e. 1 or 2 rows away from the A1 fold) were slightly taller and larger than folded cells, but these dimensions were still less than those for non-folded cells.
Sqh protein is distributed as junctional and medial-apical species in the EAD. Junctional Sqh was present in all cells, whereas medial-apical Sqh was observed in cells undergoing apical constriction (Fig 3F, arrow, and S1 Movie). Medial-apical Sqh accumulated periodically as apical size decreased (S1 Movie), probably through a mechanism similar to that reported to drive cell invagination during mesoderm formation [42, 43]. In cells at the A1 fold, junctional Sqh was uniformly presented (Fig 3G, marked by stars), unlike the cable-like structure of actomyosin that is enriched at opposing interfaces of cells along the A/P boundary [27]. Mitotic cells were frequently observed in the A1 fold (Fig 3H, arrow), suggesting that lineage restriction at the A1 fold is not likely due to a zone of quiescent cells.
We next tested whether actomyosin is responsible for the apical constriction and formation of the A1 fold. Actomyosin is composed of actin, non-muscle myosin II heavy chain (Zipper, Zip), and regulatory light chain (Spaghetti-squash, Sqh). Spaghetti-squash activator (Sqa) is a myosin like chain kinase (MLCK)-like kinase required for non-muscle myosin activation [47]. Both zip2 and sqaf01512 mutant clones at the A1 and Ar folds showed reduced fold (S4A–S4C Fig, compare yellow and white arrows), while maintaining apical-basal polarity. Larger mutant clones showed apical swelling and/or delamination, as has been previously reported (S4D Fig) [48].
We then examined if lineage restriction is affected when epithelial fold is disrupted. Lim1- and Dll-expressing cells are well segregated in the L3 antenna (Fig 1E). In zip2 MARCM and sqaf01512 clones that span the A1 fold, mixing of Dll- and Lim1- expressing cells was observed within the clones (Fig 4A and 4B, 18/23 in zip2 and 11/15 in sqaf01512) and occasionally outside of the clones (S4E–S4E’ Fig, arrow), implying a breakdown of the boundary. zip2 or sqaf01512 clones located exclusively within the A1 or A2-Ar domains did not show altered expression of Lim1 or Dll (S4F Fig), suggesting that cell mixing in these mutant clones was not due to altered cell fates, but to loss of positional restriction. The mislocalized cells were maintained in the epithelial sheet and were not sorted out basally for elimination (S4G Fig). Cleaved caspase 3 in the mutant larval EAD was rarely detected (S4H Fig). Together, these finding imply that the cell mixing phenotype may be observed in adults. It was difficult to observe cell mixing between antennal segments in the adult head. However, adult heads with zip2 or sqaf01512 clones consistently showed mislocalized ommatidia in head cuticle and antennae (Fig 4C and 4D, highlighted in red), and antennal-like tissue at the borders of compound eyes (Fig 4E), indicating a breakdown of the E/C boundary, which is also characterized by an epithelial fold (Fig 2E). We occasionally observed necrotic scar-like cells in zip2 or sqaf01512 adults (Fig 4F), suggesting that some elimination of mutant cells takes place during or after the pupal stage. Knocking down (KD) of zip and sqh by hth-GAL4 from the L2 stage, which covers the A1-A3 region in the antennal disc (Fig 4G) [36], revealed disorganization and mixing of Lim1 and Dll cells (Fig 4H and 4I, frequency of cell mixing in zip KD: 100%; sqh KD: 95.5%). Again, these mislocalized cells were properly integrated in the epithelial sheet for both zip and sqh knockdown (S4I–S4J Fig). We observed Dll cells in the Lim1 field (Fig 4I and S4G–S4I Fig) and Lim1 cells in the Dll field (Fig 4A, 4B, 4H and S4J Fig), so mislocalization between different fields due to disruption of the A1 fold is reciprocal. Collectively, these results support a role for the epithelial fold in acting as a boundary to separate different cells in the EAD.
We also tested a number of proteins known to interact with actomyosin and that are involved in boundary formation to clarify their roles in the A1 fold formation. Knockdown by hth-GAL4 of the basal focal adhesion components integrin (encoded by myospheroid, mys) and talin (encoded by rhea) (S5A–S5D Fig), the Hippo-regulating LIM protein Ajuba (jub) (S5E–S5F Fig) [49], and the adherens junction component Echinoid (ed) [50] (S5G–S5H Fig) did not affect formation of the A1 fold or segregation of Lim1 and Dll cells. These results show that integrin, talin, Ajuba, and Ed are not likely to be involved in A1 boundary formation. However, these mutant cells showed various morphological defects (e.g. swelling, enlargement or delamination) to a similar extent as zip2 cells (compare S4D Fig to S5I–S5L Fig; quantitation in S5M Fig). Our results imply that drastic changes in cell shape per se do not affect cell segregation at fold-mediated boundaries. Therefore, the mixing of Dll and Lim1 cells in zip, sqh, and sqa mutants is not due to altered cell size or morphology, but due to disruption of the epithelial fold.
To assess the effect of acute blockage of Sqh on the formation of the A1 boundary, we used chromophore-assisted laser inactivation (CALI) [27, 51, 52] to specifically inactivate Sqh-GFP in ex vivo e-L3 EAD. Indeed, Sqh-GFP inactivation by CALI caused a significant reduction in the extent of epithelial fold in the A1 fold (Fig 5A–5A’, compare aPKC and Coracle signals in the boxed regions for CALI and control). In contrast, the same CALI treatment on Moe-ABD::GFP did not cause a similar effect (Fig 5B–5B’), indicating the high specificity of CALI [27, 53]. Clones expressing RFP were induced at L2. Cells adjacent to, but not including RFP-labeled clones, were subjected to CALI treatment (S2 Movie). When CALI was applied to the A1 fold (Fig 5C, yellow boxed region), a few cells from an adjacent RFP clone crossed the disrupted A1 fold to the adjacent field (S2 Movie). Cells that crossed the A1 fold still maintained their Dll expression (yellow arrow in Fig 5C and 5D), indicating that cell fate had not changed (at least for the time span of our observations).
Due to the EAD curvature, some RFP-labeled cells from the peripodial membrane appeared in several time points that may confuse the observation (white arrow head in Fig 5C”; cross section in 5F and 5G,). Individual cell tracking was performed over time to unambiguously show border crossing (S2 Movie, overall trajectory in Fig 5E). The CALI inactivation of Sqh-GFP only lasted less than 5–6 hours, after which the endogenous Sqh-GFP expression was recovered and the A1 fold was reformed (S6A Fig, arrow). Hence the time window for RFP cell across boundary was less than the first 6 hours post CALI treatment (Fig 5E, S2 Movie). In the EAD ex vivo culture, we also noticed some small nuclei appeared at later time points (from post CALI 10h), probably result from impaired growth and gross morphological changes under ex vivo condition [46]. Indeed, the EAD cultured for more than 12 hours showed notable cellular architectural and morphological alterations, which was not observed at 6h post-CALI time point (compare S6A–S6B Fig). Therefore, the EAD deterioration after long-term culture is unlikely to contribute to border crossing. RFP clones in the EAD without CALI treatment was unable to cross the A1 fold (S3 Movie). We also analyzed the trajectory of RFP clones that crossed (CALI) or not crossed (non-CALI) the A1 fold (S6C Fig). The orientation and displacement are comparable between crossed and not crossed cells, indicating that CALI treatment did not cause significant side effects and further lead to additional behavior changes.
Since Notch (N) signaling is involved in the wing D/V boundary, mediated through intercellular actomyosin cables, we checked whether N might also be involved in formation of the A1 fold. N activity, based on anti-Nintra and N reporters E(spl)mβ-lacZ and Su(H)Gbe-lacZ, showed ring like patterns in l-L2 antenna discs (Fig 6A, 6C and 6E). In e-L3, N activity was enhanced in the A1 fold (Fig 6B and 6D). We examined in detail the relative timing of segregated expression of Dll, Lim1 and the N ligands, Delta (Dl) and Serrate (Ser) in l-L2. Disc sizes in groups 1, 2, and 3 (see material and methods) were 4835 ± 328, 6058 ± 231, and 7065 ± 309μm2 (mean ± stdev), respectively. Dl was mostly expressed in the central region, whereas Ser was expressed at the periphery of discs (S7A–S7C Fig cross-sections). These patterns echo the expressions of Dll and Lim1 during l-L2, respectively (Fig 6F). Lim1/Dll segregation is apparent in group 1, whereas segregated expression of Dl/Ser begins later in group 2 and is more pronounced in group 3 (Fig 6F; S7A’–S7C’ Fig). Fringe (Fng) is a glycosyltransferase that can modulate the interaction between N and its ligands [54]. Timing of fng-lacZ differential expression correlated with Dl/Ser segregation (Fig 6F and S7D’–S7F’ Fig). We further analyzed expression levels of these proteins in single cells from different groups. The correlation of Dll/Lim1 segregation and Dl/Ser segregation within single cells increased significantly with increasing disc size (Fig 6G). Taken together, these results suggest that sharp segregation of Dll and Lim1 expression precedes the differential expression of Ser, Dl and Fng, and thereby define the localization of N activation.
N is activated during e-L3 at the A1 fold, hence we tested whether N signaling is responsible for the epithelial fold formation. An N dominant-negative (NDN) mutant was expressed by hth-GAL4 and examined at l-L3. The A1 fold of this mutant was disrupted with high penetrance (Fig 7A, arrow, 86%) and was always accompanied by mixing of Dll and Lim1 cells (Fig 7A’). NDN clones that did not span the A1 fold presented normal Dll and Lim1 expressions (Fig 7B–7B’), indicating that reduced N activation did not alter cell fates. The cell morphology of ex vivo-cultured hth>NKD EAD was monitored by Sqh-GFP for more than 5 hours (Fig 7C–7D, compare to Fig 3B). These EAD failed to form the A1 fold. Most antennal cells exhibited a fluctuating apical area (Fig 7C, blue cells; 7D, quantitation), and a few scattered cells underwent apical constriction (Fig 7C, red cells). Since the N ligands Dl and Ser are differentially expressed in the Dll and Lim1 domains, respectively, we generated DlRevF10 SerRX82 double mutant clones so that for any clone spanning the A1 fold, no N ligand could activate Notch. Indeed, in such mutant clones, the A1 fold failed to form (Fig 7E, compare white and yellow arrows). Cells in the NDN clones showed less apical constriction (Fig 7G), and the cell volumes of their apical and basolateral domains were similar to those of non-folded cells (Fig 7H, compare with Fig 3E). These results indicate that N signaling is required for the formation of the A1 fold. In contrast, clonal expression at l-L2 of constitutively-activated N (Nact, the Notch intracellular domain [55]) caused ectopic tissue fold when located in a non-fold region (Fig 7F–7F’, arrow, 76%). The cells at the ectopic fold showed apical constriction and reduced apical and basolateral volumes similar to cells at the A1 fold (Fig 7G, 7H, and Fig 3E). The reduced cell volume in Nact cells is likely due to shrinkage, since the volume of these cells in L3 (Fig 7H) is smaller than normal cells in L2 (Fig 3E). Prolonged N activation did not cause further changes in cell volume (Fig 7H, compare 48h and 72h), suggesting that these drastic changes in cell morphology were stable upon N activation. Together, these loss-of-function and gain-of-function results show that N signaling drives the formation of stable tissue folds in the antennal disc.
N signaling is important for the establishment of the D/V boundary in wing disc [56, 57]. There, it represses the micro-RNA bantam, which itself represses its target Enabled (Ena), that is a positive regulator of actin polymerization. By repressing bantam, N enhances Ena expression, thereby establishing the actomyosin cable-based D/V boundary [19]. We assessed endogenous bantam level by RNA in situ hybridization in combination with a N activity reporter, Su(H)Gbe-lacZ, and Ena to study their relative expressions in the EAD. The bantam RNA in situ signals recapitulated the patterns reported previously in the wing D/V boundary [19] (S8A Fig). In l-L2 antennal discs, bantam and Ena levels were generally low, with little correlations with Su(H)Gbe-lacZ level (S8C Fig). In e-L3, a relatively lower bantam level was observed in the A1 fold region, whereas Su(H)Gbe-lacZ and Ena level were both elevated (S8D and S8D” Fig arrows). bantam-overexpressing clones showed significantly reduced Ena levels and inhibited EAD fold (Fig 7I), as well as mixing of Lim1 and Dll cells (Fig 7J). The bantam-overexpression clones within a single field did not exhibit altered Lim1 and Dll expression, indicating that the cell mixing phenotype was not the result of a changed cell fate (Fig 7K). Concomitant blocking of N signaling (through NDN) and a reduction of bantam (by expressing bantamsponge) in hth>NDN+bantamsponge mutants rescued the disrupted A1 fold and the lineage mixing phenotype (Fig 7I and 7K, 23% phenotype, compared to 81% in hth>NDN in Fig 7A). The Notch/bantam axis has been shown to regulate cell proliferation and apoptosis [58, 59]. We further tested if such regulation also exists and may potentially affect A1 fold formation. Mitosis (phospho-Histone H3) and apoptosis (cleaved caspase 3, S8E and S8F Fig) were examined in NDN or bantam overexpression mutants driven by dpp-GAL4 from L2 (dppL2). Cell proliferation was reduced by about 30% in both mutants, whereas there were no significant changes in apoptosis. In contrast to the nearly complete absence of proliferation reported in the DV boundary of wing disc [17, 60], this 30% reduction may not significantly affect the formation of epithelial folds. Our results suggest that N acted through bantam and Ena (possibly by repressing bantam to allow Ena expression) to induce actomyosin assembly and thus epithelial constriction and formation of the A1 fold.
Even after formation of the A1 fold, N activity is sustained during e-L3 (Fig 6D). We found that blocking epithelial fold, by knock down of zip and sqh in the dpp expression domain that spans the dorsal A1 fold reduced the expression level of the N reporter Su(H)Gbe-lacZ (Fig 8A–8C; 8G, quantitation). Interestingly, levels of the same N reporter were not affected in conditions where D/V boundaries were disrupted in wing discs (Fig 8D–8F; 8G). This suggests that N activity is sustained by the epithelial fold, possibly representing positive feedback regulation.
In this study, we tried to unravel the molecular and cellular mechanisms of boundary formation in the Drosophila head. We focused our analysis on the antennal A1 fold that separates the A1 and A2-Ar segments. Our results showed that the expression of the selector genes Lim1 and Dll, which are expressed in A1 and A2-Ar, respectively, was sharply segregated. This step was followed by differential expression of Dl, Ser and Fng, as well as activation of N signaling at the interface between A1 and A2 (Fig 9). N signaling then induced apical constriction and epithelial fold, possibly through repression of bantam to allow levels of the bantam target Ena to become elevated, with this latter inducing the actomyosin network. The actomyosin-dependent epithelial fold then provided a mechanical force to prevent cell mixing. When N signaling or actomyosin was disrupted, or when bantam was overexpressed, the epithelial fold was disrupted and Dll and Lim1 cells become mixed. Thus we describe a clear temporal and causal sequence of events leading from selector gene expression to the establishment of a lineage-restricting boundary.
Sharp segregation of Dll/Lim1 expressions began before formation of the A1 fold, suggesting that fold formation is not the driving force for segregation of Dll/Lim1 expression. Instead, the fold functions to safeguard the segregated lineages from mixing. Whether Dll/Lim1 segregated expression is due to direct or indirect antagonism between the two proteins is not known.
Actomyosin-dependent apical constriction is an important mechanism for tissue morphogenesis in diverse developmental processes, e.g. gastrulation in vertebrates, neural closure and Drosophila gastrulation, as well as dorsal closure and formation of the ventral furrow and segmental groove in embryos (see reviews [61, 62]). Our study describes a new function of actomyosin, i.e., the formation of lineage-restricting boundaries via apical constriction during development.
This actomyosin-dependent epithelial fold provides a mechanism distinctly different from other known types of boundary formation. We found that the cells at the A1 fold still undergo mitosis, suggesting that mitotic quiescence is not involved. Perhaps epithelial fold as a lineage barrier is needed in situations in which mitotic quiescence does not happen. Mechanically and physically, epithelial folds could serve as stronger barriers than intercellular cables when mitotic activity is not suppressed. The drastic and sustained morphological changes, including reduced apical area and cell volume, may be accompanied by increased cortical tension of cells along the A1 fold [63, 64], with such high interfacial tension then preventing cell intermingling and ensuring Dll and Lim1 cell segregation [30, 65]. Although similar to actomyosin boundaries, the epithelial fold in the A1 boundary is distinctly different from the supracellular actomyosin cable structure in fly parasegmental borders, the wing D/V border, and the interrhombomeric boundaries of vertebrates [19, 25–27] (see review [66]). The adherens junction protein Ed, which is known to promote the formation of supracellular actomyosin cables [50], is not involved in A1 fold formation (S5G Fig). Although actomyosin is enriched in a ring of cells in the A1 fold, it does not exert a centripetal force to close the ring, unlike the circumferential cable described in dorsal closure and wound healing (see review [67]). In the A1 fold, the constricting cells become smaller in both their apical and basolateral domains, thus differing from ventral furrow cells where cell volume remains constant [68, 69].
A tissue fold probably provides a strong physical or mechanical barrier to prevent cell mixing. In addition, whereas in a flat tissue where the boundary involves only one to two rows of cells, the tissue fold involves more cells engaging in cell-cell communication. The close apposition of cells within the fold may allow efficient signaling within a small volume [70]. This may be an evolutionarily conserved mechanism for boundary formation that corresponds to stable morphological constrictions such as the joints in the antennae and leg segments (see below).
Although N signaling has been reported to be involved in many developmental processes, a role in inducing actomyosin-dependent apical constriction and epithelial fold is a novel described function for N. For the A1 boundary, N activity is possibly mediated through repression of bantam and consequent upregulation of Ena. In the wing D/V boundary, N signaling is also mediated through bantam and Ena, but the outcome is formation of actomyosin cables, i.e., without apical constriction and epithelial fold [19]. Thus, the N/bantam/Ena pathway for tissue morphological changes is apparently context-dependent.
Tissue constriction also occurs later in joint formation of the legs and antennae. N activation also occurs in the joints of the leg disc and is required for joint formation [71–74]. This role is conserved from holometabolous insects like the fruitfly Drosophila melanogaster and the red flour beetle Tribolium castaneum [75] to the hemimetabolous cricket Gryllus bimaculatus [76]. It is possible that for segmented structures that telescope out in the P/D axis, like the antennae, legs, proboscis and genitalia, N signaling is used to demarcate the boundaries between segments, which are characterized by tissue constriction. N-dependent epithelial fold morphogenesis has also been reported in mice cilia body development without affecting cell fate [77], suggesting that such N-dependent regulation in morphogenesis is evolutionarily-conserved.
We propose that N signaling is important in all boundaries that involve stable tissue morphogenesis. For those boundaries corresponding to stable morphological constrictions, e.g. the joints in insect appendages, N acts via actomyosin-mediated epithelial fold. The wing D/V boundary represents a different type of stable tissue morphogenesis. It becomes bent into the wing margin and involves N signaling via actomyosin cables, rather than apical constriction. In contrast, actomyosin-dependent apical constrictions do not involved N signaling and are involved in transient tissue morphogenesis, such as gastrulation in vertebrates, neural closure, Drosophila gastrulation, dorsal closure, as well as formation of the ventral furrow, eye disc morphogenetic furrow, and segmental groove in embryos (see review [61]).
N signaling is also involved in the boundary between new bud and the parent body of Hydra, where it is required for sharpening of the gene expression boundary and tissue constriction at the base of the bud [78]. Whether the role of N in these tissue constrictions is due to actomyosin-dependent apical constriction and epithelial fold is not known.
Boundaries may be established early in development. As the tissue grows in size through cell divisions and growth, boundary maintenance become essential. We found that N activity is maintained by actomyosin, suggesting feedback regulation to stably maintain the boundary. Mechanical tension generated by actomyosin networks has been suggested to enhance actomyosin assembly in a feedback manner (see review [79]). Interestingly, the N-mediated wing A/P and D/V boundaries, which form actomyosin cables rather than tissue folds, did not exhibit such positive feedback regulation (Fig 8D–8F). Instead, the stability of the Drosophila wing D/V boundary is maintained by a complex gene regulatory network involving N, Wg, N ligands and Cut [80, 81]. Perhaps this is necessary for a boundary not involving tissue morphogenesis.
The segmented appendages of arthropods (antennae, legs, mouth parts) are homologous structures of common evolutionary origin ([82, 83]). Snodgrass (1935) proposed that the generalized arthropod appendage is composed of a proximal segment called the coxopodite and a distal segment called the telopodite, either of which can further develop into more segments. The coxopodite is believed to be an extension of the body wall, whereas the telopodite represents the true limb, and thus represents an evolutionary addition [84, 85]. Dll mutants lack all distal segments except for the coxa in legs and the A1 segment in antennae [84, 86, 87]. Lineage tracing studies have shown that Dll-expressing cells contributed to all parts of the legs except the coxa [87, 88]. These results indicate that the leg coxa and antenna A1 segment correspond to the Dll-independent coxopodite, and that Dll is the selector gene for the telopodite. Therefore, the antennal A1 fold is the boundary between the coxopodite and telopodite. We postulate that the same N-mediated epithelial fold mechanism also operates in the coxopodite/telopodite boundary of legs and other appendages.
Flies were cultured in 25°C according to standard procedure unless otherwise noted. w1118 larvae were used for expression pattern analysis. Fly stocks were: sqhAX3; sqh-SqhGFP42 (Sqh–GFP) [89], Moe-ABD::GFP (also known as sGMCA [53]) was from Dan Kiehart (Duke University, North Carolina), hth-GAL4 [90] was from Richard Mann (Columbia University, New York), tub-GAL4 [91] was from Tzumin Lee (Janelia Farm Research Campus, HHMI, Virginia), dpp-GAL4c40.6. was from Jessica Treisman (New York University), fng-lacZ [92], Su(H)Gbe-lacZ was from Sarah Bray (University of Cambridge, UK), E(spl)mβ-lacZ [93], UAS-Nact [55], UAS-NDN [94], UAS-bantam and UAS-bantamsponge [19] were from Marco Milán (Institute for Research in Biomedicine, Barcelona). UAS-RNAi stocks were from VDRC (zip: 7819, sqh: 7916, mys: 29613, rhea: 40399, jub: 38442, ed: 104279/3087), NIG (N: 3936-R2), and Bloomington (N: 7870, zip: 36727, sqh: 32439, mys: 27735, rhea: 28950).
Genotypes for the mutant and MARCM clonal analysis were: hs-FLP1; UAS-rCD2-RFP, UAS-miR-GFP, FRT40A/ UAS-mCD8-GFP, UAS-miR-CD2, FRT40A; tub-GAL4/+ [41], hs-FLP; FRT42B, zip2/FRT42B, ubi-GFP, hs-FLP; FRT42B, zip2/FRT42B, tub-GAL80; tub-GAL4/ UAS-GFP [48], hs-FLP; FRT42D, sqaf01512/FRT42D, ubi-GFP (DGRC114526, sqaf01512 is a PiggyBac insertion in sqa [95]), hs-FLP; tub-GAL4, UAS-mCD8GFP/+; FRT82B, DlRevF10, SerRX82/FRT82B, tub-GAL80 (Bloomington 6300).
Positive labeled clones were induced using hs-FLP122; +; Act5C>CD2>GAL4, UAS-RFP [96]. Induction of hs-FLP122 was conducted at 38°C for 8 min at 24 or 48h after egg-laying (AEL). For lineage tracing experiments using Twin-Spot MARCM [41], newly-hatched first instar larvae were collected every two hours from juice plates, and kept in 25°C before heat shock (38°C for 10min). Larvae were raised under conditions of 25°C except for heat-shock at the indicated stage. Clonal induction was performed at L1 (AEH 18-20h), mL2 (AEH 26-28h), l-L2 (AEH 38-40h), or e-L3 (AEH 48-50h) stage. The discs were dissected and examined at l-L3.
We use AEH (after egg-hatching) for Twin-spot MARCM, and Dll/Lim1 expression pattern analysis, for which more precise timings are required. AEL (after egg-laying) was used for genomic mutant (zip2, sqaf01512, and DlRevF10, SerRX82, induced at L1), Ay (induced at L1/ L2), and tub-GAL80ts experiments.
Antibody staining was performed according to a procedure described previously [36]. Primary antibodies from DSHB (Developmental Studies Hybridoma Bank, University of Iowa) were mouse-anti-Coracle (C615.16, 1:20), mouse-anti Cut (2B10, 1:100), mouse-anti-Dl (C594.9B, 1:300), mouse-anti-Dlg (4F3, 1:200), mouse-anti-Ena (5G2, 1: 100), mouse-anti-FasIII (7G10, 1:50), mouse-anti-GFP (12A6, 1:100), mouse-anti-Nintra (C17.9C6, 1:200), mouse-anti-Ptc (Apa1, 1:100). Other primary antibodies included rabbit anti-Lim1 (1:400, from Dr. Juan Botas), rat-anti-Serrate (1:1000, preabsorbed, from Dr. Kennith Irvine), rabbit anti-aPKC (C-20, 1:50, Santa Cruz), rabbit-anti-caspase3 (cleaved) (1:200, Cell Signaling), goat-anti-Dll (F-20) (1:100, Santa Cruz), rabbit-anti-GFP (1:1000, Invitrogen), rabbit-anti-β-gal (1:5000, Cappel), rabbit-anti-phospho-Histone 3 (1:200, Millipore), rat-anti-RFP (5F8) (1:1000, Chromotek), Phalloidin (F-actin, Alexa 488-/555- or 647-conjugated) (1:100, Life Technologies). Species-matched Alexa 488-/561- or 633-conjugated secondary antibodies were from Jackson ImmunoResearch. Alexa Fluor 405-donkey anti-rabbit was from Abcam (ab175651). Images were acquired using a Zeiss LSM 780 or 710 with appropriate GaAsP detectors. Objectives were Plan-Apochromat 20x/0.8, Plan-Apochromat 40x/1.4 Oil, C-Apochromat 40x/1.2W Korr, and Plan-Apochromat 63x/1.4 Oil (Zeiss). All the images in this study were oriented dorsal-face up and with the posterior end to the right. Optical sections were oriented with the apical face of the disc proper to the right or top.
Images were processed with ZEN (Zeiss) with minimal brightness/contrast adjustments. To analyze the pixel intensities of Dll, Lim1 and N related (Dl, Ser and fng-lacZ) expression patterns, optical sections of 60μm were manually positioned with the center (0 in the X axis) placed at the fold (eL3) or at the Dll-Lim1 overlapping regions (L2). Although the larvae were collected at 1 hour intervals, there were still variations in developmental timing. Therefore, more than ten EADs were imaged, and only those of similar size were chosen for further analysis. Disc sizes in groups 1, 2, and 3 were, respectively: 4835 ± 328, 6058 ± 231, and 7065 ± 309μm2 (mean ± stdev). More than five EADs were quantified and 2–3 optical sections were analyzed per EAD. The signal intensity was established from the histogram analysis module in ZEN (Zeiss) and normalized to the basal level in non-expressing cells. The center (0 in the X axis) was manually positioned at the center of the Dll-Lim1 overlapping region. Correlations of ratios between Dll/Lim1 and Delta/Serrate were achieved by individual mean intensities from single cells. The stack images of 16–18μm were projected to ensure coverage of ligands and to identity genes. Cells with Dll-only, Dll+Lim1, and Lim1-only expressions were collected from the three groups. To establish Su(H)Gbe-lacZ levels in sqh and zip knockdown experiments, the pixel intensity of lacZ from optical sections across the A1 fold was quantified using the average pixel intensity of dpp-expressing regions normalized with non dpp-expressing regions in the same discs.
Time-lapse imaging to track cell morphology (Sqh-GFP and Sqh-mCherry) from l-L2 EAD ex vivo cultures was processed in Imaris software (Bitplane). 3D-projected images from time-lapse stacks were acquired using the surpass mode. A total of 5 hours of stack images were rotated and cropped in 3D to remove the peripodial membrane and basolateral regions. Segmentation of individual cells was carried out using the filament module with minimal manual corrections. The surface module was further applied to the post-filament images to obtain cell sizes and automatic tracking over time. Each cell was pre-processed for its absolute apical area value over time to determine whether it belonged to the constant (δArea < 10μm2), fluctuating (δArea ≥ 10μm2), or decreasing (initial apical area around 20–40μm2, and final < 10μm2) groups. For individual cells (as indicated by “i”), apical areas at each time point (Ati) were subtracted from the respective mean area over time (Aavgi) before normalization with the respective mean ((Ati-Aavgi)/Aavgi) to represent the proportional change. Proportional changes of cells in the same group were plotted as total mean and stdev. Trajectories of RFP clones in Sqh-GFP were accessed by spot tracking module in Imaris software (S2 and S3 Movie). The spot detection diameter was set to 1μm (shown as center point), with maximum distance between time points for 2μm. Autoregression motion algorithm were used to track RFP signal over time. 3D surpass time-lapse images were shown in spot center point with trajectory in dragon tail mode (for 20 time points). The overall trajectories of individual cells were presented in color-coded time map.
Apical (aPKC) and basolateral (FasIII) cell volumes were acquired from serial sections of fixed EAD using the Imaris surface module. Individual cell contours along the XY plane were outlined using the autofit module through all stack images, with settings of full accuracy and least impact. Stack contours from single cells were further processed to generate a 3D surface render and to acquire apical and basolateral volumes. For the basolateral domain, pinhole = 0.9 μm, optical interval = 0.47μm (total z = 30–40μm). For the apical domain, pinhole = 0.5 μm; optical interval = 0.27 μm (total z = 4–7μm).
Data sets were analyzed and plotted in Prism 6 using two-tailed un-paired t tests (S2C Fig, S6C Fig), linear regression analyse (Fig 6G), ANOVA-Tukey’s multiple comparisons (Fig 7G), and ANOVA-Dunnett’s multiple comparisons (Fig 8G; S5M Fig; S8E and S8F Fig).
Adult flies were fixed in Bouin’s solution, followed by serial dehydrations in 25%, 50%, 75%, and then 100% ethanol solutions before being transferred to 100% acetone. The samples were further processed by critical-point drying with liquid CO2, followed by sputter-coating with gold. Images were acquired using an Environmental Scanning Electron Microscope (FEI Quanta 200).
Ex vivo culturing and live imaging of EAD were as described [46]. For l-L2 and e-L3 EAD, the discs were embedded in 0.6% and 0.75% low gelling agarose, respectively.
Sqh-GFP and Moe-ABD::GFP were used as target molecule for CALI. CALI was carried out using an LSM710 inverted confocal microscope (Zeiss) with a 488nm laser (25mW) set at 100% of its power for a total of five cycles with 300 iterations per cycle (20–25 minutes break between each cycle, total of CALI treatment for 2.5 hours). The numerical zoom was set to 5 using a 40x objective. The region for CALI treatment was 3μm x 20μm with differential Z adjusted manually each time. Time-lapse images were acquired pre- and post-CALI treatment, with Z-stack set to a mean of 35μm. The time interval between each stack was 6 min as indicated in the S2 Movie. The parameters were: scan speed: 6 arbitrary units; number of scans per frame: 1; scanning: bi-directional; pinhole: 1.2μm; objectives: C-Apochromat 40x/1.2W Korr (Zeiss).
The EAD was dissected in DEPC-PBS, followed by fixation (4% PFA and 1% DMSO in PBS) for 20 min. Samples were washed in PBT (0.1% Tween20 in PBS) before proteinase K permeabilization (2 μg/mL in digestion buffer for 3 min, digestion buffer: 50 mM Tris-HCl, pH7.5 and 50 mM EDTA). After proteinase K inactivation (0.2% of glycine in PBS), samples were post-fixed with 4% PFA for 20 min. Samples were prehybridized in hybridization buffer (HYB: 50% formamide, 5x SSC, 0.1% Tween20, 100 μg/mL denatured salmon DNA, 100 μg/mL yeast tRNA, and 50 μg/mL heparin) for more than 1 hour at 60°C. The DIG-labeled probe (stock: 50 ng/μL, dilute stock 1:250 in HYB) was hybridized overnight at 60°C. After hybridization, samples were washed in 100% HYB, 66% HYB-PBT, 33% HYB-PBT, then PBT at 60°C for 1 hour each, then at room temperature for 4 more washes in PBT (5 min each). Samples were treated with 3% H2O2 in PBS to reduce endogenous HRP activity. Samples were then blocked in blocking solution (2% blocking reagent, 20% normal horse serum in PBT) for 30 min before overnight incubation with anti-Dig-HRP (POD Roche 1207–733, 1:100 dilute in blocking solution) at 4 oC. TSA amplification (PerkinElmer, NEL745001KT) was used to enhance the hybridization signals before protein detection. Protein immunofluorescence was performed after RNA in situ hybridization as described previously [36], except all steps were conducted in the dark.
The 5’- or 3’-DIG-labeled probes for bantam detection and the control sequence were: aatcagctttcaaaatgatctcacttgtatg (bantam), and gtgtaacacgtctatacgccca (scramble-miR, EXIQON).
|
10.1371/journal.pgen.1008177 | A mutation in the endonuclease domain of mouse MLH3 reveals novel roles for MutLγ during crossover formation in meiotic prophase I | During meiotic prophase I, double-strand breaks (DSBs) initiate homologous recombination leading to non-crossovers (NCOs) and crossovers (COs). In mouse, 10% of DSBs are designated to become COs, primarily through a pathway dependent on the MLH1-MLH3 heterodimer (MutLγ). Mlh3 contains an endonuclease domain that is critical for resolving COs in yeast. We generated a mouse (Mlh3DN/DN) harboring a mutation within this conserved domain that is predicted to generate a protein that is catalytically inert. Mlh3DN/DN males, like fully null Mlh3-/- males, have no spermatozoa and are infertile, yet spermatocytes have grossly normal DSBs and synapsis events in early prophase I. Unlike Mlh3-/- males, mutation of the endonuclease domain within MLH3 permits normal loading and frequency of MutLγ in pachynema. However, key DSB repair factors (RAD51) and mediators of CO pathway choice (BLM helicase) persist into pachynema in Mlh3DN/DN males, indicating a temporal delay in repair events and revealing a mechanism by which alternative DSB repair pathways may be selected. While Mlh3DN/DN spermatocytes retain only 22% of wildtype chiasmata counts, this frequency is greater than observed in Mlh3-/- males (10%), suggesting that the allele may permit partial endonuclease activity, or that other pathways can generate COs from these MutLγ-defined repair intermediates in Mlh3DN/DN males. Double mutant mice homozygous for the Mlh3DN/DN and Mus81-/- mutations show losses in chiasmata close to those observed in Mlh3-/- males, indicating that the MUS81-EME1-regulated crossover pathway can only partially account for the increased residual chiasmata in Mlh3DN/DN spermatocytes. Our data demonstrate that mouse spermatocytes bearing the MLH1-MLH3DN/DN complex display the proper loading of factors essential for CO resolution (MutSγ, CDK2, HEI10, MutLγ). Despite these functions, mice bearing the Mlh3DN/DN allele show defects in the repair of meiotic recombination intermediates and a loss of most chiasmata.
| Meiosis is a specialized cell division whereby a diploid cell undergoes one round of DNA replication followed by two rounds of division, yielding up to four haploid gametes. This process depends on tethering of maternal and paternal homologous chromosomes, and by the formation of crossovers (COs) between homologs during prophase I. COs arise from programmed double-strand breaks (DSBs), and can form via one of at least two mechanisms (class I, class II). In mouse, class I represents the major CO pathway, with the MLH1-MLH3 (MutLγ) complex being critical. MLH3 contains a conserved metal binding motif, DQHA(X)2E(X)4E, required for its endonuclease function, and this activity is postulated to represent a “resolvase”activity for class I COs. We generated a point mutant (Mlh3DN) in the endonuclease domain without altering the overall structure of MutLγ. Mlh3DN/DN males have no spermatozoa and are infertile, yet spermatocytes have grossly normal DSBs and chromosome pairing. The MLH3DN mutation permits normal loading of MutLγ, but key DSB repair factors persist in Mlh3DN/DN males, indicating a temporal delay in repair and suggesting a mechanism by which alternative DSB repair pathways may be selected. Thus, the endonuclease domain of MLH3 is important for normal processing of DSB repair intermediates.
| Meiosis is a specialized cell division process in which a diploid parental cell undergoes one round of DNA replication followed by two rounds of division, resulting in up to four haploid gametes. Successful halving of the genome during meiosis I depends on the tethering of maternal and paternal homologous chromosomes during meiotic prophase I, and their subsequent release at the first meiotic division. This tethering is ensured by homologous recombination, leading to the formation of crossovers; by synapsis, the formation of a tripartite proteinaceous structure, the synaptonemal complex, or SC between homologous chromosomes; and by cohesion between replicated sister chromatids that ensures appropriate tension on the metaphase I spindle [1,2]. Thus, recombination and synapsis are hallmarks of prophase I, and are both essential for ensuring homolog interactions leading to the formation of at least one crossover event per chromosome pair. Moreover, the correct placement, frequency, and distribution of crossovers is critical for ensuring appropriate disjunction at metaphase I and for maintaining genomic stability [1,3].
Meiotic recombination begins with the introduction of a large number of programmed double-strand breaks (DSBs), which are repaired as non-crossovers (NCOs) or crossovers (COs). Evidence for distinct NCO versus CO pathways was obtained in S. cerevisiae, where it was shown that the former occur earlier in meiotic prophase I, and subsequent work suggested that they appeared primarily through synthesis-dependent strand annealing (SDSA)[4,5]. In M. musculus, only 10% of DSBs are repaired as COs, while the majority are mostly repaired as NCOs, presumably via SDSA or other pathways, [6–9].
COs can form via one of at least two distinct mechanisms (referred to as class I and class II), each of which is used in varying degrees in different eukaryotic organisms [6,10,11]. The class I CO pathway is also known as the ZMM pathway, named after the major genes discovered in yeast that regulate this mechanism [12–18]. Class II COs, on the other hand, do not involve the ZMM proteins, but instead appear to rely on the structure-specific endonuclease (SSN), MUS81/EME1 (Mus81/Mms4 in S. cerevisiae) [6,10,11]. Class I COs also differ from class II COs in that the former are regulated by interference, the process by which placement of one CO prevents the nearby localization of a second CO, thus resulting in CO events spaced further apart than expected by chance [19].
In the class I CO pathway, DSBs are processed and resected to form single-end invasion (SEI) intermediates. This is followed by displacement of single strand DNA (ssDNA) from the recipient homolog to produce a double Holliday junction (dHJ). The yeast ZMM proteins Msh4 and Msh5 form a complex known as MutSγ that associates with a subset of these intermediate structures [20–22]. At least in yeast, this recruitment may be dependent on the STR complex, consisting of Sgs1 (BLM in mammals), Top3 and Rmi1 [23,24]. STR is proposed to act by disassembling the early recombination intermediates that would otherwise be processed through SSN-directed recombination pathways, thereby promoting either early NCO formation via SDSA, or CO formation through the capture of these recombination intermediates by the ZMM proteins, including MutSγ [23]. MutSγ is then thought to stabilize the dHJs, leading to the recruitment of a second MMR complex, MutLγ, consisting of the MutL homologs, Mlh1 and Mlh3 [25,26]. The mouse MutSγ complex associates with chromosome cores in zygonema [27], recruiting the MutLγ complex in pachynema. However, MutLγ associates with only a subset of MutSγ sites (~24–26 and 150 foci/nucleus, respectively), designating these events as class I COs [28,29].
Though not formally considered to be ZMM proteins, MLH1 and MLH3 are critical for most, if not all, class I CO events in numerous organisms [26,29–36]. In fact, the M. musculus MLH1-MLH3 heterodimer localizes to sites that are destined to become class I COs and the absence of either subunit in male spermatocytes leads to a dramatic decrease, but not complete absence, of chiasmata (the physical manifestation of a CO) [28,29,37–40]. While MutLγ is known to be recruited to sites that are preloaded with MutSγ, recent studies have shown that S. cerevisiae MutLγ can bind to single and double-stranded DNA (ssDNA, dsDNA), as well as a variety of branched DNA structures [33,41–43]. How such binding properties relate to the in vivo functions of MutLγ remains unclear.
Class I CO formation in M. musculus is dependent on MLH3, and on its heterodimeric interaction with MLH1 [8,29,37]. Interestingly, MLH3 recruitment precedes that of MLH1 [28]. Further analysis of MutLγ has shown that MLH3 contains a conserved metal binding motif, DQHA(X)2E(X)4E, originally discovered in the human MutL homolog, PMS2, and found to be required for human MutLα (hMLH1/hPMS2) endonuclease function [44]. This putative endonuclease motif is highly conserved in eukaryotic homologs of human PMS2 and MLH3, but not in homologs of human MLH1 and PMS1. The expectation for MLH3 is that this endonuclease function might represent a “resolvase”activity for class I COs. Studies in S. cerevisiae have shown that a single point mutation in the endonuclease motif of yeast Mlh3 (mlh3-D523N) disrupts its endonucleolytic activity and results in meiotic crossover defects similar to full mlh3 (mlh3Δ) null mutants, yet does not affect the protein stability of Mlh3 or its interaction with Mlh1 [32]. Further analysis of the entire endonuclease domain in S. cerevisiae revealed that mutation of any conserved residue results in a null or near-null phenotype with respect to crossing over [34]. Biochemical analysis reveals that the Mlh1-mlh3D523N protein lacks the ability to nick closed circular double stranded DNA, indicating loss of endonuclease activity [33,43]. Collectively, these studies in S. cerevisiae suggest that MutLγ plays a direct role in resolving dHJs to generate COs through its endonuclease activity.
To investigate the function of the putative endonuclease domain of MLH3 in mammalian meiotic recombination, we generated a point mutant mouse (termed Mlh3DN) in which the endonuclease domain was disrupted at the orthologous residue to the D523N mutation in yeast, allowing the overall structure of MLH3 to remain intact, as determined by the ability to form a stable complex with MLH1. By mutating the catalytic domain of MLH3, we hypothesized that the mutant MutLγ complex would remain structurally intact and thus might reveal a functional interplay with other meiotic CO functions. We demonstrate that normal function of the MLH3 endonuclease domain is required for resolution of DSB repair intermediates towards CO formation and thus for late meiotic recombination events. Mlh3DN/DN spermatocytes exhibit grossly normal DSB formation and early processing events, and normal timing of synapsis through early prophase I. Mlh3DN/DN spermatocytes exhibit appropriate localization of MLH3 and MLH1 to the synaptonemal complex during pachynema, along with pro-crossover factors HEI10 and CDK2, phenotypes that are clearly different from that observed in Mlh3-/- males. However, Mlh3DN/DN diakinesis-staged spermatocytes show significantly fewer chiasmata compared to wild-type mice (WT), but significantly more when compared to Mlh3-/- males, suggesting either that the MLH3DN protein retains partial endonuclease activity, or that the presence of the MutLγ complex, albeit altered in its endonuclease capacity, can invoke MLH3-independent repair pathways to become active by interfering with normal resolution of recombination intermediates. In line with these suggestions, we find that the RecQ helicase, BLM, is upregulated throughout prophase I in Mlh3DN/DN spermatocytes, perhaps aiding the recruitment of other repair proteins. To explore the increase in residual chiasmata observed at diakinesis in Mlh3DN/DN males relative to that of Mlh3-/- males, we demonstrate that co-incident loss of the class II CO pathway in Mlh3DN/DNMus81-/- double mutant males results in altered distribution of MutLγ, with an increased proportion of synapsed autosomes bearing no MutLγ foci. Furthermore, the proportion of chiasmata remaining in these double mutants is between that of Mlh3DN/DN and Mlh3-/- males, suggesting that MUS81-EME1 may account for only a proportion of these additional chiasmata, the mutant MutLγ retains residual resolvase activity, and/or mutant MutLγ can recruit other proteins to perform this resolvase activity at a subset of recombination intermediate sites. Collectively, our data show that the endonuclease activity of MLH3 is important for normal processing of DSB repair intermediates through the Class I pathway.
To investigate the meiotic requirement for the presence of a functional endonuclease domain in mammalian MLH3, we generated a mouse line with a point mutation in a conserved endonuclease motif located in the M. musculus protein: DQHAAHERIRLE [44,45]. Specifically, we replaced the aspartic acid "D" in amino acid position 1185, with an asparagine "N" by changing GAC to AAC in the genomic sequence, termed MLH3DN throughout. Extrapolating from an analogous mutation in the S. cerevisiae gene, this D-to-N replacement is predicted to disrupt the endonuclease function of MLH3 while maintaining its ability to interact with MLH1 ([32] S1 Fig). Mice were maintained on a C57Bl/6J background throughout the study.
Male Mlh3+/DN mice were phenotypically similar to WT littermates and displayed full fertility. Mlh3DN/DN males are also grossly normal when compared to WT littermates, survive into adulthood, and live normal lifespans. Mlh3DN/DN males also exhibit normal mating behaviors as determined by observing a vaginal plug in WT females the morning after mating. However, breeding between multiple sets of Mlh3DN/DN males and WT females never resulted in offspring over a four-year period.
Similar to the situation seen for Mlh3-/- males [29], Mlh3DN/DN males show complete infertility, accompanied by significantly reduced testes size when compared to WT (Fig 1A and 1B; p < 0.0001) and the absence of spermatozoa in the epididymides (Fig 1C; p < 0.0001). Whereas histological cross-sections of testes stained with hemotoxylin and eosin from WT males showed the presence of meiotic and post-meiotic cells within the seminiferous epithelium, testis sections from Mlh3DN/DN males were devoid of spermatids, but showed the presence of spermatogonia and spermatocytes (Fig 1D–1G). In addition, metaphase I spermatocytes were observed in the tubular lumen of Mlh3DN/DN mice (Fig 1G, black arrows). Thus, mutation of the endonuclease domain of Mlh3 in the mouse results in a sterility phenotype grossly similar to that seen in Mlh3-/- mice.
To investigate the progression of meiotic recombination, prophase I chromosome spreads were prepared from WT, Mlh3DN/DN, and Mlh3-/- adult males and stained for a variety of markers involved in synapsis and recombination. Chromosome spreads were stained with antibodies against γH2AX, the phosphorylated form of histone H2AX, as a marker of DSBs [46,47]. In spermatocyte preparations from WT males, γH2AX signal is abundant throughout the nucleus at leptonema, coincident with the induction of several hundred DSBs [1,47]. The γH2AX signal declines in zygonema as DSBs are processed for repair [47,48]. In pachynema and diplonema, γH2AX signal is absent from the autosomes, but emerges throughout the sex body due to meiotic sex chromosome inactivation (MSCI) ([49]; S2A–S2D Fig). Spermatocytes from both Mlh3DN/DN and Mlh3-/- males exhibit the same γH2AX signal and temporal dynamics as observed in WT spermatocytes, with abundant staining in leptonema, slightly reduced signaling in zygonema, followed by the absence of γH2AX signal on the autosomes of pachytene and diplotene spermatocytes, except at the sex body (S2F–S2I and S2K–S2N Fig). We do not see specific persistent γH2AX signal on the autosomes at pachynema in Mlh3-/- spermatocytes [50], unless we markedly increase our imaging exposure time γH2AX (S2E, S2J and S2O Fig; white arrows). Under these conditions, we see persistent foci of γH2AX in spermatocytes from WT and from Mlh3DN/DN spermatocytes also. Thus, in our hands, we see no specific persistence in autosomal γH2AX signal through pachynema in mice lacking MLH3 or harboring a mutation within the endonuclease domain of MLH3.
Spermatocyte chromosome spreads from WT and Mlh3DN/DN males were stained with antibodies against synaptonemal complex (SC) components, SYCP3 and SYCP1, marking the axial/lateral elements and the transverse filaments, respectively. Prophase I progression in WT spreads is characterized by the initial accumulation of SYCP3 signal in discrete dots along chromosomes at leptonema, and these dots gradually coalesce into continuous filaments along the chromosome cores in zygonema (S2Q Fig). At this time, SYCP1 appears in patches along the SYCP3 signal, indicating that synapsis is occurring. By late zygonema, most of the chromosome core is now labeled with SYCP1, and by pachynema synapsis is complete, as demonstrated by complete overlap of the SYCP3/SYCP1 signals on the autosomes. For the sex chromosomes, synapsis only occurs at the pseudoautosomal region (PAR). After meiotic recombination occurs, the SC begins to degrade in diplonema, and the homologs are no longer tethered to one another except at CO sites (S2P–S2S Fig).
Synapsis appears normal in Mlh3DN/DN spermatocytes with discrete accumulation of SYCP3 on the chromosomes in leptonema, followed by continued accumulation of SYCP3 along the chromosomes as SYCP1 appears in patches in zygonema (S2T and S2U Fig). Complete synapsis of the autosomes and the PAR is observed in pachynema with co-localization of SYCP1 and SYCP3 (S2V Fig). Desynapsis is then observed in diplonema with the degradation of the SC (S2W Fig). Thus, synapsis in Mlh3DN/DN spermatocytes appears unaffected by loss of the endonuclease activity of MLH3, a result similar to that seen for complete loss of MLH3 protein.
Early DSB repair events were monitored by examining localization of the RecA strand exchange protein, RAD51, on chromosome cores of the autosomes throughout prophase I [51,52]. In WT mice, RAD51 localizes to chromosome cores of early and late zygotene cells as discrete foci at a high frequency (EZ and LZ, respectively; Fig 2A and 2G). Compared to WT littermates in early and late zygonema, RAD51 counts in spermatocytes from Mlh3DN/DN males were significantly elevated (Fig 2C and 2G; p<0.001 and p<0.01, respectively, by unpaired t-test with Welch’s correction). However, while early zygotene RAD51 counts were indistinguishable in Mlh3-/- spermatocytes compared to WT (Fig 2E and 2G), they were significantly lower than that seen at the equivalent stage in Mlh3DN/DN males (p<0.001 by unpaired t-test with Welch’s correction). By late zygonema, the RAD51 counts were significantly lower in Mlh3-/- spermatocytes compared to WT and Mlh3DN/DN animals (p<0.001 by unpaired t-test with Welch’s correction). By pachynema, RAD51 foci frequency in spermatocytes from WT mice decreased to very low numbers, as did that of Mlh3-/- males (Fig 2B, 2F and 2H; p = 0.55 unpaired t-test). In contrast, focus counts in pachytene spermatocytes from Mlh3DN/DN males remained significantly elevated following the pattern first seen in zygonema (Fig 2D and 2H; p<0.0001).
The localization and accumulation of single strand DNA binding protein RPA, which associates with chromosomes from zygonema through until early pachynema, was also explored. (S3 Fig). In leptonema and zygonema, RPA focus counts on chromosome cores were significantly elevated in Mlh3DN/DN animals compared to WT (p<0.01 and p<0.001, respectively, unpaired t-test with Welch’s correction), similar to the increased focus frequency observed for RAD51. However, unlike RAD51, the RPA focus counts were not significantly different between genotypes at pachynema and diplonema (S3 Fig). These observations suggest either that there is a prolonged period of DSB induction in Mlh3DN/DN animals, or that there is a lag time in the turnover of DSB repair intermediates in early prophase I. Taken together with the persistent RAD51 localization, these observations suggest that, in Mlh3DN/DN spermatocytes, there is a persistence of DSB repair intermediates loaded with RPA in zygonema, and that these intermediates continue to persist as they accumulate RecA homolog proteins, with RAD51 remaining on chromosome cores of Mlh3DN/DN spermatocytes through pachynema. We hypothesize, based on these observations, that RPA accumulation in leptonema and RAD51 accumulation in zygonema are affected by loss or mutation of MLH3 protein, suggesting an early function for MutLγ in establishing appropriate DSB repair intermediates that is not confined to CO pathway fate [42,53]. We hypothesize that early DSB repair events occur within the normal timeframe in mice lacking MLH3 protein entirely, but at an even faster rate than in WT spermatocytes, because in late zygotene RAD51 counts in Mlh3-/- mutants have declined to levels seen in pachytene. The repair of these DSBs in Mlh3-/- mutants may occur through repair pathways that differ from those utilized in WT-derived spermatocytes. We hypothesize that in Mlh3DN/DN mutants DSBs are not repaired efficiently, or there is an extended period of DSB induction resulting from feedback mechanisms that lead to a persistence of RPA and RAD51 foci.
Bloom's syndrome mutated (BLM) is a mammalian RecQ DNA helicase whose S. cerevisiae ortholog, Sgs1, was shown to promote the resolution of complex multi-chromatid joint molecule intermediates, that may result from SEI events, into both NCOs and COs [23,24]. During prophase I in WT male spermatocytes, BLM localizes to the chromosomal cores at a high frequency in zygonema and diminishes to a few foci in pachynema [54–56]. Recently, we showed that loss of MLH3 results in up-regulated BLM localization during prophase I, along with persistence of BLM on chromosome cores through late pachynema [56].
To determine if the disruption of the MLH3 endonuclease domain affects the localization of BLM in a similar fashion, to Mlh3-/-, we stained prophase I chromosome spreads with an antibody against BLM. In zygonema, as previously reported, WT cells show the accumulation of BLM foci on the cores in high numbers, and this frequency is elevated in spermatocytes from both Mlh3DN/DN and Mlh3-/- spermatocytes (Fig 3A, 3B, 3E, 3F, 3I, 3J and 3M; p<0.0001 unpaired t-test). This is similar to that reported previously for Mlh3-/- spermatocytes [56]. In early to mid-pachynema, BLM localization on chromosome cores persists in a small percentage of WT spermatocytes, but the number of foci is very much reduced at this stage (Fig 3C, 3D and 3M). In contrast, all spermatocytes from Mlh3DN/DN and Mlh3-/- spermatocytes show persistent BLM focus localization along chromosome cores (Fig 3G, 3H, 3K and 3L) at a frequency that is elevated above that of WT spermatocytes (Fig 3M, p<0.0001 unpaired t-test). Moreover, the number of BLM foci at pachynema in Mlh3DN/DN spermatocytes is significantly elevated relative to that seen in Mlh3-/- spermatocytes (Fig 3M, p<0.05 unpaired t-test). By diplonema this difference appears even greater, with BLM localization in Mlh3DN/DN spermatocytes persisting in stretches along the cores, and being lost entirely in Mlh3-/- spermatocytes (Fig 3D, 3H and 3L). Thus, altered MLH3 endonuclease function, like complete loss of MLH3, leads to persistence of BLM helicase on chromosome cores in late prophase I, but at an elevated frequency in Mlh3DN/DN spermatocytes relative to Mlh3-/- spermatocytes.
“Crossover designation” is defined as the process by which class I COs are selected from an excess pool of DSB repair intermediates. In mouse, the 250+ DSBs are processed through zygonema into various repair pathways, and only a subset of these will proceed towards a class I CO fate [1]. These sites become “licensed” for crossing over through the accumulation of the MutS homolog heterodimer, MutSγ (MSH4 and MSH5; [27,57]). The MutSγ complex then serves as an early pro-crossover factor by recruiting the MutLγ complex to a select subset of sites and it is these sites that will become “designated” as class I CO events. Notably, all of the 150+ MutSγ sites must be repaired, either as a CO or an NCO, which means that approximately ~125 MutSγ sites must leave the class I CO pathway and undergo repair through an alternate CO pathway or via an NCO pathway, a situation that is unlike that seen in S. cerevisiae where the number of MutSγ sites appear to correspond more closely to the number of CO events [20].
While the mechanism by which only a subset of MutSγ foci are retained through pachynema remains unclear, studies from a number of groups have implicated the Zip3-like protein, RNF212, in this process [58,59]. RNF212 has been shown to co-localize with the majority of MutSγ foci in spermatocytes from WT males and is thought to act as a pro-crossover factor by stabilizing these MutSγ-loaded events [59]. As such, the number of RNF212 foci on chromosome cores is pared down through pachynema in a similar fashion to that of MutSγ [59,60]. Moreover, in mouse mutants that disrupt this paring down process, both RNF212 and MutSγ focus counts remain elevated, but equivalent, throughout prophase I [50,61].
To investigate how loss of MLH3 endonuclease function could affect this paring down process, we analyzed RNF212 and MSH4 focus dynamics on chromosome spreads throughout prophase I from WT, Mlh3DN/DN, and Mlh3-/- adult male mice (S4 Fig). For both RNF212 (S4A–S4M Fig) and MSH4 (S4N–S4Z Fig), we find the expected paring down of focus counts from early pachynema (EP) to late pachynema (LP) in spermatocytes from WT, Mlh3DN/DN, and Mlh3-/- adult males. In all three cases, RNF212 and MSH4 foci appear on chromosome cores in zygonema (S4B, S4F, S4J, S4O, S4S and S4W Fig), persist at high levels in early pachynema (S4C, S4G, S4K, S4P, S4T and S4X Fig), and then are reduced to approximately 1–2 foci per chromosome in late pachynema (S4D, S4H, S4L, S4Q, S4U and S4Y Fig). Quantification of RNF212 and MSH4 focus numbers in early and late pachytene spermatocytes from WT, Mlh3DN/DN, and Mlh3-/- adult male reveals the expected statistically significant decline in these foci through pachynema (S4M and S4Z Fig; p<0.0001 Mann-Whitney U Test for all). However, the levels of RNF212 foci in both early and late pachynema are significantly higher in spermatocytes from Mlh3DN/DN and Mlh3-/- adult males compared to that seen in WT spermatocytes (S4M Fig; p<0.001 Mann-Whitney U test for all). Thus, while the dynamics of RNF212 (high in early and low in late pachynema) are evident in Mlh3DN/DN and Mlh3-/- adult males, their focus counts at each of these stages are significantly elevated compared to equivalently-staged WT RNF212 counts. By contrast, at both early and late pachynema, MSH4 counts did not differ between spermatocytes from WT, Mlh3DN/DN and Mlh3-/- adult males at both early and late pachynema (S4Z Fig). Thus, mice bearing no MLH3 or catalytically defective MLH3 show a phenotypic divergence in RNF212 and MSH4 focus counts in pachytene spermatocytes compared to WT.
MutLγ represents the ultimate marker of DSB repair events that have adopted a class I CO fate, and has been used as a CO proxy marker in many organisms [29,62–64]. We anticipated that the D1185N endonuclease mutation in MLH3 would not affect localization of this complex. In WT spermatocytes, MLH3 localizes on the chromosomes during early pachynema, remaining associated with SYCP3 signal through to diplonema (Fig 4A). In pachytene spermatocyte preparations from Mlh3DN/DN mice, MLH3 signal remains associated with the autosomal chromosome cores from early pachynema at a focus frequency that is statistically indistinguishable from that of WT cells (Fig 4A–4C, p = 0.36 by unpaired t-test). MLH3 association with the PAR of the synapsed X and Y chromosomes was similarly unaffected in Mlh3DN/DN pachytene spermatocytes. In addition, the timing of MLH3 appearance, in early pachynema and prior to that of MLH1, was normal in Mlh3DN/DN pachytene spermatocytes.
Localization of MLH1 was similarly explored in spermatocytes from Mlh3DN/DN mice. As with MLH3, there was no difference in the timing of MLH1 accumulation on chromosome cores between WT and Mlh3DN/DN mice (Fig 4D–4F). Moreover, when autosomal MLH1 foci were quantified, no statistical difference was observed in MLH1 focus frequency between WT and Mlh3DN/DN pachytene cells (Fig 4D–4F, p = 0.2 by unpaired t-test). These data suggest that disruption of the endonuclease domain of MLH3 does not alter recruitment of MutLγ to chromosomes in pachynema.
To observe class I CO events in pachynema, we employed two well characterized markers of these sites: the putative ubiquitin E3 ligase, Human Enhancer of Invasion-10 (HEI10), and cyclin-dependent kinase-2 (CDK2) [50,65,66]. In WT prophase I cells, CDK2 localizes to the telomeres (Fig 5A, yellow arrows) as well as on the chromosome cores (Fig 5A, white arrows) during mid to late pachynema and remains associated with SYCP3 signal through to diplonema [66]. The localization of CDK2 along chromosome cores parallels the localization of MLH1 and MLH3, both temporally and quantitatively (Fig 5A and 5D), and is associated with nascent class I CO events. In pachytene spermatocyte preparations from Mlh3DN/DN mice, CDK2 signal remains associated with both the telomeres and chromosome cores at a frequency and intensity that is reminiscent of that seen in WT spermatocyte spreads (Fig 5B and 5D). This is in contrast to the situation in spermatocyte preparations from Mlh3-/- males, in which CDK2 association with the telomere persists, but is lost from nascent CO sites (Fig 5C and 5D).
HEI10 was recently shown to co-localize with MutLγ at sites of class I CO, and its localization is dependent on Cyclin N-terminal Domain-containing-1 (CNTD1)[50,61]. HEI10 is thought to play a key role in CO designation/maturation [50]. As previously reported for WT cells at pachynema, HEI10 localizes with similar frequency to that of CDK2 and MutLγ (Fig 5E [pink arrows], 5H). Similar localization patterns and frequency were observed for Mlh3DN/DN mice, with a frequency of one to two foci per chromosome (Fig 5F [pink arrows] and H), indicating normal recruitment of HEI10 on pachytene chromosome cores in Mlh3DN/DN males. This is in contrast to the pattern of HEI10 staining in spermatocytes from Mlh3-/- mice, where there is an increased accumulation of HEI10 foci (Fig 5G [pink arrows], 5H), as previously reported [50]. For both CDK2 and HEI10, the altered frequencies of foci observed in spermatocytes from Mlh3-/- mice were significantly different from that of WT or Mlh3DN/DN males (Welch’s T-test, p<0.0001). CDK2 and HEI10 focus counts for WT or Mlh3DN/DN males were not statistically different from each other. Taken together, these observations demonstrate that loading of HEI10 and CDK2 on class I CO designated sites is affected differently by mutation of Mlh3: complete loss of MLH3 results in failure to load CDK2 and hyper-accumulation of HEI10, while altered endonuclease activity of MLH3 results in normal loading of both CDK2 and HEI10. Thus, the physical accumulation of MutLγ is required for normal loading of associated pro-crossover maturation factors.
Mouse Mlh1 and Mlh3 were amplified from cDNA and cloned into pFastBac1 vectors as described in the Methods. The MLH1-MLH3 and MLH1-MLH3-D1185N complexes were expressed from Sf9 cells infected with baculoviruses containing MBP-Mlh1 and His10-Mlh3 or His10-Mlh3-D1185N constructs (Fig 4G). Extracts from these cells were applied to a Ni-NTA column. Fractions containing induced proteins were pooled and then applied to an amylose column. Two major bands of molecular weights predicted for an MBP-MLH1-His10-MLH3 complex were detected on SDS-PAGE after amylose chromatography (Fig 4H). These bands were further analyzed by mass spectrometry, and the results from this analysis confirmed their identity (Fig 4I). Importantly, MLH1-MLH3 and MLH1-MLH3-D1185N eluted with an apparent 1:1 stoichiometry in both chromatography steps, indicating that the heterodimers were stable, and the protein yields of the two complexes after amylose chromatography were similar (Fig 4H).
Chiasmata are the physical manifestations of crossing over and, as such, can inform the process of DSB repair via all pathways. Diakinesis-staged spermatocytes from WT and Mlh3DN/DN males were used to quantify chiasmata. WT cells exhibited a chiasmata frequency of 23.5 ±1.3 per nucleus (Fig 6A and 6D) whereas Mlh3DN/DN spermatocytes exhibited a dramatically reduced chiasmata count of 5.2 ± 1.7 chiasmata per nucleus (Fig 6B and 6D; p < 0.0001 by unpaired t-test). Chiasmata counts for Mlh3-/- males were even more dramatically reduced at 2.8 ± 1.1 chiasmata per nucleus, a value that is significantly lower than both WT and Mlh3DN/DN spermatocytes (Fig 6C and 6D; p <0.0001 by unpaired t-test). Thus, complete loss of MLH3 protein leads to the loss of approximately 88% of chiasmata, while loss of endonuclease activity, but retention of MutLγ heterodimer results in only a 78% loss. Thus, the number of residual chiasmata observed in Mlh3DN/DN spermatocytes is higher than the expected number of chiasmata achieved through the MUS81-EME1-driven class II CO pathway (~2–3, assessed both cytologically and genetically; [8,28]).
The increased residual chiasmata observed in Mlh3DN/DN males compared to Mlh3-/- animals prompted us to ask whether some or all of these crossovers were dependent on the activity of the MUS81-EME1 heterodimer. Previous studies in our lab showed that Mus81-/- animals show increased accumulation of MutLγ, resulting in normal chiasmata counts, suggesting that class I CO events are up-regulated in the absence of the class II machinery [6]. Co-incident mutation of one or both Mlh3 alleles to the Mlh3DN variant on the Mus81-/- mutant background yielded MLH1 focus counts that were significantly reduced compared to those observed in Mlh3+/+Mus81-/- mice (Fig 6E, p<0.005 by unpaired t-test with Welch’s correction), and instead resembled MLH1 focus counts observed in spermatocytes from Mlh3+/+ mice (Fig 4F). Thus, the upregulation of MLH1 foci at pachynema requires both a defective MUS81-EME1 dimer and the presence of only functional MutLγ heterodimer.
Interestingly, pachytene spermatocytes from Mlh3+/DN Mus81-/- and Mlh3DN/DN Mus81-/- males show an abnormal distribution of MLH1 foci across all autosomal pairs (Fig 6F). In Mlh3+/+Mus81-/- males, 17% (5/29) cells showed synapsed chromosomes without any MLH1 foci in pachynema (so-called “no exchange” or “E0” chromosomes), while in Mlh3+/DN Mus81-/- males, this proportion increased to 67% (8/12), with up to 3 E0 chromosomes per cell (Fig 6F; examples shown in S5 Fig). In Mlh3DN/DNMus81-/- males, 94% of cells had E0 chromosomes (29/31), with as many as 6 E0 bivalents being observed (Figs 6F and S5). The higher proportion of MLH1-devoid autosomes in Mlh3+/DN and Mlh3DN/DN males on the Mus81 null background was statistically significant in all pairwise comparisons (Fig 6F, unpaired t-test with Welch’s correction with Bonferroni adjustment), indicating that the placement of an obligate crossover is perturbed in mice having one or two copies of the Mlh3DN allele on a Mus81 null background, and suggesting that mechanisms that ensure correct CO placement require a function MutLγ complex. In yeast, and also probably in mice, CO interference requires a functional MutSγ complex [67]. However, MutSγ alone is not sufficient to ensure appropriate CO placement since Mus81-/- males exhibit disrupted interference despite appropriate MutSγ loading [6].
Assessment of chiasmata counts in single (Fig 6D) and double mutants (Fig 6G) revealed the expected normal chiasmata frequency in spermatocytes from Mus81-/-Mlh3+/+ and Mus81-/-Mlh3+/DN males, and the loss of most chiasmata in spermatocytes from Mus81-/-Mlh3DN/DN males. However, whereas the loss of chiasmata structures in Mlh3-/- and Mlh3DN/DN single mutants was observed to be 88% and 78%, respectively (Fig 6D), the loss of chiasmata in Mus81-/-Mlh3DN/DN males was 83%. The frequency of chiasmata in cells from these double mutants was statistically different from both Mlh3 single homozygous mutant animals (p<0.05, unpaired t-test with Welch’s correction). Thus, loss of the class II pathway in addition to the mutation of the endonuclease domain of Mlh3 only partially reduces residual chiasmata counts to the levels observed in Mlh3-/- animals. Interestingly, despite reduced chiasmata, the overall number of bivalent structures observed in diakinesis preparations from Mus81-/-Mlh3DN/DN males is the same as that seen in Mlh3DN/DN males, and is significantly elevated above that seen in Mlh3-/- single mutant males (Fig 6H). Taken together, these observations suggest two important features regarding class I/II interactions: (1) that a fully functional class I and class II machinery is required for appropriate distribution of MutLγ foci across the genome; and 2) that MUS81-EME1 activity cannot fully account for the residual chiasmata count observed in Mlh3DN/DN males.
Studies in S. cerevisiae and M. musculus have implicated MutLγ as the major resolvase of dHJs in the class I CO pathway [29,32–35,41,43,44,68]. The current study examines the importance of an intact endonuclease domain for the proper functioning of MLH3 during prophase I of mammalian meiosis, and is the first exploration of a point mutation for MutLγ in the mouse. We generated a mouse with a mutation in the MLH3 endonuclease domain that affects its catalytic activity while allowing for heterodimer assembly. We found that, as in Mlh3-/- mutant males, Mlh3DN/DN males are infertile, exhibit significantly smaller testes than their WT litter mates, and have no epididymal spermatozoa. Beyond this, our data reveal important similarities and differences in the meiotic phenotypes to that observed with a nullizygous Mlh3-/- allele, as articulated below.
Importantly, the phenotypic consequences of loss of a functional MutLγ complex occur despite normal accumulation of both MutSγ and MutLγ indicating that the physical presence of these complexes is not sufficient to ensure complete CO resolution. However, CO licensing (defined by MutSγ deposition) and designation (defined by MutSγ and MutLγ deposition) are both normal in Mlh3DN/DN males, indicating that a functional MutLγ complex is not required for these CO-defining processes. Similarly, although normal accumulation of pro-crossover factors, CDK2 and HEI10, is observed in Mlh3DN/DN males (unlike the situation in Mlh3-/- males), this is not sufficient to drive CO resolution along the class I CO pathway.
We demonstrate that an intact endonuclease domain within MLH3 is not required for DSB or synaptonemal complex formation in early prophase I, similar to that seen in Mlh3-/- males [29]. However, there are distinct differences in RAD51 accumulation and persistence in spermatocytes from Mlh3-/- and Mlh3DN/DN males, suggesting that the effect of MutLγ loss on DSB repair processing is quite different from the presence of a defective MutLγ complex. Most importantly, while RAD51 is recruited in elevated numbers to chromosome cores of Mlh3DN/DN males, it fails to be cleared effectively in pachynema, perhaps because the defective MutLγ complex blocks subsequent processing of DSB repair intermediates. Intriguingly, the significantly altered RAD51 accumulation in leptonema in both Mlh3 mutants indicates a role for MutLγ prior to pachynema, far earlier than has been defined thus far. Indeed, analogous early pairing roles for MutLγ have been proposed for M. sordaria and S. cerevisiae [42,53]. In yeast, Al-Sweel et al. constructed whole genome recombination maps for wildtype, endonuclease defective, and null mlh3 yeast mutants. Both the endonuclease defective and null yeast mutants for mlh3 showed increases in the number of NCO events, consistent with recombination intermediates being resolved through alternative recombination pathways [34]. Thus, in the case of yeast, loss of Mlh3 protein, or the production of an endonuclease defective protein, increases the frequency of other recombination outcomes, most notably including earlier NCO events.
The absence of either component of MutLγ results in the loss of 90–95% of chiasmata, consistent with the established dogma that class I COs account for the majority, but not all, chiasmata in mammalian meiosis [6,8,29,39,68]. By contrast, loss of MUS81, the major class II CO regulator, results in normal chiasmata levels as a result of up-regulation of class I events, as evidenced by a ~10% increase in MutLγ localization during pachynema [6], suggesting that loss of the class II pathway leads to a compensatory increase in class I events. Furthermore, our previous analysis of Mlh3-/-Mus81-/- double mutant animals revealed a very small (<1 on average), but consistent, number of residual chiasmata, indicating the existence of other resolvase complexes [6], as has been demonstrated for yeast and plants [10,11]. Taken together, these observations have two important implications for crossing over in the mouse: first, additional class I CO events can be achieved through recruitment of additional MutSγ-designated precursor sites in the absence of the class II pathway (and possibly under other circumstances too), and second, a few crossovers can be achieved without implementing either MutLγ or MUS81-EME1.
In the current study, we show that spermatocytes from Mlh3DN/DN males show normal accumulation of both MLH1 and MLH3 at pachynema, but this results in an increase in the residual chiasmata count at diakinesis relative to that seen in Mlh3-/- mice: approximately 78% of COs are lost in Mlh3DN/DN spermatocytes, leading to 22% residual chiasmata. This suggests several possibilities: either that the endonucleolytic function of MLH3 does not account for the resolution of all class I COs under wildtype situations, and/or that other resolvases can be recruited under certain circumstances once MutLγ loads, irrespective of whether this complex is endonucleolytically competent. Alternatively, the point mutation in the endonuclease domain does not completely eliminate endonucleolytic activity in the mouse, resulting in partial class I resolvase activity. We find this latter possibility unlikely due to the severity of the defect in endonuclease activity in the S. cerevisiae Mlh1-mlh3-D523N complex [32,33], but we were unable to test this in the current analysis.
Another explanation for the difference in chiasmata counts between Mlh3DN/DN and Mlh3-/- mice is that, in the former, the existence of a defective MutLγ prevents most class I-type COs, but facilitates the resolution of recombination intermediates through alternative pathways. Thus, despite the presence of the MLH3DN protein, some class I CO events can be processed by other CO machinery under conditions of normal accumulation of pro-crossover factors, MutSγ, HEI10 and CDK2. While we cannot assess the recruitment of the class II machinery to sites of DSB repair in prophase I in the mouse (due to the lack of available reagents), there is evidence to support the idea that MUS81-EME1 might participate in this CO resolution crosstalk. Specifically, double mutants lacking Mus81 and bearing a homozygous Mlh3DN allele show reduced chiasmata relative to Mlh3DN/DN males, indicating that MUS81 may account for at least some of the increase in chiasmata above that of Mlh3-/- males. Arguing against this idea is the observation that the loss of Mus81 on the Mlh3-/- background results in a similar drop in residual chiasmata from Mlh3-/- single mutant males alone to that observed in Mlh3DN/DNMus81-/- double mutant animals compared to Mlh3DN/DN single mutant animals.
Previous analysis of Mus81-/- single mutant males revealed a compensatory increase in MutLγ in pachynema that resulted in normal crossover numbers as a result of upregulation of class I CO events [6]. The persistence of RAD51 at several foci in late pachynema in these Mus81-/- males suggested that the class II-destined DSBs were not repaired, but that additional COs were derived from increased CO designation from the larger pool of MutSγ-loaded DSB repair intermediates, suggesting some crosstalk between CO pathways to achieve CO homeostasis. Further evidence for crosstalk between the two major crossover pathways is provided in the current study. However, this elevated MLH1 focus frequency is not observed in Mlh3DN/DNMus81-/- and Mlh3+/DNMus81-/- males, suggesting that a functional MutLγ is required for this crosstalk. Intriguingly, loss of Mus81 on the Mlh3DN/DN and Mlh3+/DN backgrounds results in altered distribution of MLH1 across the genome, resulting in elevated numbers of chromosomes lacking an MLH1 focus entirely. This increase in “E0” chromosomes does not occur in either Mus81-/- or Mlh3DN/DN single null males, indicating that CO distribution is dependent on the functionality (or partial functionality) of both pathways. While we cannot fully explain the reason for altered MLH1 distribution in both Mlh3DN/DNMus81-/- and Mlh3+/DNMus81-/- males, these observations point to complex interplay between crossover pathways in achieving normal distribution of crossover events in mammalian meiosis. Indeed, our recent studies involving a mouse model harboring a point mutation within Msh5 indicates that altered MutSγ function affects both crossover pathways [69].
Our previous studies, along with the current one, indicate a role for BLM helicase in modulating the pathway choice in DSB repair during mouse meiosis. In Mlh3-/- [56] and Mlh3DN/DN (current work) males, prophase I spermatocytes show increased and persistent accumulation of BLM helicase through until late pachynema. A similar increase in BLM localization was also noted in Mus81-/- spermatocytes [6]. In S. cerevisiae, loss of class I CO pathway components (for example, in msh4/5 or mlh1/3 mutants) is suppressed by mutation of the BLM ortholog, Sgs1, highlighting the role of Sgs1 as an anti-crossover factor. However, the additional CO events that arise in these double mutant yeast strains are presumed to be processed via Mus81-dependent resolution, and thus via class II CO events [13,36,70]. In msh4/5 sgs1 double mutant strains, the restoration of COs occurs without any concomitant decrease in NCO events, suggesting either that other CO pathways account for the non-class I COs, or that these DSBs are repaired via inter-sister repair processes. In this sense, Sgs1 has been proposed to be master orchestrator of recombination pathway choice [36], while the Sgs1-Top3-Rmi1 complex as a whole can regulate CO formation both positively and negatively in yeast [23,24]. The situation we observe in Mlh3-/- and Mlh3DN/DN males with respect to BLM persistence may be similar to that seen in yeast for Sgs1, in that up-regulation of BLM foci is observed in both Mlh3 mutant lines from zygonema onwards, but is significantly higher in Mlh3DN/DN males at pachynema. Thus, residual chiasmata counts in Mlh3-/- and Mlh3DN/DN males are proportional to pachytene BLM focus counts (lower in full nulls, higher in Mlh3DN/DN males). Thus, we can postulate that the loss of MLH3 protein entirely in Mlh3-/- males results in a compensatory, but ineffective, increase in BLM that cannot overcome the failure to process class I COs sufficiently (a situation that is different to yeast). In the presence of intact, but catalytically inert MutLγ, on the other hand, the availability of additional BLM foci can then direct DSB repair in favor of other CO pathways in a similar fashion to the situation in yeast, where the engagement of Sgs1 promotes alternative repair mechanisms, primarily through the recruitment of structure specific nucleases, and the resolution of some dHJs through a class II (or other) CO pathway [13,36,70–72].
We provide evidence that COs are achieved in Mlh3DN/DN spermatocytes in a manner that may be dependent on the MUS81-EME1 endonuclease, or on other resolvase complexes that have yet to be determined in mammalian meiosis. Indeed, our previous analysis of Mlh3-/-Mus81-/- males indicated the existence of additional CO events that were independent of the class I and class II pathways [6]. Additional resolvases in yeast include SLX1-SLX4 and YEN1/GEN1[35,73,74]. The persistence of DSB repair intermediates into pachynema, along with the upregulated and persistent BLM might suggest that the defective MutLγ complex prevents accumulation of other such resolvase complexes. This might, in turn delay CO maturation until later in prophase I when, for example, GEN1 can be invoked to resolve the CO [75]. Thus, we propose that the timing of MutLγ activity, and its clearance from nascent COs is an important factor in the recruitment of alternative CO processing machineries, but in a manner that is not dependent on its endonuclease activity.
Work performed in this manuscript was approved by the Cornell Institutional Animal Care and Use Committee, under protocol 2004–0063.
A PL253 targeting vector containing the Mlh3-D1185N point mutation in the potential endonuclease domain and a loxP-neo-loxP cassette in intron 5–6 of Mlh3 was incorporated into an embryonic stem cell line. Mlh3DN transgenic mice were crossed with a Spo11-Cre mouse line to remove the neo cassette [76], and then maintained on an inbred background through backcrossing on to the C57Bl/6J line (Jackson Laboratory, Bar Harbor, ME). Genotyping of WT, Mlh3+/DN, and Mlh3DN/DN mice was performed using the following PCR primer pairs: forward (5’-AAGCCAAGTCTGCATGAGTA-3’) and reverse (5’-TAAATGTGCCACTGACTAAAT-3’) followed by a restriction enzyme digestion with Sau96I (New England Biolabs) at 37°C for 2–3 hours, which results in 439-bp and 263-bp fragments from the WT allele and a 702-bp fragment from the mutant allele. Fertility tests were performed by breeding Mlh3DN/DN adult males with WT females. At least 3 males of each genotyped were evaluated. Presence of a copulation plug the following morning counted as a successful mating event. Pregnancy was confirmed by gentle palpation of the abdomen after gestation day 11 or on delivery date of litters. Mus81-/- animals were generated from our breeding stock of such mice, as previously described [6]. Mice were housed and utilized under the guidance and approval of the Cornell University Institutional Animal Care and Use Committee.
Testes from adult mice were fixed in Bouin’s solution overnight at room temperature and then washed 3 x 10 min with 70% ethanol at room temperature with agitation. Fixed and paraffin-embedded testes were section at 5 μm. H&E staining was performed on Bouin’s fixed testes using standard methods. At least 6 males of each genotyped were evaluated.
Caudal epididymides were removed from adult males and placed in pre-warmed 1X PBS containing 4% bovine serum albumin. Sperm were released into solution by squeezing epididymis with tweezers and incubated for 20 min at 32°C/5% CO2. After incubation, 20 μL of sperm suspension was re-suspended in 480 μL of 10% formalin. Sperm counts were performed with a hemocytometer. At least 10 males of each genotype were evaluated for sperm counts and testis weights.
Prophase I chromosome spreads from adult testes were prepared as previously described [28,61]. For all experiments, at least 6 males of each genotyped were evaluated. Chromosome slides were then washed in 0.4% Kodak Photo-Flo 200/1X PBS for 2 x 5 min, 0.4% Kodak Photo-Flo 200/dH2O for 2 x 5 min, then air-dried for approximately 10 min and stored in -80°C or used immediately for staining. Primary antibodies used were: anti-γH2AX (Millipore, NY, #05–636 1:10,000), anti-SYCP3 (Abcam, MA, #97672, 1:5000), anti-SYCP1 (Abcam, MA, #15087, 1:1000), anti-RAD51 (Calbiochem, #PC130, 1:500), anti-BLM (generous gift from Dr. Ramundo Freire; 1:100;), anti-CDK2 (Santa Cruz, TX, sc-163; 1:250), anti-MLH3 ([61]; 1:1000), anti-RNF212 (generous gift from Dr. Neil Hunter), anti-RPA (generous gift from Dr. Jeremy Wang; 1:500), anti-MSH4 (Abcam, MA, #58666; 1:500), anti-HEI10 (Anti-CCNB1IP1, Abcam, MA # 71977) and anti-MLH1 (BD Biosciences Pharmingen, CA, #550838, 1:100). Secondary antibodies used were: goat anti-mouse Alexa Fluor 488 (#62–6511), goat anti-mouse Alexa Fluor 555 (#A-10521), goat anti-rabbit Alexa Fluor 488 (#65–6111), goat anti-rabbit Alexa Fluor 555 (#A-10520; all Invitrogen, 1:2000).
Diakinesis chromosome spreads were prepared as previously with slight modifications [61,77]. Slides were stained with 10% Giemsa for 10 mins, washed, air-dried and mounted with Permount.
All chromosome spread slides were visualized using the Zeiss Imager Z1 microscope (Carl Zeiss, Inc.). Images were captured with a high-resolution microscopy camera AxioCam MRM (Carl Zeiss, Inc.) and processed with ZEN Software (version 2.0.0.0; Carl Zeiss, Inc.).
Focus counts were performed manually by at least two people, and the results averaged before analysis. For RPA, we employed and ImageJ algorithm for automated counting, as described [69], and compared this automated calculation to manual counts for consistency. Manual and automated counts were not statistically significantly different to each other.
cDNA was synthesized from total testis RNA from wildtype C57B/6J adult males using the SuperScript III Reverse Transcriptase Kit from ThermoFisher. Mlh1 and Mlh3 open reading frames were PCR amplified from cDNA using Expand High Fidelity DNA polymerase using primer pairs AO3365 (5’GCTAGCAGCTGATGCATATGGCGTTTGTAGCAGGAG) and AO3366 (5’TACCGCATGCTATGCATTAACACCGCTCAAAGACTTTG) for Mlh1, and AO3367 (5’ACGTCGACGAGCTCATATGCATCACCATCACCATCACCATCACCATCACATCAGGTGTCTATCAGATGAC) and AO3368 (5’CGAAAGCGGCCGCGATCATGGAGGCTCACAAGG) for His10-Mlh3. Each fragment was cloned into the Spe1 site of pFastBac1 (ThermoFisher) using Gibson assembly PCR (NEB) to create pEAE393 (Mlh1) and pEAE397 (His10-Mlh3). Constructs were verified by DNA sequencing with NCBI reference sequences NM_026810.2 and NM_175337.2 for Mlh1 and Mlh3, respectively. These constructs were then modified as follows:
Sf9 cells were transfected with pEAE397 (His10-Mlh3), pEAE413 (His10-Mlh3-D1185N) and pEAE395 (MBP-Mlh1) using the Bac-to-Bac baculovirus infection system (Invitrogen). Fresh Sf9 cells were co-infected with both viruses (containing Mlh1 and Mlh3 or Mlh3-D1185N). Cells were harvested 60 hours post infection, washed with phosphate buffered saline, and kept at -80°C until use.
Cell pellets from 250 ml of cells as thawed, resuspended in 60 ml hypotonic lysis buffer (20 mM HEPES-KOH pH 7.5, 5 mM NaCl, 1 mM MgCl2, 1 mM PMSF and EDTA free protease inhibitor mixture from Roche and Thermo Scientific) and incubated for 15 min on ice. The suspension was adjusted to 250 mM NaCl, 15 mM imidazole, 10% glycerol, 2 mM ß-mercaptoethanol (BME), and clarified by centrifugation at 17,000 g for 20 min at 4°C. The supernatant was mixed with 6 ml of 50% nickel-nitrolotriaceticacid-agarose (Ni-NTA) resin and allowed to bind for 2 hours or overnight followed by centrifugation to remove the unbound fraction. The resin was packed onto a column and washed with 7–10 column volumes of wash buffer (50 mM HEPES-KOH pH 7.5, 250 mM NaCl, 40 mM imidazole, 10% glycerol, 2 mM BME, 1 mM PMSF). Protein was eluted with 15 ml of 300 mM imidazole in 50 mM HEPES-KOH pH 7.5, 250 mM NaCl, 40 mM imidazole, 10% glycerol, 2 mM BME and 1 mM PMSF. Elution fractions containing MLH1-MLH3, determined by SDS-PAGE, were pooled and loaded onto 1 ml 100% amylose resin (NEB). The resin was washed with 10 column volumes of wash buffer (50 mM HEPES-KOH pH 7.5, 250 mM NaCl, 10% glycerol, 2 mM BME, 1 mM PMSF) and eluted with 6 ml wash buffer containing 10 mM maltose. Fractions containing MLH1-MLH3 were pooled and aliquots were flash frozen and stored in -80°C. The protein yield, following amylose chromatography, was similar for wild-type and mutant complexes (approximately 120–150 μg per 250 ml cells).
It is important to note that we were unable to detect a specific endonuclease activity for the mouse MBP-MLH1-MLH3 complex, suggesting that the MBP tag interferes with MLH1-MLH3 functions. We were unable to test this directly because, despite numerous attempts, we were unable to efficiently remove the MBP tag from MLH1 by treating MBP-MLH1-MLH3 with TEV protease.
SDS-PAGE bands following amylose chromatography predicted to contain MBP-MLH1 and His10-MLH3 were excised and analyzed by the Cornell University Proteomics facility using a Thermo LTQ Orbitrap Velos Mass Spectrometer.
The majority of comparisons involved with unpaired parametric t-test with Welch's correction or nonparametric Mann-Whitney U-test, depending on the data distribution. Where necessary, Bonferroni’s adjustment was used for multiple comparisons. All statistical analysis was performed with GraphPad Prism Version 7.00 for Mac, Graphpad Software, La Jolla California USA, www.graphpad.com. P-values less than 0.05 were considered statistically significant.
|
10.1371/journal.pntd.0005819 | Widespread Trypanosoma cruzi infection in government working dogs along the Texas-Mexico border: Discordant serology, parasite genotyping and associated vectors | Chagas disease, caused by the vector-borne protozoan Trypanosoma cruzi, is increasingly recognized in the southern U.S. Government-owned working dogs along the Texas-Mexico border could be at heightened risk due to prolonged exposure outdoors in habitats with high densities of vectors. We quantified working dog exposure to T. cruzi, characterized parasite strains, and analyzed associated triatomine vectors along the Texas-Mexico border.
In 2015–2016, we sampled government working dogs in five management areas plus a training center in Texas and collected triatomine vectors from canine environments. Canine serum was tested for anti-T. cruzi antibodies with up to three serological tests including two immunochromatographic assays (Stat-Pak and Trypanosoma Detect) and indirect fluorescent antibody (IFA) test. The buffy coat fraction of blood and vector hindguts were tested for T. cruzi DNA and parasite discrete typing unit was determined. Overall seroprevalence was 7.4 and 18.9% (n = 528) in a conservative versus inclusive analysis, respectively, based on classifying weakly reactive samples as negative versus positive. Canines in two western management areas had 2.6–2.8 (95% CI: 1.0–6.8 p = 0.02–0.04) times greater odds of seropositivity compared to the training center. Parasite DNA was detected in three dogs (0.6%), including TcI and TcI/TcIV mix. Nine of 20 (45%) T. gerstaeckeri and T. rubida were infected with TcI and TcIV; insects analyzed for bloodmeals (n = 11) fed primarily on canine (54.5%).
Government working dogs have widespread exposure to T. cruzi across the Texas-Mexico border. Interpretation of sample serostatus was challenged by discordant results across testing platforms and very faint serological bands. In the absence of gold standard methodologies, epidemiological studies will benefit from presenting a range of results based on different tests/interpretation criteria to encompass uncertainty. Working dogs are highly trained in security functions and potential loss of duty from the clinical outcomes of infection could affect the work force and have broad consequences.
| Chagas disease, a potentially deadly cardiac disease of humans, canines and other mammals is caused by the parasite Trypanosoma cruzi. The parasite is primarily transmitted to dogs by ingestion of infected triatomine ‘kissing bug’ vectors or through contact with the insect’s feces. Previous studies concluded that stray and shelter dogs are at high risk of infection in the southern U.S. We proposed that high-value U.S. government working dogs along the Texas-Mexico border may also be at high risk because of their activities in regions with established, infected vector populations. We sampled 528 working dogs along the Texas-Mexico border, and found that 7.4–18.9% of dogs were positive for T. cruzi antibodies and a small proportion (0.6%) also had parasite circulating in the blood. We collected two species of kissing bugs from the canine environments and used molecular approaches to determine that 45% were positive for T. cruzi and the majority had recently fed on canines. We highlight the need for better diagnostic tools for canine Chagas disease research and diagnosis. The widespread burden of T. cruzi infection in the government working dogs could be associated with far-reaching consequences for both animal and human well-being.
| Chagas disease, a potentially deadly cardiac disease of humans and dogs, is caused by the flagellated protozoan parasite Trypanosoma cruzi. The parasite is transmitted by infected hematophagous triatomine insects, commonly known as ‘kissing bugs’. Chagas disease is estimated to infect nearly 6 million people throughout Latin America, and occurs across the southern US in enzootic cycles [1,2], where raccoons and other wildlife serve as reservoirs [2,3]. In many areas of Latin America, such as in the Gran Chaco ecosystem, domestic dogs are an important reservoir of T. cruzi and domestic vectors that fed on dogs showed higher infection prevalence than vectors that fed on other domestic hosts [4,5]. The importance of canines in the T. cruzi transmission cycle in the US is not yet understood.
The occurrence of T. cruzi infected canines in the USA is especially high in the state of Texas [1,6,7], where 439 cases were reported across 58 counties between 2013–2015 when there was mandatory reporting of T. cruzi infected dogs [8]. Texas harbors at least seven established species of triatomine vectors capable of transmitting T. cruzi [3] and infected wildlife are widespread [1]. The high frequency of canines infected with T. cruzi likely reflects robust enzootic transmission in the state. Outside of Texas, dogs infected with T. cruzi have been reported in Louisiana [9,10], Oklahoma [11,12], Tennessee [13] and Virginia [14]. Across the studied populations, apparent seroprevalence ranged from 3.6–57.6% and predispositions of infection status with certain breeds or types of dogs do not appear to be strong, with hunting dogs, working dogs, household pets, shelter and stray dogs all impacted [6,7,9,12,14,15].
T. cruzi infection can occur by vector-mediated transmission through the introduction of infected bug feces into the bite site or mucous membrane or through the ingestion of infected bugs or their feces [5]. Additionally, congenital transmission may occur [3]. Dogs are more likely to become infected than humans [16,17], which could be from dog’s affinity to consume bugs [12,18–21]. T. cruzi-infected dogs may be asymptomatic or may develop debilitating acute or chronic cardiac disease, characterized by myocarditis, hepatomegaly, ascites, cardiac dilatation, or sudden death [22]. There are currently no vaccinations or approved anti-parasitic treatments for T. cruzi infections in dogs in the US, and infected dogs are treated symptomatically.
The Department of Homeland Security (DHS) of the US government manages over 3,000 working dogs in various capacities including the Transportation Security Authority, Coast Guard, Secret Service, Federal Protective Services, Customs and Border Protection, and Federal Operations. These dogs are highly trained in working duties performed in the indoor and outdoor environment including search and rescue functions as well as detection of concealed persons, narcotics, or explosives. DHS working dogs may be at increased risk for contact with vector species from working and sleeping outdoors. Some of the working dogs are kept in group kennels, which have previously been shown to be a risk factor for T. cruzi infection [7]. Their working environment could further be an attractant to the vector, where there is high vehicle traffic emitting CO2- a known attractant [23], bright lights at night, and concentrations of animals and people in otherwise rural areas. In order to provide a baseline for conducting clinical assessments and developing disease management strategies, we conducted a seroepidemiological investigation to quantify the prevalence of T. cruzi infection in populations of working dogs along the Texas-Mexico border. Additionally, we aimed to determine the infection status and feeding patterns of triatomine vectors in the environments where these dogs work and are kenneled.
All canine samples were collected in adherence with animal use protocols approved by Texas A&M University’s Institutional Animal Care and Use Committee on 08/17/2015 under the number 2015–0289. Written consent was received for each canine sampled from DHS personnel.
Sampled DHS working dog breeds were predominantly Belgian Malinois and German Shepherds. Most dogs were bred in Europe, and less commonly dogs came from vendors within Texas or other parts of the US. Dogs receive over 6 months of training at either a training facility in El Paso, Texas, or Front Royal, Virginia, and specialize in various jobs such as track and trail, detection of humans, narcotics, currency, or agricultural products, and search and rescue. After training, dogs are typically assigned to a specific management area and have limited travel. The dogs in our study perform working duties either immediately adjacent to the geopolitical border (ports of entry) or north of the border (checkpoints). Off-duty canines are either kenneled individually at their handler’s residence or in a group kennel. Residential kennels are indoor-outdoor metal kennels raised 2 feet from the ground, giving the dog the option of sleeping inside or outside. Group kennels are indoor-outdoor, concrete kennels, and dogs are confined inside during the night.
We used a cross sectional study design to collect blood samples from DHS working dogs during November 2015 and April 2016. Working dogs were sampled from all 5 management areas, with a goal of sampling at least 60% of the dogs that occurred within each management area. Additionally, we sampled DHS canines that were in training at a training facility in management area #1 (Fig 1). Sample criteria included dogs over 6 months in age and on active duty or in training. Demographic information was collected on all dogs sampled including age, sex, breed, canine job, sleeping location and station of duty. A minimum of 1 ml of blood was collected by venipuncture and aliquoted into serum and EDTA tubes.
Samples were screened for anti-T. cruzi antibodies by Chagas Stat-Pak rapid immunochromatographic test (ChemBio, NY) which was designed for use in humans and has been validated in dogs [9]. Stat-Pak assay uses three T. cruzi recombinant antigens that are bound to the assay membrane solid phase. Serum or plasma samples were tested according to manufacturer’s protocol and read for result determination after 15 minutes. Tests were considered negative when no color developed and positive when a clear line developed. Additionally, very faint bands that were not perceptible enough to be consider a clear positive, yet with some low level of color development to differentiate them from negative, were tracked as ‘inconclusive’ and subjected to additional testing.
All positive or inconclusive samples as determined by Stat-Pak plus 10% of the negatives were tested by both indirect fluorescent antibody (IFA) test and Trypanosoma Detect (InBios, International, Inc., Seattle, WA). IFA detects anti-T. cruzi IgG antibodies and was performed by the Texas Veterinary Medical Diagnostic Laboratory (TVMDL, College Station, TX) on serum or plasma samples. Titer values of 20 or higher were considered positive per TVMDL standard protocol; this titer value cutoff has also been used in human medicine [24]. IFA readers were blinded to previous serologic results. Trypanosoma Detect is a rapid immunochromatographic dipstick assay that employs a multi-epitope recombinant antigen for the detection of anti-T. cruzi antibodies. The Trypanosoma Detect test was designed for use in humans but has been found to have high sensitivity and specificity for use in dogs [25]. Serum or plasma were tested according to manufacturer's protocol and read for result determination after 20 minutes. Test results were scored as positive, inconclusive, or negative using the same criteria as described above for the Chagas Stat-Pak. Serological positive status was assigned to samples that tested positive on at least two independent tests.
Amplification of parasite DNA from blood samples by real time PCR was performed on all sampled dogs. DNA was isolated from 250 uL of buffy coat by using E.Z.N.A. Tissue DNA kit (Omega Bio-Tek, Norcross, GA). Negative controls (phosphate buffered saline or water template) were included in the DNA extractions and the PCR. To determine if analysis of clot rather than buffy coat may result in a greater ability to detect parasite DNA, we conducted additional work with a subset of samples as follows. From 12 dog samples, we extracted DNA from 1 mL of clot for PCR analysis. These 12 dogs comprised 10 that were seropositive and PCR negative based on buffy coat; 1 that was seropositive and PCR positive based on buffy coat; and 1 that was seronegative and PCR positive based on buffy coat analysis.
Samples were first screened for presence of T. cruzi satellite DNA using the Cruzi 1/2 primer set and Cruzi 3 probe in a real-time assay to amplify a 166-bp segment of a repetitive nuclear DNA [26, 27]. Reactions consisted of five microliters of extracted DNA, primers I and II each at a concentration of 0.75 μM, 0.25 μM of probe, and iTaq University Probes Supermix (BioRad Laboratories, Hercules, CA), in a 20 μL reaction volume. Previously published thermocycling parameters were followed except with a 3-minute initial denaturation using a Stratagene MxPro3000 (Agilent Technologies, Santa Clara, CA). T. cruzi DNA extracted from isolate Sylvio X10 CL4 (ATCC 50800, American Type Culture Collection [ATCC]) was used for a positive control. Machine-calculated thresholds and reaction curves were visually checked for quality. Samples with Ct values less than 34 were considered suspect positive and subjected to further testing.
Suspect positive samples by qPCR were run on a second, independent PCR using T. cruzi 121/122 primers to amplify a 330-bp region of kinetoplast DNA [28,29]. Reactions included 1μL template DNA, primers at final concentrations of 0.75 μM each, and FailSafe PCR Enzyme Mix with PreMix E (Epicentre, Madison, WI) in a final reaction volume of 15 μL. Amplicons were visualized on 1.5% agarose gels stained with GreenGlo safe DNA dye (Denville Scientific Inc., Metuchen, NJ). Samples that yielded a band of the appropriate size were interpreted as positive in this assay. Parasite positive dogs were defined as those that tested positive on both the rt-PCR screening and the secondary PCR assays.
We used a multiplex quantitative, real time PCR to determine T. cruzi discrete taxonomic unit (DTU) of samples that were positive or suspect positive on the screening assay based on amplification of the nuclear spliced leader intergenic region (SL-IR) [30]. Using a QIAGEN Multiplex PCR Kit (QIAGEN, USA) reactions were performed using 2μL template DNA in a final volume of 20 μl and run on a BioRad CFX96 (Hercules, CA, USA). The only deviation from the previously described protocol was the extension of cycles from 40 to 45 and substitution of dyes as previously described [7]. Positive controls consisted of DNA from triatomines collected across Texas that were previously characterized as infected with TcI or TcIV based on amplification and sequencing of the TcSC5D gene [31]. Samples with Ct values less than 34 were considered positive, and fluorescence signal determined the strain type.
Triatomine bugs were opportunistically collected by dog handlers in summer 2016 from group kennels, outside handler’s residence around canine housing, and at stations where dogs worked. To encourage collections, outreach materials with photos of triatomines and look-alike species were disseminated by email and in printed format to dog handlers prior to the summer peak of adult triatomine activity. Bugs were identified to species using morphologic features [32] and sexed. After bugs were washed in 10% bleach solution and rinsed in distilled water, sterile instruments were used to dissect the bugs, isolate hindgut material and evidence of a recent bloodmeal was noted. DNA was extracted from hindguts and tested for T. cruzi DNA and determination of T. cruzi DTU using the same methods as the above testing of dog samples. In order to determine the source of recent bloodmeals, hindgut DNA was subjected to PCR amplification of vertebrate cytochrome B sequences using previously published primers and cycling conditions [33,34]. Reactions included 3 μL template DNA, primers at final concentrations of 0.66 μM each, and FailSafe PCR Enzyme Mix with PreMix E (Epicentre, Madison, WI) in a final reaction volume of 50 μL. Amplicons were visualized on 1.5% agarose gel, prepared for sequencing using ExoSAP-IT (Affymetrix, Santa Clara, CA, USA), and Sanger sequencing was performed (Eton Bioscience Inc., San Diego, CA, USA). Resulting sequences were compared to existing sequences using Basic Local Alignment Search Tool (National Center for Biotechnology Information, US National Library of Medicine). In recognition of the potential for contamination from the environment, samples that aligned to human were re-run on another PCR assay to provide a secondary line of evidence.
Due to the uncertainty of sample serostatus associated with the inconclusive band development, antibody-positive dogs were defined using two methods; a) in the conservative method, inconclusive band development was interpreted as negative, and b) in the inclusive method, inconclusive band development was interpreted as positive. In the absence of gold standard serological methodology, these two different criteria of positivity (method A and B) were analyzed separately to provide a range of results.
To evaluate the relationship between potential risk factors and the serostatus of canines, data were imported into R software [35] for analysis. Assessed variables were dog age (young = 6 months to <3 years, middle age = ≥ 3 years to <6 years, senior = ≥ 6 years), sex, breed, sleeping location (individual kennel at handler’s residence or group kennel) and management area (locations 1–5 or training center). Due to the small sample size of dogs in some jobs, canine job was dichotomized based on type of detection. Bivariable analysis using the chi-squared or Fisher’s exact was used to identify putative risk factors. Factors with a p≤ 0.25 from the initial screening were used in a logistic regression model, while controlling for management area as a random effect. Generalized linear mixed models were calculated and factors with values of p < 0.05 were considered significant. Odds ratios and 95% confidence intervals were calculated. To determine variation in seroprevalence across management areas, a logistic regression model was used in which the training center served as the referent to which all five management areas were compared. Kappa index was used to test the agreement between each pairwise combination of the results of the three serological assays for the samples that were tested on all three assays; this sample set was biased toward Stat-Pak positive samples.
A total of 528 dogs from along the Texas-Mexico border were evaluated using a variety of serologic and molecular techniques to detect T. cruzi exposure and infection. Distribution of samples among the five management areas ranged from 47 (8.9%) to 135 (25.6%), and 86 (16.3%) dogs were sampled from the training center. The most common breeds were Belgian Malinois and German Shepherd, which together comprised 86% of the sampled dogs, with Dutch Shepherds, Sable Shepherds, Groenendael and Labrador Retriever comprising the remainder. Age ranged from 6 months to 13 years with a median of 4.47 and a mean of 4.79. There were 351 males (66.5%) and 177 females (33.5%). Of the dogs sampled, 55.9% spend their off-duty time in individual residential kennels whereas 44.1% were group kenneled. The sample sizes of dogs within each canine job category or management unit are not disclosed because it is law enforcement sensitive information.
In considering inconclusive bands on immunochromatographic tests as negative, 39 of 528 (7.4%) of dogs were seropositive for antibodies to T. cruzi on at least 2 assays. Across management areas and the training center, seroprevalence ranged from 4.3% to 10% (Fig 1). In the bivariable analysis, T. cruzi seroprevalence was significantly different across dog breed (p = 0.03), with seroprevalence of German Shepherds being lowest (3.7%) and ‘other’ breeds being highest (14.3%; Table 1). Dogs that spent off-duty time in residential kennels had a significantly higher seroprevalence (29/295, 9.8%, p = 0.02) than those that were group-kenneled (10/233, 4.3%). Seroprevalence was significantly different among age groups (p = 0.04), where senior dogs had a seroprevalence of 10.4%, middle age dogs a seroprevalence of 7.9% and young dogs 3.2%. Seroprevalence did not vary significantly by sex or canine job.
Multivariable logistic regression analysis showed a significant association (odds ratio [OR] 0.41, 95% CI 0.17–0.99, p = 0.047, Table 2) between breed and seropositive dogs, after controlling for management areas as a random effect (Table 2), in which German Shepherds were associated with a significantly lower seroprevalence (3.7%) than Belgian Malinois (8.6%). No significant association was found between age, job, or sleeping location and seroprevalence.
In considering inconclusive bands on serologic tests as positive, 100 of 528 (18.9%) of dogs were seropositive for antibodies to T. cruzi on at least 2 assays. Seroprevalence ranged from 11.6% to 26.7% across management areas and the training center (Fig 1). When running bivariable analysis, dogs that spent off-duty time in residential kennels (65/295, 22%) were marginally (p = 0.09, Table 2) more likely to be seropositive than dogs sleeping at a group kennel (36/233, 15.4%). Seroprevalence did not vary significantly by age, breed, sex or canine job.
Multivariable logistic regression analysis showed that there was no association between age, job, or sleeping location and seroprevalence. Backwards elimination was performed and when only age was included in the model there was a marginal association in which old dogs had a higher seroprevalence (39/182, 21.4%) than young dogs (22/156, 14.1%; p = 0.09), after controlling for management areas as a random effect.
While seroprevalence did not significantly differ across management areas and the training center when positivity was defined according to Method A, dogs from management area #2 (OR 2.6, 95% CI 1.0–6.7, p = 0.04) and #3 (OR 2.8, 95% CI 1.2–6.8, p = 0.02) had significantly higher seroprevalence compared to the training center when seropositivity was determined according to Method B (Table 3). This indicates that area #2 and #3 were both associated with many samples that produced very faint (inconclusive) bands on the immunochromatographic tests.
In comparing the results across all three serological testing platforms (Table 4), all IFA positive samples are positive on Trypanosoma Detect, and all but two samples are Stat-Pak positive-both of these samples having a titer of 20. When comparing the IFA negative samples 71.3% are positive or inconclusive on Stat-Pak and 48.4% are positive or inconclusive on Trypanosoma Detect. From the 528 dog samples in the study, 215 samples were tested on all three serology assays. Overall test agreement ranged from slight to moderate agreement based on the Kappa Indices (Table 5), with agreement between tests being better when interpreting immunochromatographic test results using the conservative method A (kappa range 0.37–0.48) compared to inclusive method B (kappa range 0.05–027). The best agreement was using method A between Stat- Pak and Trypanosoma Detect, with a Kappa index of 0.48 (moderate agreement).
Of the 57 randomly-selected Stat-Pak negative samples that were subjected to additional serologic testing, one was positive on both IFA (titer 20) and Trypanosoma Detect; this sample was counted as positive in the seroprevalence estimates. Nine (15.8%) samples that were both Stat-Pak and IFA negative were positive on Trypanosoma Detect; these dogs were counted as negative in the seroprevalence estimates, but could be false negatives. When applying this prevalence of potential false negatives to the total number of dogs that were negative by Stat-Pak, an additional 49 dogs are extrapolated to be potential false negatives; including these samples as positive would increase seroprevalence to 15.9% (84 dogs total) by conservative method A, and 25.4% (149 dogs total) by inclusive method B.
Inconclusive bands were reported from 108 (20.5%) samples screened on Chagas Stat-Pak. When tested on IFA only 1 (0.9%) inconclusive tested positive with a titer of 20. When inconclusive samples were run on Trypanosoma Detect, 37 (29.6%) had inconclusive bands on Trypanosoma Detect, 20 (18.5%) were positive, and 51 (47.2%) were negative.
T. cruzi DNA was detected in the buffy coat fraction of the blood in three of 528 (0.6%) dog samples according to our diagnostic method which included amplification in both a screening and confirmatory assay. The first PCR-positive dog was sampled from area # 5 in November and was positive for antibodies by all three serology assays with a relatively high titer (640) on IFA. Using the multiplex real time PCR to determine T. cruzi DTUs, we found that this dog harbored DTU TcIV. The second PCR-positive dog was from the canine training center, sampled in April, positive on all serology assays with a titer of 320 and harbored a mix TcI/TcIV. The third dog was from area # 2, sampled in April, was negative by all serological assays, and strain type could not be determined. When this PCR positive yet serologically-negative dog was included in binomial analysis of risk factors and the logistical regression model, no difference was found in significant associations. The subset of 12 samples that were subjected to an additional DNA extraction from 1mL of clot produced PCR results that were identical to the results obtained from the 250 uL buffy coat extractions with the exception of the sample from the seronegative, buffy coat-positive sample. This sample was negative based on clot analysis.
In the summer of 2016, a total of 20 adult triatomine bugs of two species (18 Triatoma gerstaeckeri and 2 T. rubida) were opportunistically collected by canine handlers from three management areas (Table 6). Kissing bugs were collected from stations where dogs and handlers work (n = 6), handler’s residence near canine housing (n = 7), group kennels (n = 4), from the field (n = 2) and 1 bug was removed from a dog while working. Nine (45%) triatomines were positive for T. cruzi including half of the T. gerstaeckeri specimens but neither of the two T. rubida specimens. Of the 9 positive bugs, parasite strain typing revealed DTU TcI in 6, TcIV in 1, and a mixed TcI/TcIV coinfection in 2. From dissection, 13 of the 20 bugs had evidence of a recent blood meal in their hind gut, and 11 of these yielded results after the blood meal analysis protocols, revealing human, canine, coastal-plain toad (Bufo nebulifer) and rat (Rattus rattus) DNA (Table 6).
We found widespread T. cruzi infection in government working dogs along the Texas-Mexico border. DHS working dogs play an important role in detection and security functions in the Unites States and the clinical manifestation of infection may be associated with significant future economic and security consequences. We are aware of only two prior epidemiological investigations of T. cruzi infection in working dogs in the US. In 2007, a serological survey was conducted on military working dogs (MWD) in San Antonio, TX, after veterinarians noted an increase in Chagas disease diagnoses, revealing 8% of the kenneled dogs were positive by IFA [36]. Such findings are of utmost importance in these dogs; in 2009, MWDs deployed in Iraq were evacuated due to cardiac symptoms and diagnosed with T. cruzi infection leaving troops vulnerable without explosive detection dogs [36]. Recently, populations of working hound dogs in south central Texas that are used for scent detection and track/trail were characterized with an extremely high seroprevalence of 57.6% (n = 85) in which positive dogs were reactive on both Stat-Pak and IFA [7]. The study population also included many dogs with parasite DNA in the blood and other organs, and infected triatomines collected from the dog kennels were determined to have fed on dogs, allowing the authors to conclude that multi-dog kennels can be high risk environments of T. cruzi transmission [7].
Exposed dogs were present in all five management areas and the canine training school, with an overall apparent seroprevalence of 7.4–18.9%. This seroprevalence is similar to that reported from dogs in Chagas-endemic areas in Latin America including populations in Peru (12.3%) [37], Argentina (45.6%) [38], Panama (11.1%) [39], Costa Rica (27.7%) [19], Yucatan State, Mexico (9.8%-14.4%) [40] and Mexico State, Mexico (10%-15.8%) [41]. Previous epidemiological investigations of T. cruzi in canines in the US are limited, and most have focused on stray dogs or those sampled from animal shelters, which may be considered as high risk populations due to outdoor activity. A serosurvey of high risk kenneled dogs in southern Louisiana found that 22.1% [9] of dogs tested positive for T. cruzi antibodies using the same three serology assays performed in this study. A study in Oklahoma sampling shelter dogs and pet dogs concluded that 3.6% dogs were seropositive when testing by radioimmunoprecipitation assay (RIPA) [12]. Earlier studies in southern Texas stray dogs 375 dogs were tested and 7.5% were positive by indirect immunofluorescence [15]. Similarly, across Texas shelter dogs had a seroprevalence of 8.8% when testing dogs on Chagas Stat-Pak [6]. These studies and ours suggest that despite the regular veterinary care, quality food and shelter, highly-valued working dogs can have similar or greater T. cruzi infection than stray and shelter dogs in the US and free roaming or pet dogs in endemic countries.
Both population-level and individual-level T. cruzi studies of naturally-infected hosts suffer from a lack of gold standard tests or diagnostic recommendations. Discordance among tests results is prevalent in human and veterinary Chagas diagnostics. For example, a study looking at seroprevalence in people from Veracruz, Mexico used 5 assays and found that test agreements ranged from 0.038–0.798 on the Kappa index [42]. Similarly, using the Kappa index we found a high discordance among serology assays used, with agreement ranging from slight to moderate depending on the interpretation method. Assay discordance could be affected by the single freeze-thaw cycle, or the age of the sample. These diagnostic challenges make it difficult to directly compare seroprevalence across populations and diagnostic methods, and presents a challenge in clinical settings for diagnosis. Two of the three serological tests we used are only available for research use for dogs in the US, and a limited number of commercial laboratories offer canine T. cruzi diagnostic test services. As in most diagnostic tests, there is some subjectivity in the interpretation of results, and the development of very faint ‘equivocal’ serological bands on both Chagas Stat-Pak and Trypanosoma Detect posed particular complexities in our analysis. The Stat-Pak and Trypanosoma Detect instructions state that band intensity will vary, but faint bands should be interpreted as positive [43,44] and that variation is dependent on the concentration of antibodies present [44]. However, some previous canine studies have counted faint bands as negative [6,9] while others have interpreted them as positive for analysis [45]. Our presentation of a seroprevalence calculated both conservatively (very faint bands interpreted as negative) and inclusively (very faint bands interpreted as positive) is an effort to account for imperfect diagnostics. Until refined T. cruzi diagnostic tools are available, we encourage transparency in presenting results on single vs. multiple tests across all strengths of test response.
The discordance between test results and the within assay variation could be caused by parasite heterogeneity [45]. T. cruzi is notably heterogeneous with seven major genotypes or discrete typing units (DTUs) described as TcI-TcVI and TcBat which vary be region [46,47]. Additionally, a notable intra-DTU variability has been found [47,48]. Previous research has found that assay reactivity varies by geographic origin of the patient [49]. O’Connor and others found that strain TcI clusters geographic between North and South America [50]. The Chagas Stat-Pak was validated with human sera from Central America to detect strains circulation in that region [51] and may not be optimized for T. cruzi clones from Texas. When very faint bands were interpreted as positive (method B), seroprevalence was significantly higher in two western management areas (OR 2.6–2.8, 95% CI: 1.0–6.8 p = 0.02–0.04) compared to the training center (Table 3), whereas this difference was not evident when very faint bands were interpreted as negative (method A). The disproportionate abundance of very faint bands in this geographic area may be driven by differences in the locally-circulating T. cruzi clones. Diosque et al. performed a genetic survey of T. cruzi isolates within a restricted geographical area (~300 km2) and found five different clones circulating [52]; such findings are clinically and diagnostically relevant because parasite heterogeneity has been shown to cause varying infectivity and immune response [52–54]. In addition, host biological factors (exposure history, coinfection, genetic makeup) could also cause reaction variability within and across serology assays.
Sleeping location (group housed indoors vs. individually housed outdoors) appeared to be independently associated with T. cruzi status with a higher seroprevalence in dogs sleeping outdoors than indoors by method A (p = 0.02), and marginally significant by method B (p = 0.09) in bivariable analysis. Previous studies have indicated dogs housed outdoors where vector contact is more likely to be at a higher risk for exposure [6,9,39]. Dogs in Tennessee spending 100% of their time outdoors were significantly more likely to be seropositive for T. cruzi than dogs spend ≤50% of their time outdoors [13]. Seroprevalence did increase with age in both method A and B, but was only significant in bivariable analysis in method A, where senior (>6 years) and middle age (≥ 3 years to <6 years), were more likely to be seropositive than young dogs (<3 years old) (Table 1). This is anticipated in infectious disease since exposure increases with age and has been found in previous studies [7,13,17,55]. We found that German Shepherds were associated with a significantly lower seroprevalence (3.7%) than Belgian Malinois (8.6%) in our study; although the driving factors for this difference are currently unknown, it may relate to host behavior, differences in host immune response, or a physical characteristic.
We found three dogs (0.6%) harbored parasite PCR in their blood, suggesting that these dogs are parasitemic. While two of the three PCR-positive dogs also harbored detectable anti T-cruzi antibodies, one did not, suggesting this dog may have been in the acute stage of infection [56]. The two dogs with successfully typed infections harbored DTUs TcIV and a TcI/TcIV mix, consistent with previous studies on dogs in the US [57,58]. Both strain types infect a variety of hosts and vectors in the southern US [3]. DTU TcI is an ancient strain found throughout South and Central America and the predominant strain infecting humans in the US [3], where it is also associated with wildlife reservoirs including opossums (Didelphis virginiana) [57]. TcIV is also associated with wildlife, especially raccoons (Procyon lotor) [3] and to our knowledge has not been implicated in the small number of typed human infections in the US. This study found a lower prevalence of dogs PCR positive then previous studies, which likely reflects the time of sampling (November and April) when the vector is less active and dogs in Texas are less likely to come in contact with the kissing bug [58]. In recognition of other datasets that have shown that analysis of clot, rather than buffy coat, may afford a greater the chance of detecting parasite DNA [59], we subjected 12 clot samples to PCR and compared results to previous results from analysis of buffy coat. We found that buffy coat and clot results were identical across this subset with the exception of a sample from a single seronegative dog which was positive from buffy coat and negative from clot. Based on this small comparison trial, we suggest that the low frequency of encountering PCR-positive dogs in our study was not due to the blood fraction used in the analysis.
We found an infection prevalence of 45% in the kissing bugs collected from areas where the working dogs frequent, including kennels, stations and handler’s residence, including DTUs TcI, TcIV, and TcI/IV mix. This infection is slightly lower than previous estimates across the state of Texas of 63% and 51% [59,60]. Bloodmeal analysis revealed canine, human, and wildlife DNA within the hindguts of these insects, underscoring the generalist feeding strategies of triatomines that often use the most locally abundant hosts. Strict protocols were used to reduce the risk of contamination of samples by exogenous DNA (i.e., human DNA), including surface sterilization of vectors and dissection of the hindguts. It is biologically plausible that the insects associated with suspected human blood feeding encountered humans at their residence or station or work. A study in California and Arizona that collected bugs by light traps found that 5 of 13 bugs (38%) bugs were positive for a human blood meal, 4 fed on canine and 1 each for rat, pig, chicken and mouse [61]. In Texas, Gorchakov et al. found 65% (n = 62) of bugs positive for human bloodmeal and 32% for canid bloodmeal [62]; in contrast, Kjos et al. found only 1% of vectors (n = 96) collected from residential settings had fed on a human, and 20% on dogs [63]. Larger sample sizes of engorged vectors from the working dog environments will assist in learning the local vector-host interactions that sculpt disease risk.
Using dogs as sentinels has been suggested for targeted vector control programs endemic areas such as Peru [37] and to monitor transmission in Argentina [55]. However, the relative importance of dogs as reservoirs, and whether or not they can be a sentinel species for human disease risk in the US, is unknown. Further, because the triatomines in the US tend not to be colonized within homes, dogs are less likely to be useful sentinels at the household level. Nonetheless, given these infected working dogs signal the presence of infected vectors in the environment, there are public health implications of these findings especially with respect to the human handlers who are exposed to the same environments.
Because not all T. cruzi-infected dogs will develop disease [21], the prognosis and clinical implications of the widespread presence of T. cruzi-infected government working dogs along the US-Mexico border is unknown. Nonetheless, the potential loss of duty days resulting in an inadequate canine workforce must be considered. Additionally, given that the canine training school in west Texas (Fig 1) occurs in an area where triatomines are endemic, vector and canine surveillance must be conducted to determine if young dogs may be exposed to the parasite while in training, which would not only have implications for the health of the dog but also potentially afford dispersal of the parasite to the new areas across the US where these dogs are stationed. Understanding the epidemiology of T. cruzi infection is the first step toward implementing control measures to protect the health of these high-value working dogs.
|
10.1371/journal.pbio.0050237 | High-Throughput In Vivo Analysis of Gene Expression in Caenorhabditis elegans | Using DNA sequences 5′ to open reading frames, we have constructed green fluorescent protein (GFP) fusions and generated spatial and temporal tissue expression profiles for 1,886 specific genes in the nematode Caenorhabditis elegans. This effort encompasses about 10% of all genes identified in this organism. GFP-expressing wild-type animals were analyzed at each stage of development from embryo to adult. We have identified 5′ DNA regions regulating expression at all developmental stages and in 38 different cell and tissue types in this organism. Among the regulatory regions identified are sequences that regulate expression in all cells, in specific tissues, in combinations of tissues, and in single cells. Most of the genes we have examined in C. elegans have human orthologs. All the images and expression pattern data generated by this project are available at WormAtlas (http://gfpweb.aecom.yu.edu/index) and through WormBase (http://www.wormbase.org).
| Knowing where a protein is expressed provides an important clue about its potential function. As critical as this information is, we have complete developmental expression profiles for only a small fraction of all genes expressed in any metazoan. Here, we have generated spatial and temporal tissue expression profiles for 10% of all genes in the nematode Caenorhabditis elegans. Worms expressing putative gene regulatory elements fused with green fluorescent protein were analyzed at each stage of development from embryo to adult. Among the regulatory regions identified are sequences that regulate expression in all cells, in specific tissues, in combinations of tissues, and in single cells. Most of the genes we have examined in C. elegans have human orthologs. Our analysis of complex expression patterns for so many genes may not only facilitate functional analysis in C. elegans, but also create a foundation for decoding the informational hierarchies governing gene expression in all organisms.
| Determining when and where genes are expressed is often key to determining their function. Although expression profiling of genes using Serial Analysis of Gene Expression (SAGE) and microarrays is now routine, we still have complete developmental expression profiles for only a small fraction of all genes expressed in any metazoan. The spatial resolution of these two techniques is limited unless purified cell populations can be isolated in sufficient abundance to provide the necessary RNA (for examples, see [1–3]). How then do we gain expression information on the thousands of human genes that are still largely uncharacterized? One approach is to use high-throughput RNA in situ hybridization as has recently been done for brain tissue in the mouse [4]. In this study, 20,000 genes were assayed in the adult male mouse brain, and their distribution in many cases was resolved to the level of a single cell. Another complementary approach involves employing green fluorescent protein (GFP) [5] as a marker to monitor gene expression in a specific cell or tissue. The GenSAT project [6] uses Bacterial Artificial Chromosomes (BACs) with GFP-marked genes in transgenic mice to monitor tissue and cell expression. About 2,000 gene expression patterns are described at the GenSAT site (http://www.gensat.org/). Because gene functions were largely maintained during evolution, yet another possible approach is to first study orthologs of these genes in less complex organisms. Knowing what tissue or cell type expresses a particular gene in a simpler system such as Caenorhabditis elegans or Drosophila melanogaster could help drive the analysis of this gene in a more complex tissue or organ system, as is found in mice and humans. In Drosophila, a large-scale in situ hybridization study has now documented the expression pattern of close to 3,000 genes in the developing embryo ([7]; http://www.fruitfly.org/cgi-bin/ex/insitu.pl). The goal of our study was to characterize the temporal and spatial expression pattern of human orthologs in the nematode C. elegans down to the resolution of a single cell. Specifically, we determined the expression profile of individual genes throughout the whole organism and across all life stages. Independent of the biomedical aspects of our approach, the analysis of complex expression patterns of many genes may not only facilitate functional analysis in C. elegans and other organisms, but also create a foundation for decoding the informational hierarchies governing gene expression.
C. elegans has several advantages as a venue for expression studies at this resolution. The main advantages are that it is one of the simplest multicellular organisms with a complete genome sequence available [8] and a completely documented cell lineage [9,10]. In addition, the small size, transparency, and limited cell number of the worm allow for the easy observation of many complex cellular and developmental processes that are difficult to observe in higher eukaryotes, and morphogenesis can be observed at the level of a single cell [11].
Besides ourselves, only two groups have attempted large-scale expression profiling in C. elegans at this resolution. Hope and colleagues in the past have used lacZ reporters and currently are using the newly developed “promoterome” to characterize gene expression [12–14]. Another approach, developed by Yuji Kohara's group in Japan, uses in situ hybridization to fixed animals at different developmental stages (http://nematode.lab.nig.ac.jp). Our approach was to examine expression in living animals transformed with GFP fused to DNA 5′ of genes with human orthologs. For gene fusion and amplification, we used “PCR stitching” [15], which proved to be a fast, efficient, and economical method for obtaining such constructs, and we have demonstrated that the method is scalable [1]. Because of the relatively small intergenic regions in the C. elegans genome, typically less than 3 kb, PCR stitching did not have to be done over large intervals. These small intergenic intervals illustrate yet another advantage of doing this type of study in the nematode. This is a key advantage that sets our project apart from previous high-throughput expression projects done in other organisms. Our overall approach takes advantage of the transparency of the nematode and allows us to visualize gene expression in vivo, in real time, in a living animal. This method allowed us to determine the temporal and spatial distribution of the expressed GFP in close to 10% (1,886) of all genes identified in this organism.
Expression patterns analyzed for the 1,886 genes in this study were primarily, but not exclusively, from nematode orthologs of human genes (>80%). Our target genes were drawn from nematode–human ortholog groups in the InParanoid database [16] (http://inparanoid.sbc.su.se), selecting primarily genes for which no function is known. To analyze the in vivo spatial and temporal expression profiles of thousands of genes, we needed a high-throughput approach for GFP fusion constructs. GFP has been shown to be an effective cell marker in C. elegans [5,17], and because of the need for cost-effectiveness and scalability, we chose to use the promotor::GFP fusion technique “PCR stitching” [15]. The 5′ regulatory regions examined in this study extend a maximum of 3 kb upstream of the predicted ATG initiator site for a targeted gene. Most often, an upstream gene was nearer than 3 kb and we did not extend our analysis into or past this adjacent gene. As a benchmark and internal control, 10% of our analysis included genes with expression annotation in WormBase. We used half of these benchmark genes and found that 80% of our observations on expression matched the annotated expression patterns. For another 10% of the benchmark genes, we found some overlap, and for about 10%, we found little or no agreement with expression patterns compiled at WormBase. (Table S1).
Transformants carrying GFP fusions were subject to detailed in vivo analysis as outlined in Figure 1. We have observed GFP expression for 1,886 genes. Because we only sampled 10% of the genes in this organism, we wanted to ensure that specific functional categories were not overrepresented in our dataset. We used Gene Ontology (GO) annotation to examine the genes in our set relative to the whole genome and found that the representation of most functional groups reflected their frequency within the genome (Figure S1). Besides the genes for which we detected expression, there were another 516 genes for which we did not detect any expression (see Discussion). At present, only 15% of the strains exhibiting expression are in stable strains (possibly chromosomal integrants). As is usual for microinjected transgenes, most strains carry unincorporated concatamer arrays, and we detected mosaicism in many of these strains. To compensate for this mosaicism, and to ensure that we did not miss expressing cells, at least 20 replicates were analyzed for each developmental stage. Only GFP-expressing cells and tissues that showed consistent expression in 50% of the animals at any given developmental stage were recorded.
Two subclasses of expressing strains were further analyzed: (1) those with rare or complex expression patterns and (2) those that showed embryonic expression before the comma stage of embryogenesis. In the former case, the strains underwent their final analysis via 2-D and 3-D imaging on a confocal microscope before being submitted to the public Web site. In the latter case, the embryonic strains were first integrated (see Materials and Methods) and then recorded during development using a four-dimensional (4-D) microscope system (multifocal, time-lapse video recording system) developed for the purpose of tracking embryonic cell identities and movements [18,19]. Since the cell lineage of C. elegans is invariant [10], we could use these recordings in conjunction with Simi BioCell software [19] to retrace the cell lineages and determine the identity of the cells expressing GFP. This has resulted in 95 embryonic recordings, two examples of which are illustrated in Figure 2. In the first, pC45G9.13 (Figure 2A–2D), expression is initially detected in three cells, ABprappppa, M5, and MSpapaapa, but later expands to include several other cells. In the second example, pZK637.11 (Figure 2E–2H), expression is detected early during embryogenesis, and includes the AB and MS lineages. At present, only a portion (10%) of the embryonic recordings have been completely analyzed and the lineage of all GFP expressing cells determined.
The data from this project are publicly available at WormBase and interactively at WormAtlas (http://gfpweb.aecom.yu.edu/index). All strains are available from the Caenorhabditis Genetics Center (http://www.cbs.umn.edu/CGC/CGChomepage.htm) (currently, strain requests go through R. Johnsen [[email protected]]). Our Web site (http://gfpweb.aecom.yu.edu/index) provides the user with two formats for accessing the data: (1) a Browse page (Figure 3) to display all strains and data, with a search option for stage or tissue, and (2) a Gene Search page (Figure 4) that enables the user to recover selected information on specific genes of interest, or identify a subset of genes from the entire dataset (e.g., show genes that are unc and have associated movies). Each gene displayed has links through the gene name and location to WormBase's Gene Summary and mapping pages (Figures 3C and 4C). The strain name has a link to a comprehensive summary page containing all data relevant to that strain (Figure 3D). Along with the data present on the initial search readout page, other information included are the primers used to amplify the promoter, whether the strain is stabilized, and links to additional images of the strain.
A survey of temporal and spatial GFP expression patterns for all 1,886 genes is shown in Table 1 and Figure 5, and some illustrative examples in different tissues are displayed in Figure 6. We have detected GFP at all developmental stages and have identified expressed GFP in all major tissues except the germinal gonad. Most GFP fusions express across all developmental stages with 1,781 (95%) showing expression in adults, 1,835 (97%) in larval animals, and 1,556 (83%) expressing during embryogenesis. A majority of the 5′ regulatory DNA sequences examined drive GFP expression in the nervous system (63%), the intestine (63%), the pharynx (40%), and the body-wall muscle (32%) (Table 1). Subsets of cells and tissues within these broad categories are also delineated; we have observed GFP expression specific to the nerve ring, sensory neurons, ventral nerve cord, pharynx, seam cells, the excretory canal and excretory gland cells, the spermatheca, and coelomocytes, to list a few.
Over the course of our analysis, we observed GFP expression in 38 tissues and cell types throughout all developmental stages: embryo, larval (L1–L4) and adult (Figure 6; Table 1). We observed many examples of temporal expression stability and examples where the expression pattern changed during development. For example, pF26F4.6::GFP exhibited hypodermal expression during the larval stages, but no GFP was detectable in adult hypodermis; pY61A9LA.10::GFP showed intestinal and neural expression during early developmental stages, whereas adults lacked any GFP expression at all. Conversely, we observed cases where GFP expression was turned on later in development, as in the case of pF11F1.1::GFP, where no GFP was detected until the animals matured to adults, at which point hypodermal and intestinal expression were observed. Examples of changing patterns of expression formed a minority of our dataset. This, in some respects, was to be expected because we used the enhanced form of GFP (EGFP), which has a long half-life. Early expressed embryonic GFP could persist through the 14 h (22 °C) duration of embryogenesis [20] and possibly past hatching. Similarly, GFP expressed during larval development may persist in adult tissues. Also, embryonic expression was never detected earlier than the 50–100 cell stage of embryogenesis, possibly a consequence of our inability to detect maternal RNA contributions to the developing embryo [21] (see Discussion).
Although the concatameric arrays may have led to germline silencing in the gonad [21], they may also have contributed to increasing the sensitivity of detecting an expression signal in other tissues. As described in Materials and Methods, each array has several copies of the fusion GFP construct. Several of these GFP fusions can express simultaneously in a particular cell. As a test of the sensitivity of GFP fusions, we used them to see if we could detect expression from genes with low numbers of SAGE tags. Specifically, we were able to detect a GFP signal for 232 genes that only had a single tag in either the embryo or one of the following tissues: neurons, hypodermis, intestine, or muscle. In each case, GFP expression was detected in the tissue for which only a single SAGE tag had been recorded (Table S2). The SAGE data can be viewed at http://tock.bcgsc.ca/cgi-bin/sage170 and WormBase. The source of this material is from different developmental stages and different purified tissue and cell populations during early development ([1] and unpublished data).
To measure the reliability and accuracy of the reporter expression patterns described in this study, we took further advantage of the existing SAGE data for specific tissues. We have compared the intersects between our GFP expression patterns for muscle, gut, and the nervous system against SAGE data for stage-specific purified cells populations for each of these tissues. In each case, we can identify about 70% of GFP reporter genes in the corresponding SAGE library (e.g., genes for 71% of GFP reporters expressed in muscle are detected in the muscle SAGE library; unpublished data). Considering that the SAGE libraries are limited to embryonic tissue only and that half of the SAGE tags are present in single copies, we believe this is a reasonable validation of the GFP reporter expression patterns observed.
Of the 1,886 genes with analyzable expression, only one in five was found to be tissue specific and only a very few were found to be cell specific (Figure 5; Table 1). Cell-specific promoters were found in a few special cases, as in the excretory cell in which we identified six 5′ regulatory regions that drove expression in only this cell (Figure 5). In another example, we found four specific cases of 5′ sequences limiting expression to the head mesodermal cell (Figure 5). In this study, we did not find any examples where individual cells belonging to a larger tissue group such as body-wall muscle, or hypodermis, or the intestine expressed by themselves. Tissue-specific GFP expression accounted for 20% of our samples, and all major tissues in this organism are represented in our dataset (Figure 6; Table 1). Of the 414 tissue-specific regulatory regions identified, the majority are expressed exclusively either in neural (l55; Figure 7 displays several examples of the complexity of the nervous system) or intestinal tissue (136). Other tissues or cell groupings that exhibited exclusive expression include the pharynx (40), body-wall muscle (18), reproductive system (14), hypodermis (10), hypodermal seam cells (8), pharyngeal gland cells (4), and the arcade cells (2).
When we examine the remaining genes, we observe that 321 of these regulatory regions drive expression in only two tissues. In the majority of these examples (72%), one of the two tissues involved is neural. We detected no bias for specific combinations of tissues or specific exclusions (Figure 8). Co-expression in nerve and muscle (604 examples), nerve and intestine (698 examples), or intestine and muscle (532 examples) are all roughly equivalent, with relatively little contribution from hypodermal expression. Cell- and tissue-specific regulatory regions clearly account for a minority of our expression examples, because the majority of 5′ regulatory regions we have analyzed, 1,151 (61%), drive expression in several tissues. (This is reflected in the Venn diagram of Figure 8 in which 493 examples express in at least three of the tissues being examined). A portion of this last group may represent ubiquitous expression, but it is not always possible to conclude that every cell expresses GFP. Widespread expression in an animal can make it extremely difficult to detect expression in each cell. In these cases, mosaicism of expression, rather than a hindrance, can be helpful. Figure 9 illustrates how mosaic expression can be used to advantage to obtain images of structures within the somatic gonad (Figure 9A, 9B, 9C, 9E, and 9G) and individual cells of the gonad (Figure 9D, 9F, 9H, and 9I). All of the examples in this figure are for genes that show expression in many different cells and tissues (see database at http://gfpweb.aecom.yu.edu/index for details on each gene.)
A large source of regulatory sequences and expression data permits investigation of regulatory sequences required to drive expression in a specific cell or tissue type. We use muscle as an example of how this dataset can be employed. We first identified several 5′ sequences capable of driving expression of GFP in body-wall muscle. We next took a subset of these sequences (four) and mapped out the region responsible for muscle expression by constructing a deletion series (Table S3 lists primers used for this deletion series). These deletion constructs determined the minimal 5′ DNA sequence required to drive muscle expression. The four gene promoters analyzed in our study were those of F15G9.4a, C34E10.6, T04A8.4, and T27A1.4 (Figure 10). From the deletion series, we found that the minimal length required for muscle expression varied between the promoters, the longest being 326 bp (Figure 10B), whereas the shortest was only 143 bp (Figure 10D). When compared to each other, except for T27A1.5 which contains an E box consensus sequence, the minimal promoters were found to contain neither any shared motifs nor any of the previously identified muscle motifs [22,23] (unpublished data).
Within this database, there are representatives of many of the expression patterns that are possible in this organism. We have identified 5′ DNA sequences that drive expression in single cells, in single tissues, in multiple tissues, and in all tissues. The dataset is large enough so that one can make some general statements about patterns of expression in this organism. One conclusion from these data is that expression within only a single cell using extant 5′ sequences is rare. The examples that exist in our dataset are usually examples in which a single cell is equivalent to a tissue, as in the case of the excretory cell. However, tissue-specific 5′ regulatory regions are abundant. We found many examples of expression limited to a single tissue, and this included such tissues as the intestine, muscle, and the nervous system, the primary tissues arising from the three primordial germ layers, of endoderm, mesoderm, and ectoderm (Figure 8). There are also expression patterns that represent subsets of these tissues and expression patterns that are specific to organs or specialized groupings of cells within these broader tissue categories, for example, expression in the pharynx, but not other muscle, or expression in the amphids/phasmids, but not other cells of the nervous system. We also identified 5′ regulatory regions that are not limited to regulating expression in a single tissue, but may include two or more tissues and even cells from several tissues. We also identified several 5′ sequences that apparently permit ubiquitous expression (at least 1%). Finally, we observed 516 5′ regulatory regions that did not exhibit any detectable expression. Although there are several trivial explanations for why these regions do not promote expression, there is also the possibility that these are conditional promoters. Several laboratories have requested these strains to test for expression in different genetic (male vs. hermaphrodite) or environmental backgrounds. So far, none have been shown to be conditional promoters.
The paucity of 5′ regulatory regions that drive expression in a single cell is perhaps disappointing, but it should not be a surprising result. At least one quarter of the genome is expressed in any particular tissue or cell type ([1], unpublished data; http://tock.bcgsc.ca/cgi-bin/sage170) which, as we observed, suggests even tissue-specific control regions will be relatively infrequent. To identify regulatory regions that drive expression in only single cells in this organism may require other approaches. In our experience, single unique genes predominantly express in multiple cell types. One possible way to identify cell-specific control elements may be to focus attention on gene families or alternative splice forms of a single gene. The seven transmembrane domain and guanyl cyclase gene families of receptors are excellent examples of gene families in which isoforms are specific for separate sensory neurons [24,25]. In this study, we did not focus on gene families, but it may be the approach one should take if the objective is to identify cell-specific markers.
The database should not be viewed as the final arbiter of complete expression for any specific gene. As we have only included DNA 5′ of a particular ORF, we may not have the complete “promoter” or all possible “enhancer” elements that impinge on the regulation and expression of this gene when located at its proper location within the chromosome. Our analysis misses any downstream, intronic, or more than 3-kb upstream elements important for proper gene expression. Because of this, a gene's complete expression pattern may differ from that observed using our reporter constructs. As well, 85% of the strains we examined had concatamer arrays with multiple copies of the regulatory region of the gene. This led to mosaic expression when the concatamer was lost, which meant that we had to be sure to examine several animals to ensure that we described all expression patterns possible using this stretch of DNA as a control element. Stably inherited constructs were made for about 15% of the samples, including those from which we desired to make an embryo 4-D recording. Note that the aforementioned caveats are not unusual, as most single gene studies reported in most C. elegans publications work with the same limitations (see expression report summaries in WormBase). If one uses reproducibility as a benchmark, then the data reported here are quite reliable. First, we compared our GFP expression data to expression data using SAGE to detect tissue-specific transcripts and found that about 70% of the genes found expressed in a particular tissue by our GFP reporter assay were also detected using SAGE analysis. We also included in our analysis several genes whose expression was previously characterized, either by GFP promoter constructs or protein fusions or by antibodies. For more than 80% of the previously characterized genes we examined, the expression pattern is in good agreement with published observations. In some cases, we observed a wider range of expression, and in some cases, we observed less. In less than 10% of cases, our observations were completely at odds with what has previously been published. Due to the possible differences in size of 5′ promoter regions, differences in concatamer arrays, or even entirely different methodologies, these discrepancies should not be too surprising. In regard to this benchmark set of genes, often it is not clear whether our observations are the correct ones, or whether previous observations are correct, or if neither reflect the full range of expression of the gene in question. What we are certain of is that the annotation of tissue and cell identity is correct in our study. We have called upon experts within the C. elegans community and the staff of WormAtlas in every instance in which there was a question of cell identity. If cell identity could still not be resolved, this was indicated in the annotation. If there are errors, they are errors of omission, not errors of commission.
With almost 2,000 expression profiles, the database is an excellent resource for examining the expression profile of a previously uncharacterized gene, even with the caveats stated above. However, we do not feel this is the only possible use of the data. The data reflect expression from less than 5 Mb of DNA, less than 5% of the genome of this organism, and yet we see expression in almost every tissue and cell type in the organism. We think this is fertile ground for researchers interested in identifying motifs regulating gene expression. In many cases, the DNA segment regulating precise cellular and temporal expression is considerably shorter than our maximum size fragments of 3 kb. The ability to search this database for short DNA sequences controlling specific expression patterns should make it easier to identify transcription factor binding sites for a particular organ, tissue, or cell type.
Our survey of a few regions determining expression within muscle serves as a case study. We first identified several genes expressed within body-wall muscle. We then picked a subset of 5′ regions and did promoter deletions in order to map essential sites for muscle expression. Curiously, we did not find any single motif, but in fact, found several potential sequences that each could direct expression in muscle (unpublished data). The implication of these observations is that different 5′ sequences can lead to expression in the same tissue, in this case muscle, and we suspect this multiplicity of transcriptional control regions may occur in other tissues as well. This adds a level of complexity to gene regulation that many researchers fail to take into consideration. Our findings of multiple different sequences controlling muscle expression are similar to results reported previously [22,23], but the sequences we have identified are different from those reported in these earlier studies. Even though a MyoD homolog (hlh-1) [26–28] is expressed in C. elegans muscle, it does not seem to be the major transcription factor, because no MyoD binding site has been found in three of four control regions we analyzed. Recently, it has been shown that MyoD acts as part of a trio of transcription factors to regulate muscle differentiation in C. elegans [29].
Many of the genes in this expression database have human orthologs, and for a number of these genes, these expression data are the first indication of where these genes may be expressed in humans. We think this is an important resource to help direct studies of these genes in mammals. Considering the complexity of the mammalian nervous system, any gene that we can identify in a particular subset of neurons may be especially useful. Another use of the database has been to confirm an expression profile of a specific gene identified by other methods. Studies of adult intestine and ciliated neurons have used the GFP strains described in this database as confirmation of tissue-specific expression of genes identified by SAGE tags found in these tissues [30,31].
The GFP constructs described in this study are relatively easy to make and thus lend themselves to a high-throughput strategy. The PCR-stitching strategy we used [15] has proven robust and efficient. This approach has at least one advantage over the newly developed “promoterome” [12], which is that significantly larger 5′ regions can be used for stitching when necessary. Many regulatory regions are close to the ATG start site, as shown for the four genes we analyzed for muscle expression (Figure 10), but this is not always the case. A further complication with plasmids is that they often contain cryptic promoter elements, which one can avoid by using the PCR-stitching approach. The use of freely segregating concatamer arrays for this study had three implications. It appears from a comparison of GFP expression with low tag-frequency SAGE data that concatamer arrays of GFP may be a sensitive tool for detecting genes with a low level of transcription. We also demonstrated that mosaicism due to loss of the array often led to expression in small groups of cells or single cells, and thus allowed us to obtain a detailed image of these cells. This has been an invaluable aid to the WormAtlas project (http://www.wormatlas.org/). On the other hand, an unfortunate consequence of using a concatamer array was that it excluded us from recording germline expression and thus monitoring the maternal contribution to early development. Germline silencing of genes is well documented [21], and this silencing led to us not detecting germline expression in any of the genes we tested. It also meant that we could not detect expression in the early embryo (before 50 cells) in most cases.
In addition to the approaches described in this study, other approaches to monitor gene expression will be required if we are to monitor gene expression for the whole genome throughout all of development. The technique of homologous recombination in E. coli called recombineering [32–36] is a promising approach because it allows the modification and manipulation of large genomic clones. Larger DNA clones would remove some of the doubt about whether all control elements for transcription regulation are included. Recombineering in bacteria to construct GFP::protein fusions using fosmids with 35- to 40-kb DNA inserts should cover all control elements for most genes in C. elegans. We have built a C. elegans fosmid library, and clones from this library are being used for recombineering (http://elegans.bcgsc.bc.ca/perl/fosmid/CloneSearch) ([33] and unpublished data). If these GFP-engineered fosmids are introduced to the worm using a Biolistic gun [37,38], there is a higher probability of generating a transformed animal with a single or low copy number level of the gene. This should allow expression in the germline and the early embryo of any gene in which these are the normal sites of expression. Coupling these strategies to the newer methods of lineaging early cell division [39] should cover the stages in development overlooked in our study.
Our list of target genes was based on the 4,367 C. elegans proteins identified from a comparison of C. elegans and human predicted proteomes with InParanoid [16] (http://inparanoid.sbc.su.se), Most of the genome annotations used in the selection of our list of target genes were obtained from WormBase [40,41] (http://www.wormbase.org). The list was filtered to remove rRNA genes and genes with SL2 trans-splice acceptor sites, which are associated with operons [42,43]. Also removed were genes with characterized mRNAs, an indication that the gene was already well studied. Preference was given to genes with EST-confirmed 5′ ends and those identified as embryonically expressed in Intronerator [44]. We kept genes for which other researchers have constructed reporter fusions as a control set for our study. Our final set of targets consisted of a gene pool enriched for, although not exclusive to, human orthologs with unknown function.
The promoter::GFP fusion constructs were generated using the PCR stitching method from Hobert [15]. The PCR experiments were designed to capture putative 5′ DNA regions by amplifying about 3 kb of genomic DNA sequence immediately upstream of the predicted ATG initiator site. When an upstream gene was within 3 kb, the size of the amplicon was adjusted downward. We set the maximum primer length to be 25 nucleotides, and in order to eliminate false-positive PCR products, we designed a nested primer immediately downstream from the most 5′ primer for the second-round reaction. Where the primer encompassed the ATG initiator site, the G was mutated to a C, to ensure there was only one start codon in the promoter::GFP fusion.
Early PCR experiments were designed semimanually with the aid of primer3 [45]. To facilitate scale-up, we used Perl and AcePerl [46] to extract C. elegans genomic DNA sequence, and annotations from WormBase to tie them together with the primer design and validation programs primer3 and e-PCR [47]. An interactive version of the GFP primer design program is available at http://elegans.bcgsc.bc.ca/promoter_primers.
We used pPD95.67 variant S65C (developed by Dr. Andrew Fire, Carnegie Institution, http://www.addgene.org/pgvec1?f=c&cmd=showcol&colid=1) as our GFP source because it contains a GFP-cassette and a region that has sequence overlap with the 3′ primer, thus allowing for PCR stitching. 5′ DNA regions from target genes were amplified from C. elegans N2 (Bristol) genomic DNA.
DNA amplification mixtures consisted of Mix 1: 0.5-μl dNTP (10 mM), 1-μl N2 genomic DNA, 21.5-μl double-distilled H2O (ddH2O), 5′ and 3′ primers (1 μl of 12.5 μM each); and Mix 2: 0.75-μl Long Taq (Expand Long Template PCR System made by Roche Diagnostics, http://www.roche.com), 5-μl 10× Long PCR buffer (#2 from kit), 19.25-μl ddH2O. Mix 1 and Mix 2 were combined, and PCR was carried out for 30 cycles under the following conditions. Step 1: (1 cycle) 94 °C for 1 min. Step 2: (30 cycles) denaturation at 94 °C for 10 s, anneal at 56 °C for 30 s, and elongation at 68 °C for 2.5 min (depending on amplification fragment size). Step 3: 68 °C for 5.5 min. Stitched PCR product was constructed as follows: Mix 3: 5′ and 3′ primers (1 μl of 12.5 μM); 0.5-μl 5′ regulatory DNA PCR product, 0.5-μl GFP PCR product, 1.5-μl dNTP 10 mM, 21-μl ddH2O, and Mix 4: 5-μl 10× Long PCR buffer, 20-μl ddH20. Mix 3 and Mix 4 were combined. PCR was done as follows. Step 1: (1 cycle) 94 °C for 1 min. Step 2: (18 cycles) denaturation at 94 °C for 10 s, anneal at 56 °C for 30 s, and elongation at 68 °C for 2.5 min. Step 3: (10 cycles) 94 °C for 10 s, 56 °C for 30 s, and 68 °C for 2.5 min (increased by 10 s each cycle). The PCR product was stored at 4 °C.
Nematode strain maintenance and culture were carried out as described by Brenner [48]. Strains were maintained at 15 °C on OP50 plates unless otherwise specified. Strains used include dpy-5(e907) and wild-type N2 Bristol [48].
At the beginning of the project, we injected a number of strains, with-gel purified DNA, and came to a similar conclusion as Hobert [15], that gel purifying DNA for injection did not significantly change the results.
Transgenic worms were generated by a modification of the method described by Mello et al. [49]. 5′ regulatory DNA::GFP constructs and dpy-5(+) plasmid (pCeh-361) (kindly provided by C. Thacker and A. Rose; [50]) were used to construct transgenic strains. Transformants were identified by rescue of the dpy-5 mutant phenotype. The 5′ regulatory DNA::GFP fusions were co-injected with wild-type dpy-5 plasmid DNA into P0 Dpy-5(e907) gonads using one of these systems: a Olympus BH2-HLSK with a Leitz Westlab injection needle manipulator, or a Zeiss 47 3016 microscope (Carl Zeiss, http://www.zeiss.com) with a Leitz Westlab injection needle manipulator (http://www.leitz.org/leitz_english/index.html), or a MINJ-7 microinjection system with an Olympus CK40 microscope from Tritech Reseach (http://www.tritechresearch.com). Injection mixture included ddH2O, 10× TE, dpy-5 plasmid (pCeh361, concentration 5–80 ng/μl), and 5′ regulatory DNA::GFP fusion construct (concentration 50 ng/μl). A total of 1 nl of the final mix (80–90 ng/μl pCeh361 and 5–20 ng/μl DNA::GFP fusion) was microinjected into P0 worms using 1.0-mm, 6” filamented capillary tubes from World Precision Instruments (http://www.wpiinc.com) pulled on a Sutter P-97 needle puller. P0 worms were set up for microinjection on agarose pads (2%–3% agarose flattened on cover slips) in either mineral oil (Sigma) or in halocarbon oil #700 grade (Lab Scientific, http://www.labscientific.com). An injection set consists of 25–50 P0 worms injected with a given 5′ regulatory DNA::GFP construct. Wild-type F1s were set up individually and their progeny were screened for wild-type animals in the F2 generation. One or two lines yielding at least 30% wild-type progeny were maintained as transformed stocks. For promoter analysis, DNA was injected at 40–60 ng/μl for both subcloned constructs and PCR fusions, using rol-6 as an injection marker.
To determine the size of the concatemeric arrays in vivo, we used quantitative PCR to estimate the copy number of the 5′ DNA::GFP constructs and plasmids in 20 different transgenic strains. We estimated that there were about 5–10 copies of promotor::GFP and 100–600 copies of the dpy-5 plasmid in the heritable arrays, which was sufficient for the sensitivity of our GFP assay.
We constructed chromosomal integrant strains for a subset of the GFP constructs (1%) using a modified version of M. Koelle's method (http://info.med.yale.edu/mbb/koelle/). Young adult transgenic (wild-type) P0 hermaphrodites were treated with low-dose X-ray irradiation (1,500 R). After 1 h, the P0 animals were transferred to 90-mm OP50 plates—one P0 worm/plate for 12 plates for each strain. The P0 animals were allowed to lay eggs for 18–24 h and then were removed in order to limit the number of F1s laid. Seven days later, mid to late larval wild-type F2 animals were picked and set up (one/plate, 12 from each of the 90-mm plates) at room temperature (20–22 °C). Four to 5 d later, the F2 plates were screened for the absence of Dpy-5 animals, indicating stable inheritance of the array.
Strains intended for embryo recordings were outcrossed using an unc-32 marker. P0 GFP-expressing hermaphrodites were crossed with N2 males, F1 GFP males were crossed with unc-32 hermaphrodites, and then F2 and F3 GFP hermaphrodites were individually plated. Lastly, the F4 populations were screened for exclusively wild-type animals. Outcrossing was done at 15 °C.
General classification and imaging of GFP expression was done initially with a low-power GFP dissecting microscope (Zeiss stereomicroscope fitted with Kramer epifluorescence), before moving to either a Zeiss Axioplan or a Zeiss Axiophot microscope. Images were captured using a digital camera (QICAM; QImaging, http://www.qimaging.com/products/cameras/scientific/) and QCapture software. This was the first pass, where we determined the developmental stage, tissues, and, where possible, the individual cells expressing GFP. Both stable and unstable strains were evaluated on expression pattern complexity and frequency of occurrence. Unusual or complicated expression patterns, or neural expression, would undergo further analysis using an inverted Zeiss Axiovert LSM 5 confocal microscope equipped with epifluorescence, Nomarski optics, and LSM 5 Pascal software. If we detected pre-comma nonubiquitous expression, strains were put in queue for stabilization and/or outcrossing, and 4-D recording and analysis. The results of all analyses, excepting of the embryos, were curated by hand and uploaded to the project Web site (http://gfpweb.aecom.yu.edu/index) and WormBase (http://www.wormbase.org), and the strains were sent to the stock centre (http://www.cbs.umn.edu/CGC/CGChomepage.htm) and are available by request from R. Johnsen ([email protected]) (see expression pipeline in Figure 1).
The images and movies were processed using Adobe PhotoShop 7.0 (http://www.adobe.com) and the LSM 5 Pascal volume-rendering software. Single images were normalized and placed into image panels before exporting to the public domain. Movies were obtained from Z-stacks comprised of 20–60 *.lsm images, taken 0.5–1-μm intervals apart, the specifics of which were dependant on the age of the worm, the tissue of interest, and the intensity of the GFP. These stacks were then optionally volume-rendered and/or converted into QuickTime movies, normalized, and exported to the Web site.
In some cases, embryonic expression was determined without difficulty. However, in many cases, the patterns were determined to be too complex, and it was deemed necessary to have a 4-D recording and to lineage the embryo. Embryos for 4-D analysis were obtained from gravid hermaphrodites. Two embryos at the 2–4 cell stage, were transferred to a 5% agar pad and manipulated into adjacent positions with the same orientation. Detailed expression patterns and gene activation in the embryos were captured with live, two-channel, four-dimensional microscopy, on a Zeiss Axioplan microscope. The fourth dimension being time, Z-stacks (25 Z-images) of developing embryos were recorded at 25 °C using Nomarski microscopy every 30–45 s over a 7-h time course. Interspersed with the normal Z-stacks, we recorded several Z-stacks of GFP fluorescence in specific cells, which were then mapped and identified relative to the Nomarski images. Software that supports this type of microscopy recording and analysis has been developed [19,51–53]. We used programs derived from the study by Schnabel et al. [19] and the program Simi Biocell to lineage the embryos [19]. The data have not been posted to the Web site, but are available from the authors.
All strain data are in a mySQL database. All primer designs relative to genes and all annotation of genes on the Web site are based on WormBase version 140. The functionality of the Web site is based on perl/CGI and perl modules for the queries, which provide the user with three formats for accessing the data: (1) the display of all strains and data for browsing, (2) the selection of specific genes and the information the user wants to see for each gene, and (3) a gene search, based on tissue expression pattern. All of the data can be downloaded in .tab or .csv format from WormAtlas (http://gfpweb.aecom.yu.edu/index). The data are also available at WormBase (http://www.wormbase.org).
|
10.1371/journal.ppat.0030190 | A Cellular Basis for Wolbachia Recruitment to the Host Germline | Wolbachia are among the most widespread intracellular bacteria, carried by thousands of metazoan species. The success of Wolbachia is due to efficient vertical transmission by the host maternal germline. Some Wolbachia strains concentrate at the posterior of host oocytes, which promotes Wolbachia incorporation into posterior germ cells during embryogenesis. The molecular basis for this localization strategy is unknown. Here we report that the wMel Wolbachia strain relies upon a two-step mechanism for its posterior localization in oogenesis. The microtubule motor protein kinesin-1 transports wMel toward the oocyte posterior, then pole plasm mediates wMel anchorage to the posterior cortex. Trans-infection tests demonstrate that factors intrinsic to Wolbachia are responsible for directing posterior Wolbachia localization in oogenesis. These findings indicate that Wolbachia can direct the cellular machintery of host oocytes to promote germline-based bacterial transmission. This study also suggests parallels between Wolbachia localization mechanisms and those used by other intracellular pathogens.
| This study focuses on Wolbachia, a genus of intracellular bacteria carried by insect and nematode host species. It was recently shown that Wolbachia carried into the human body by the host nematode Onchocerca volvulus trigger an immune response that leads to African river blindness. Findings like these raise fundamental questions of how Wolbachia interact with host cells to perpetuate Wolbachia infection. Distinct from many pathogenic bacteria, Wolbachia are transmitted throughout host populations primarily from females to their offspring, similar to mitochondrial inheritance. The molecular basis for this transmission strategy is unclear. Here we show that Wolbachia transmission is aided by a complex mechanism in egg development. Our study suggests that Wolbachia are transported inside the egg as cargo of molecular motors that walk along microtubule filaments. This directs Wolbachia to the posterior of maturing eggs, thus placing Wolbachia at the site where reproductive cells form during embryogenesis and ensuring Wolbachia integration into those cells. Furthermore, both factors intrinsic to Wolbachia and host molecules specifying reproductive cell fates are necessary to maximize posterior concentration of Wolbachia in the egg. This suggests that Wolbachia manipulate conserved cellular machinery in egg development to direct their transmission to the next host generation.
| Wolbachia are among the most widespread intracellular bacteria, carried by an estimated 15%–76% of insect species as well as by some crustaceans, mites, and filarial nematodes [1,2]. Wolbachia are closely related to the Rickettsia family, a collection of tick-borne pathogens known for causing typhus and spotted fevers in humans. Wolbachia are also linked to human disease via a symbiotic relationship with pathogenic nematodes [3]. For example, the Wolbachia-bearing nematode Onchocerca volvulus is linked to the condition African river blindness in humans. Of the 18 million people infected by O. volvulus, nearly one million are visually impaired or already blind [4]. Recent work has implicated Wolbachia directly as the cause of ocular inflammation leading to river blindness [5].
The effect of Wolbachia infection on its host is as varied as the hosts are themselves. Wolbachia act as endosymbionts of some host organisms, such as the filarial nematode O. volvulus and the wasp Asobara tabida, which require Wolbachia in order to complete oogenesis properly [3,6]. Wolbachia appear to cause little phenotypic impact in certain hosts, such as in Drosophila melanogaster. In other cases, Wolbachia manipulate the host to their advantage. Wolbachia bias host reproduction to favor infected females by inducing phenotypes such as male-killing, feminization, sperm–egg cytoplasmic incompatibility, and parthenogenesis (virgin birth) [1,2]. This is thought to promote the spread of Wolbachia throughout host populations.
Infectious agents often spread to new hosts by becoming inhaled or ingested by that host. In the case of Wolbachia, however, bacterial transmission occurs within the host maternal germline [1,2]. Though Wolbachia are present in both male and female germlines, the bacteria are removed from sperm cysts at the end of spermatogenesis [7,8], creating a reliance upon maternal transmission. In arthropods, this maternal transmission is accomplished via incorporation of Wolbachia into germline precursor cells, also known as “pole cells” [9–11]. This ensures that infected females resulting from those embryos will carry bacteria in their germlines as well, thus perpetuating the Wolbachia transmission cycle. Wolbachia transmission rates have been reported at over 97% for wild-caught D. melanogaster flies, and at 100% for laboratory-reared D. melanogaster and D. simulans flies [12,13], suggesting that the pole cell–based transmission strategy is highly efficient.
How might Wolbachia ensure their incorporation into host pole cells? Many Wolbachia strains have been reported to concentrate at the posterior of mature oocytes [1,9–11,14–17]. Interestingly, the oocyte posterior pole corresponds to the location where pole cell formation takes place later in embryogenesis. For this reason, the posterior concentration of Wolbachia during oogenesis is thought to promote Wolbachia incorporation into the embryonic germline [9–11]. The cellular and molecular basis underlying this posterior Wolbachia localization in oogenesis is unknown to date, however.
A recent study indicated that Wolbachia can associate with host cell microtubules in D. melanogaster oocytes [18]. These oocytes contain an extensive network of microtubules that serves as a scaffold for cargo transport by motor proteins [19]. Up to stage 6 of oogenesis, microtubule minus ends are generally concentrated at the oocyte posterior with plus ends toward the anterior [20–22]. At stage 7, microtubules reorient such that minus ends are concentrated at the antero-lateral cortex of the oocyte, and plus ends are biased toward the posterior [23–27]. Work from D. melanogaster demonstrated that the wMel Wolbachia strain exhibits a microtubule-dependent concentration at the oocyte anterior from oogenesis stages 3 to 6 [18]. This anterior wMel localization requires the minus end–directed motor cytoplasmic dynein and the associated motor regulatory complex dynactin. However, the plus end–directed motor kinesin-1 is not required for anterior wMel localization [18]. These results suggest that interactions between Wolbachia and specific microtubule motors can direct the subcellular distribution of Wolbachia in oogenesis. This raises the possibility that posterior Wolbachia localization in late-stage oocytes may also rely upon interactions between bacteria, microtubules, and microtubule motor proteins. This also highlights Wolbachia as a means of understanding bacterial manipulation of host microtubules, an interaction that is considerably less well-studied than bacterial exploitation of host actin, such as in engulfment of Salmonella or intracellular propulsion of Rickettsia, Listeria, and Shigella [28,29].
How else might Wolbachia take advantage of the host cell to promote their posterior localization? It is possible that Wolbachia manipulate oocyte patterning events to their advantage. In Drosophila, the body axes are established via asymmetrical localization of determinant mRNAs in the oocyte [30,31]. For example, the posterior/germline determinant oskar (osk) mRNA concentrates at the oocyte posterior pole. The current model is that from stages 8 to 10A of oogenesis, kinesin-1 transports osk mRNA and associated Staufen (Stau) protein along microtubules toward the posterior cortex, where osk is translated [23–27]. Osk then initates recruitment of numerous mRNAs, proteins, mitochondria, and ribosomes to the oocyte posterior [32]. This multicomponent posterior assembly is referred to as “pole plasm”, and it functions in embryogenesis to specify posterior pole cell fates. Pole plasm is needed for posterior wMel localization in embryos [9]. Perhaps Wolbachia require posteriorly enriched substrates such as osk-induced pole plasm to establish their posterior localization in oogenesis as well.
This study addresses how Wolbachia posterior localization is achieved by examining the roles of microtubules, motor proteins, pole plasm assembly, and Wolbachia. Our findings indicate that during mid- to late oogenesis, kinesin-1 transports wMel Wolbachia toward the posterior cortex where pole plasm components mediate posterior wMel anchorage. The functions of kinesin-1 and pole plasm contribute independently to posterior Wolbachia localization. Furthermore, wMel can direct its localization to the oocyte posterior pole, unlike the homogeneously distributed wRi Wolbachia strain carried by D. simulans. This distinction between posteriorly concentrating and evenly dispersed Wolbachia strains may be due to different abilities of those strains to interact with posterior pole plasm.
To understand the basis for wMel incorporation into embryonic pole cells, ovaries were stained with propidium iodide. This showed wMel to be anteriorly concentrated in stage 3–6 oocytes (Figure 1A and 1B) and homogeneously distributed in stage 7–9 oocytes (Figure 1E, 1E', 1F, and 1F') [18]. From late stage 9 to stage 12, a subset of wMel bacteria concentrated at the oocyte posterior cortex (Figure 1I, 1I', 1J, and 1J'; Table 1) [10]. wMel posterior localization persisted through early embryogenesis, facilitating wMel incorporation into the pole cells (Figure S1) [9–11]. Thus, concentration of wMel at the posterior of late stage oocytes promotes germline-based transmission of wMel.
The redistribution of wMel from the oocyte anterior to posterior suggests that an active localization mechanism is involved. To test a role for microtubule-based transport in posterior wMel localization, oocytes were treated with colcemid and colchicine. Some colcemid-treated oocytes exhibited wMel at both the lateral and posterior cortex (n = 7 of 15 cases; Figure 2A and 2A'), while others displayed a non-cortical, homogeneous distribution of wMel throughout the cytoplasm (n = 8 of 15 cases; Figure 2B and 2B'). Colchicine-treated oocytes displayed similar broad cortical or homogeneous wMel localization (n = 13 of 20 and n = 5 of 20 cases, respectively). This differed from control oocytes that mainly exhibited posterior wMel localization (19 of 22 cases; Figure 2C and 2C'). These data indicate that microtubules are required for focused posterior localization of wMel.
A role for microtubules in wMel localization implies that a posteriorly directed microtubule motor such as kinesin-1 is involved. To determine if kinesin-1 participates in wMel posterior localization, we created germlines mutant for the Kinesin heavy chain (Khc) gene [23,27,33,34]. Khc27 oocytes, null for kinesin function, showed normal anterior wMel localization during early stages (Figure S2). However, stage 10A Khc27 oocytes exhibited abnormal wMel distribution, with wMel absent from the posterior cortex in 83% of oocytes (Figure 2D, 2D', 2F, and 2F'; Table 1). wMel was also strikingly depleted from the posterior half of Khc27 oocytes (Figure 2D and 2F). Thus, kinesin-1 is important to both localize wMel to the posterior cortex and redistribute wMel into the posterior region.
The role for kinesin-1 in wMel posterior localization may reflect a direct or indirect Wolbachia localization mechanism. One possibility is that kinesin-1 transports wMel to the posterior as a cargo. However, kinesin-1 also drives bulk cytoplasmic streaming during mid- to late oogenesis [27,35,36]. Perhaps streaming currents sweep wMel passively toward the posterior cortex. To test a requirement for streaming in wMel localization, we examined oocytes carrying the hypomorphic mutations Khc17 and Khc23. These alleles give rise to streaming-capable and streaming-deficient oocytes, respectively [27]. Posterior Wolbachia were exhibited by 70% of Khc17 mutant oocytes and 62% of Khc23 mutant oocytes (Figure 2E; Table 1). The similarity of posterior wMel localization in these Khc mutants suggests streaming is not needed for posterior Wolbachia localization. Rather, as both Khc17 and Khc23 oocytes retain some kinesin-1 function [27,37], these results indicate that wMel is transported toward the posterior as a cargo of kinesin-1.
A dependency of wMel on kinesin-1 for its posterior localization in oogenesis suggests wMel may rely on the kinesin-1 cargoes osk mRNA and Stau as well. Perhaps wMel hitchhikes to the oocyte posterior as a passenger on osk/Stau messenger ribonucleoprotein particles (mRNPs). Alternatively, wMel may require osk-induced pole plasm for efficient anchorage to the oocyte posterior cortex. To test these possibilities, osk and stau were disrupted with maternal-effect mutations. The majority of these mutant oocytes exhibited depletion or absence of wMel from the posterior cortex compared to wild-type (Figure 2G–2I, 2G'–2I'; Table 1), indicating that osk and stau gene products are important for efficient posterior wMel localization. Furthermore, osk and stau mutant oocytes lacking posteriorly concentrated wMel still exhibited a homogeneous bacterial distribution throughout the cytoplasm, differing sharply from the anterior wMel concentrations seen in Khc27 oocytes (compare Figure 2D to 2G). This suggests that kinesin-1 can transport wMel into the posterior half of the oocyte independently of osk/Stau mRNPs. However, kinesin-1 is insufficient to drive robust wMel concentration at the posterior cortex in oocytes with disrupted pole plasm (Figure 2G and 2G'; Table 1). This suggests that pole plasm is important for posterior wMel anchorage.
To test whether pole plasm is sufficient to drive wMel localization, we examined wMel in oocytes with anteriorly localized pole plasm. To this end, an osk-bicoid 3'UTR transgene was used to target osk mRNA to the oocyte anterior margin [38]. This ectopically localized osk is translated and assembles functional pole plasm at the antero-lateral cortex [38]. wMel co-localized with wild-type Osk protein at the oocyte posterior cortex (Figure 3A–3C, 3A'–3C'). However, wMel did not concentrate at the anterior margin with ectopically localized Osk in osk-bicoid 3'UTR oocytes (Figure 3D–3F, 3D'–3F'), suggesting that pole plasm alone is insufficient to recruit wMel from the cytoplasm. This result, taken together with those above, suggests that individual functions of kinesin-1 and pole plasm are both needed for robust posterior wMel localization in late stages 9 and 10A. This is consistent with a two-step mechanism for wMel localization: kinesin-1-mediated transport of wMel toward the oocyte posterior, followed by pole plasm-mediated anchorage of wMel to the posterior cortex (Figure 4).
The extensive requirement of host components for posterior wMel concentration raises questions about whether wMel contributes to its localization. To investigate this, a trans-infection approach was employed using the host species, D. simulans, that normally carries the wRi Wolbachia strain [39]. In D. simulans oogenesis, wRi exhibited an anterior concentration during stages 3–6 and homogeneous distribution throughout the rest of oogenesis (Figure 1C, 1G, 1G', 1K, and 1K'; Table 1) [18]. Is this lack of posterior concentration due to differences between host oogenesis machinery or between the wRi and wMel strains? To address this, we examined D. simulans oocytes ectopically transformed with wMel [40]. wMel-infected D. simulans oocytes exhibited anterior Wolbachia concentration during early stages, homogeneous distribution in middle stages, and a striking posterior localization in late stages (Figure 1D, 1H, 1H' 1L, and 1L'; Table 1). This demonstrates that host components required for Wolbachia posterior localization are present in both D. melanogaster and D. simulans oocytes. Due to strain-specific differences, however, wMel engages those host components to enhance its posterior concentration in late oogenesis, whereas wRi does not.
Which oocyte components are engaged by wMel but not by wRi? Comparing wMel in osk mutant oocytes to wRi localization in D. simulans reveals a similar homogeneous distribution (Figures 1K and 2G). A speculative interpretation of this similarity is that wMel and wRi are similarly transported into the posterior half of the oocyte by kinesin-1. A further possibility is that wRi is unable to interact with host pole plasm, unlike wMel, which requires pole plasm for efficient posterior localization (Figure 2G and 2G'; Table 1). Perhaps unlike wRi, factors intrinsic to wMel drive interactions with posterior pole plasm that facilitate posterior Wolbachia anchorage (Figure 4).
The involvement of kinesin (this study) and dynein [18] in Wolbachia localization during oogenesis is reminiscent of microtubule-based transport employed by a number of human pathogens. Viruses such as herpes simplex virus type 1 rely on dynein and dynactin for their transport to a perinuclear position referred to as their “replication site” [41]. Kinesin transports the viruses back to the cell periphery, enabling their exit from the cell. Bacteria such as Salmonella are transported toward the host cell nucleus in a dynein/dynactin-dependent manner, which then facilitates bacterial replication [41]. Salmonella also actively recruits kinesin-1 to its surrounding membrane [42]. These observations suggest some parallels with wMel, which requires dynein and dynactin for anterior localization during early oogenesis [18] and kinesin-1 for posterior localization in late oogenesis. While the function of Wolbachia anterior localization is unclear, Wolbachia titer increases substantially at that location, suggesting that dynein-driven localization creates a replication site for Wolbachia within the oocyte [18]. Once replicated, kinesin-1-based transport enables Wolbachia to traverse the entire length of the growing oocyte, promoting Wolbachia incorporation into posterior pole cells. Wolbachia may therefore have sophisticated interactions with host motor proteins analogous to those used by other bacteria and viruses. The basis for a switch between dynein- and kinesin-1-dependent Wolbachia localization is currently unknown. In some systems the dynactin complex coordinates alternation of kinesin- and dynein-driven organelle motility [43]. Perhaps a regulatory agent like dynactin directs the changing Wolbachia localization pattern in oogenesis.
Upon reaching the posterior pole, wMel becomes anchored in a pole plasm–mediated manner. How might this occur? The simplest interpretation is that wMel associates directly with pole plasm components. However, a minority of osk null oocytes exhibited weak posterior Wolbachia localization (Table 1), although pole plasm is absent in this mutant background [38]. This suggests that other factors in addition to pole plasm assist posterior Wolbachia anchorage. Perhaps wMel has a dual affinity for pole plasm and an as-yet-unidentified posterior anchor. In such a case, the combined presence of those substrates may be important for robust Wolbachia anchorage to the posterior cortex. Alternatively, pole plasm may indirectly promote Wolbachia localization by stabilizing Wolbachia anchorage sites. A recent report indicated that Osk regulates actin polymerization at the oocyte posterior cortex [44]. It may be that wMel has a high affinity for unknown factors that associate with the posterior actin cortex, creating an indirect dependency of wMel upon posterior Osk.
One apparent conflict with these selective anchorage hypotheses is the finding that some colcemid- and colchicine-treated oocytes exhibit Wolbachia in association with the lateral cortex of the oocyte (Figure 2A). One interpretation of this result is that Wolbachia may have a general affinity for cortical actin independent of pole plasm. In such a scenario, one would predict that kinesin must normally drive wMel away from the lateral cortex and restrict it to the oocyte posterior where wMel is permitted to bind actin. This type of model has previously been proposed in the context of osk mRNA localization to the posterior pole [24,27]. If this prediction is accurate for wMel also, then oocytes lacking kinesin function should exhibit wMel localization to the antero-lateral cortex. However, wMel did not concentrate on the cortex of Khc null oocytes (Figure 2D). This suggests wMel does not have a general affinity for the actin cortex analagous to osk mRNA. An alternative interpretation of cortical wMel localization in colchicine- and colcemid-treated oocytes is that the drug treatments permitted microtubule remnants to remain along the cortex of some oocytes [21]. Those microtubule remnants could serve as a substrate for short-range wMel transport by kinesin-1, giving rise to a cortical wMel localization pattern. This possibility is consistent with the other findings of this study that favor kinesin-based wMel transport to the oocyte posterior, followed by selective wMel anchorage at the posterior pole.
The study presented here is one of the few to examine host–pathogen interactions in a developmental context. What emerges from this analysis is that the Wolbachia localization pattern is unique and does not follow specific morphogens or organelles during oogenesis. The Wolbachia localization pattern is distinct from mitochondria, which are concentrated on the posterior side of the oocyte nucleus during early stages, homogeneously distributed during mid-oogenesis, and posteriorly concentrated in stages 9 and 10 [45]. The anterior localization of Wolbachia precedes that of the determinant bicoid mRNA, which concentrates anteriorly from stages 6 to 14 of oogenesis [46]. Wolbachia posterior localization also appears later than osk mRNA, which concentrates posteriorly from stages 3 to 6, anteriorly in stage 8, and posteriorly again from stages 8 to 10 of oogenesis [47,48]. Furthermore, our study indicates that Wolbachia do not localize to the posterior cortex in association with osk/Stau mRNPs. Taken together, these observations suggest that the demands of replication and localization are unique to Wolbachia and may preclude these bacteria from hitchhiking on morphogens or organelles.
The posterior localization strategy described in our report is exhibited by Wolbachia strains carried within multiple Drosophila and Hymenopteran species [1,9–11,14–17]. This recurrent localization pattern may reflect bacterial adaptations to the host environmental conditions. D. simulans allows wRi to persist at a high titer during embryogenesis, which is sufficient to promote wRi incorporation into posterior pole cells [10]. This environment may provide little incentive for wRi to evolve or retain a posterior localization strategy. The wMel strain, by contrast, is maintained at lower concentrations in D. melanogaster embryos [10]. This may pressure wMel to evolve and/or retain mechanisms that drive its posterior localization in oogenesis, thus enhancing its incorporation into embryonic pole cells. Taking advantage of kinesin-1 and pole plasm assembly at the oocyte posterior, as demonstrated by this study, provides an excellent means by which Wolbachia can accomplish this goal.
wMel Wolbachia were crossed into wild-type D. melanogaster flies carrying the markers and balancers w; Sp/Cyo, Sb/Tm6Hu. This infected stock was used to cross wMel into all the D. melanogaster mutants used for this study, ensuring that all carried wMel strains of a comparable genetic background.
Ovaries were dissected and fixed using standard methods [23], then stained and imaged as previously [18]. Rabbit anti-Osk antibodies were used at 1:3000 [49]. Embryos were dechorionated with 50% bleach, fixed 20 min in a 1:1 mixture of 3.7% formaldehyde and heptane, and devitellinized by vigorous agitation in methanol. Embryos were stained with rabbit anti-Vasa at 1:2000 [50] and mouse anti-Hsp60 (Sigma) at 1:100 [18] in PBS/0.1% Triton, followed by 1:500 dilutions of Alexa-488- and Alexa-594-conjugated secondary antibodies (Molecular Probes).
Flies were starved 18 h, then fed 24–48 h with yeast paste containing 50 μM colcemid, 50 μM colchicine in DMSO, or comparable dilutions of DMSO alone. Mispositioning of the oocyte nucleus served as an internal control to verify that microtubule disruption had occurred [21,51].
Images were acquired on a Leica DM IRB confocal microscope using a 63× oil objective and zoom factor of 1.5. Each oocyte was imaged as a z-series stack of 7–14 images spaced at 1.5-μm intervals. Optical sections deeper than 4.5 μm into the oocyte were examined for the presence of posterior Wolbachia. Oocytes were categorized in Table 1 as showing strong posterior localization if they exhibited striking Wolbachia staining, which consisted of either an intense linear array of Wolbachia puncta or a crescent-shaped area saturated with Wolbachia staining along the posterior cortex for four out of five consecutive z-sections. Oocytes were designated as showing weak posterior localization if they exhibited a.) at least one z-section with striking posterior localization, or b.) at least two z-sections with a higher Wolbachia density along the posterior cortex than in the cytoplasm of the cell. Oocytes were categorized as showing no posterior localization if they did not meet the above conditions. Wolbachia density was not analyzed in this study because oocytes carrying high bacterial loads exhibited saturation of Wolbachia labeling at the posterior pole that disrupted bacterial quantitation.
The NCBI Entrez (http://www.ncbi.nlm.nih.gov/sites/entrez?db=gene) accession numbers for the genes and gene products discussed in this paper are Hsp60 (P10809), Kinesin heavy chain (P17210), Oskar (P25158), Staufen (P25159), and Vasa (P09052). |
10.1371/journal.pcbi.1005786 | Machine learning to design integral membrane channelrhodopsins for efficient eukaryotic expression and plasma membrane localization | There is growing interest in studying and engineering integral membrane proteins (MPs) that play key roles in sensing and regulating cellular response to diverse external signals. A MP must be expressed, correctly inserted and folded in a lipid bilayer, and trafficked to the proper cellular location in order to function. The sequence and structural determinants of these processes are complex and highly constrained. Here we describe a predictive, machine-learning approach that captures this complexity to facilitate successful MP engineering and design. Machine learning on carefully-chosen training sequences made by structure-guided SCHEMA recombination has enabled us to accurately predict the rare sequences in a diverse library of channelrhodopsins (ChRs) that express and localize to the plasma membrane of mammalian cells. These light-gated channel proteins of microbial origin are of interest for neuroscience applications, where expression and localization to the plasma membrane is a prerequisite for function. We trained Gaussian process (GP) classification and regression models with expression and localization data from 218 ChR chimeras chosen from a 118,098-variant library designed by SCHEMA recombination of three parent ChRs. We use these GP models to identify ChRs that express and localize well and show that our models can elucidate sequence and structure elements important for these processes. We also used the predictive models to convert a naturally occurring ChR incapable of mammalian localization into one that localizes well.
| A protein’s amino acid sequence determines how it will fold, traffic to subcellular locations, and carry out specific functions within the cell. Understanding this process would enable the design of protein sequences capable of useful functions; unfortunately, we cannot predict in detail how sequence encodes function. However, machine-learning models have the potential to infer the complex protein sequence-function relationship by identifying patterns or features that are important for function from sequences with known functions. We used machine learning to learn about and design membrane proteins (MPs). To function, a MP must be expressed, correctly folded in a lipid membrane, and trafficked to the proper cellular location. We built predictive, machine-learning models for this complex process from a set of >200 chimeric MPs and used them to design new sequences with optimal performance on the challenging task of membrane localization. This general approach to understanding and designing MPs could be broadly useful for important pharmaceutical and engineering MP targets.
| As crucial components of regulatory and transport pathways, integral membrane proteins (MPs) are important pharmaceutical and engineering targets [1]. To be functional, MPs must be expressed and localized through a series of elaborate sub-cellular processes that include co-translational insertion, rigorous quality control, and multi-step trafficking to arrive at the correct topology in the correct sub-cellular location [2–4]. With such a complex mechanism for production, it is not surprising that MP engineering has been hampered by poor expression, stability, and localization in heterologous systems [5–7]. To overcome these limitations, protein engineers need a tool to predict how changes in sequence affect MP expression and localization. An accurate predictor would enable us to design and produce MP variants that express and localize correctly, a necessary first step in engineering MP function. A useful predictor would be sensitive to subtle changes in sequence that can lead to drastic changes in expression and localization. Our goal here was to develop data-driven models that predict the likelihood of a MP’s expression and plasma membrane localization using the amino acid sequence as the primary input.
For this study, we focus on channelrhodopsins (ChRs), light-gated ion channels that assume a seven transmembrane helix topology with a light-sensitive retinal chromophore bound in an internal pocket. This scaffold is conserved in both microbial rhodopsins (light-driven ion pumps, channels, and light sensors–type I rhodopsins) and animal rhodopsins (light-sensing G-protein coupled receptors–type II rhodopsins) [8]. Found in photosynthetic algae, ChRs function as light sensors in phototaxic and photophobic responses [9,10]. On photon absorption, ChRs undergo a multi-step photo-cycle that allows a flux of ions across the membrane and down the electrochemical gradient [11]. When ChRs are expressed transgenically in neurons, their light-dependent activity can stimulate action potentials, allowing cell-specific control over neuronal activity [12,13] and extensive applications in neuroscience [14]. The functional limitations of available ChRs have spurred efforts to engineer or discover novel ChRs [11]. The utility of a ChR, however, depends on its ability to express and localize to the plasma membrane in eukaryotic cells of interest, and changes to the amino acid sequence frequently abrogate localization [5]. A predictor for ChRs that express and localize would be of great value as a pre-screen for function.
The sequence and structural determinants for membrane localization have been a subject of much scientific investigation [15–17] and have provided some understanding of the MP sequence elements important for localization, such as signal peptide sequence, positive charge at the membrane–cytoplasm interface (the “positive-inside” rule [18]), and increased hydrophobicity in the transmembrane domains. However, these rules are of limited use to a protein engineer: there are too many amino acid sequences that follow these rules but still fail to localize to the plasma membrane (see Results). MP sequence changes that influence expression and localization are highly context-dependent: what eliminates localization in one sequence context has no effect in another, and subtle amino acid changes can have dramatic effects [5,16,19]. In short, sequence determinants of expression and localization are not captured by simple rules.
Accurate atomistic physics-based models relating a sequence to its level of expression and plasma membrane localization currently do not exist, in large measure due to the complexity of the process. Statistical models offer a powerful alternative. Statistical models are useful for predicting the outcomes of complex processes because they do not require prior knowledge of the specific biological mechanisms involved. That being said, statistical models can also be constructed to exploit prior knowledge, such as MP structural information. Statistical models can be trained using empirical data (in this case expression or localization values) collected from known sequences. During training, the model infers relationships between input (sequence) and output (expression or localization) that are then used to predict the properties of unmeasured sequence variants. The process of using empirical data to train and select statistical models is referred to as machine learning.
Machine learning has been applied to predicting various protein properties, including solubility [20,21], trafficking to the periplasm [22], crystallization propensity [23], and function [24]. Generally, these models are trained using large data sets composed of literature data from varied sources with little to no standardization of the experimental conditions, and trained using many protein classes (i.e. proteins with various folds and functions), because their aim is to identify sequence elements across all proteins that contribute to the property of interest. This generalist approach, however, is not useful for identifying subtle sequence features (i.e. amino acids or amino acid interactions) that condition expression and localization for a specific class of related sequences, the ChRs in this case. We focused our model building on ChRs, with training data collected from a range of ChR sequences under standardized conditions. We applied Gaussian process (GP) classification and regression [25] to build models that predict ChR expression and localization directly from these data.
In our previous work, GP models successfully predicted thermal stability, substrate binding affinity, and kinetics for several soluble enzymes [26]. Here, we asked whether GP modeling could accurately predict mammalian expression and localization for heterologous integral membrane ChRs and how much experimental data would be required. For a statistical model to make accurate predictions on a wide range of ChR sequences, it must be trained with a diverse set of ChR sequences [25]. We chose to generate a training set using chimeras produced by SCHEMA recombination, which was previously demonstrated to be useful for producing large sets (libraries) of diverse, functional chimeric sequences from homologous parent proteins [27]. We synthesized and measured expression and localization for only a small subset (0.18%) of sequences from the ChR recombination library. Here we use these data to train GP classification and regression models to predict the expression and localization properties of diverse, untested ChR sequences. We first made predictions on sequences within a large library of chimeric ChRs; we then expanded the predictions to sequences outside that set.
The design and characterization of the chimeric ChR sequences used to train our models have been published [5]; we will only briefly describe these results. Two separate, ten-block libraries were designed by recombining three parental ChRs (CsChrimsonR (CsChrimR) [28], C1C2 [29], and CheRiff [30]) with 45–55% amino acid sequence identity and a range of expression, localization, and functional properties (S1 Fig) [5]. Each chimeric ChR variant in these libraries is composed of blocks of sequence from the parental ChRs. These libraries were prepared by the SCHEMA algorithm to define sequence blocks for recombination that minimize the library-average disruption of tertiary protein structure [31,32]. One library swaps contiguous elements of primary structure (contiguous library), and the second swaps elements that are contiguous in the tertiary structure but not necessarily in the sequence (non-contiguous library [33]). The two libraries have similar, but not identical, element boundaries (S1A Fig) and were constructed in order to test whether one design approach was superior to the other (they gave similar results). These designs generate 118,098 possible chimeras (2 x 310), which we will refer to as the recombination library throughout this paper. Each of these chimeras has a full N-terminal signal peptide from one of the three ChR parents.
Two hundred and eighteen chimeras from the recombination library were chosen as a training set, including all the chimeras with single-block swaps (chimeras consisting of 9 blocks of one parent and a single block from one of the other two parents) and multi-block-swap chimera sequences designed to maximize mutual information between the training set and the remainder of the chimeric library. Here, the ‘information’ a chimera has to offer is how its sequence, relative to all previously tested sequences, changes ChR expression and localization. By maximizing mutual information, we select chimera sequences that provide the most information about the whole library by reducing the uncertainty (Shannon entropy) of prediction for the remainder of the library, as described in [34,35]. The 112 single-block-swap chimeras in the training set have an average of 15 mutations from the most closely related parent, while the 103 multi-block-swap chimeras in the training set have an average of 73 mutations from the most closely related parent (Table 1). While the multi-block-swap chimeras provide the most sequence diversity to learn from, they are the least likely to express and localize given their high mutation levels. The single-block-swap chimeras offer less information to learn from due to their sequence redundancies with other chimeras in the training set, but are more likely to express and localize.
Genes for these sequences were synthesized and expressed in human embryonic kidney (HEK) cells, and their expression and membrane localization properties were measured (S1B Fig) [5]. The expression levels were monitored through a fluorescent protein (mKate) fused to the C-termini of the ChRs. Plasma-membrane localization was measured using the SpyTag/SpyCatcher labeling method, which exclusively labels ChR protein that has its N terminus exposed on the extracellular surface of the cell [36]. The training set sequences displayed a wide range of expression and localization properties. While the majority of the training set sequences express, only 33% of the single-block-swap chimeras localize well, and an even smaller fraction (12%) of the multi-block-swap chimeras localize well, emphasizing the importance of having a predictive model for membrane localization.
First we explored whether ChR chimera properties could be predicted based on basic biological properties, specifically, signal peptide sequence and hydrophobicity in the transmembrane (TM) domains. Each chimera in the library has one of the three parental signal peptides. Although the signal peptide sequence does affect expression and localization (S2A Fig), chimeras with any parental signal peptide can have high or low expression and localization. Thus, the identity of the signal peptide alone is insufficient for accurate predictions of the ChR chimera properties. We then calculated the level of hydrophobicity within the 7-TM domains of each chimera. With very weak correlation between increasing hydrophobicity and measured expression and localization (S2B Fig), hydrophobicity alone is also insufficient for accurate prediction of ChR chimera properties. These models do not accurately account for the observed levels of expression or localization (S1 Fig). Therefore, we need more expressive models to predict expression and localization from the amino acid sequences of these MPs.
Our overall strategy for developing predictive machine-learning models is illustrated in Fig 1. The goal is to use a set of ChR sequences and their expression and localization measurements to train GP regression and classification models that describe how ChR properties depend on sequence and predict the behavior of untested ChRs. GP models infer predictive values from training examples by assuming that similar inputs (ChR sequence variants) will have similar outputs (expression or localization). We quantify the relatedness of inputs (ChR sequence variants) by comparing both sequence and structure. ChR variants with few differences are considered more similar than ChR variants with many differences. We define the sequence similarity between two chimeras by aligning them and counting the number of positions at which they are identical. For structural comparisons, a residue-residue ‘contact map’ was built for each ChR variant, where two residues are in contact if they have any non-hydrogen atoms within 4.5 Å. The maps were generated using a ChR parental sequence alignment and the C1C2 crystal structure, which is the only available ChR structure [29], with the assumption that ChR chimeras share the overall contact architecture observed in the C1C2 crystal structure. The structural similarity for any two ChRs was quantified by aligning the contact maps and counting the number of identical contacts [26]. Using these metrics, we calculated the sequence and structural similarity between all ChRs in the training set relative to one another (218 x 218 ChR comparisons).
These similarity functions are called kernel functions and specify how the functional properties of pairs of sequences are expected to covary (they are also known as covariance functions). In other words, the kernel is a measure of similarity between sequences, and we can draw conclusions about unobserved chimeras on the basis of their similarity to sampled points [25]. The model has high confidence in predicting the properties of sequences that are similar to previously sampled sequences, and the model is less confident in predicting the properties of sequences that are distant from previously sampled sequences.
To build a GP model, we must also specify how the relatedness between sequences will affect the property of interest, in other words how sensitive the ChR properties are to changes in relatedness as defined by the sequence/structure differences between ChRs. This is defined by the form of the kernel used. We tested three different forms of sequence and structure kernels: linear kernels, squared exponential kernels, and Matérn kernels (see Methods). These different forms represent the kinds of functions we expect to observe for the protein’s fitness landscape (i.e. the mapping of protein sequence to protein function). The linear kernel corresponds to a simple landscape where the effects of changes in sequence/structure are additive and there is no epistasis. The two non-linear kernels represent more rugged, complex landscapes where effects may be non-additive. Learning involves optimizing the form of the kernel and its hyperparameters (parameters that influence the form of kernel) to enable accurate predictions. The hyperparameters and the form of the kernel were optimized using the Bayesian method of maximizing the marginal likelihood of the resulting model. The marginal likelihood (i.e. how likely it is to observe the data given the model) rewards models that fit the training data well while penalizing model complexity to prevent overfitting.
Once trained with empirical data, the output of the GP regression model is a predicted mean and variance, or standard deviation, for any given ChR sequence variant. The standard deviation is an indication of how confident the model is in the prediction based on the relatedness of the new input relative to the tested sequences.
We used GP models to infer links between ChR properties and ChR sequence and structure from the training data. We first built GP binary classification models. In binary classification, the outputs are class labels i.e. ‘high’ or ‘low’ localization, and the goal is to use the training set data to predict the probability of a sequence falling into one of the two classes (Fig 1). We also built a GP regression model that makes real-valued predictions, i.e. amount of localized protein, based on the training data (Fig 1). After training these models, we verify that their predictions generalize to sequences outside of the training set. Once validated, these two models can be used in different ways. A classification model trained from localization data can be used to predict the probability of highly diverse sequences falling into the ‘high’ localization category (Fig 1). The classification model can only predict if a sequence has ‘high’ vs ‘low’ localization, and it cannot be used to optimize localization. The regression model, on the other hand, can be used to predict sequences with ‘optimal’ properties; for example, a regression model trained from localization data can predict untested sequences that will have very high levels of localization (Fig 1).
The training set data (S1 Fig) were used to build a GP classification model that predicted which of the 118,098 chimeras in the recombination library would have ‘high’ vs ‘low’ expression, localization, and localization efficiency. The training set includes multi-block swaps chosen to be distant from other sequences in the training set in order to provide information on sequences throughout the recombination library. A sequence was considered ‘high’ if it performed at least as well as the lowest performing parent, and it was considered ‘low’ if it performed worse than the lowest performing parent. Because the lowest performing parent for expression and localization, CheRiff, is produced and localized in sufficient quantities for downstream functional studies, we believe this to be an appropriate threshold for ‘high’ vs ‘low’ performance. For all of the classification models (Fig 2 and S3 Fig), we used kernels based on structural relatedness. For the expression classification model, we found that a linear kernel performed best, i.e. achieved the highest marginal likelihood. This suggests that expression is best approximated by an additive model weighting each of the structural contacts. Localization and localization efficiency required a non-linear kernel for the model to be predictive. This more expressive kernel allows for non-linear relationships and epistasis and also penalizes differing structural contacts more than the linear kernel. This reflects our intuitive understanding that localization is a more demanding property to tune than expression, with stricter requirements and a non-linear underlying fitness landscape.
Most of the multi-block-swap sequences from the training set did not localize to the membrane [5]. We nonetheless want to be able to design highly mutated ChRs that localize well because these are most likely to have interesting functional properties. We therefore used the localization classification model to identify multi-block-swap chimeras from the library that had a high predicted probability (>0.4) of falling into the ‘high’ localizer category (Fig 2D). From the many multi-block-swap chimeras predicted to have ‘high’ localization, we selected a set of 16 highly diverse chimeras with an average of 69 amino acid mutations from the closest parent and called this the ‘exploration’ set (S4 Fig). We synthesized and tested these chimeras and found that the model had accurately predicted chimeras with good localization (Fig 2 and Fig 3): 50% of the exploration set show ‘high’ localization compared to only 12% of the multi-block-swap sequences from the original training set, even though they have similar levels of mutation (Table 1 and S1 Data) (chimeras in the exploration set have on average 69 ± 12 amino acid mutations from the closest parent, versus 73 ± 21 for the multi-block-swap chimeras in the training set). The classification model provides a four-fold enrichment in the number of chimeras that localize well when compared to randomly-selected chimeras with equivalent levels of mutation. This accuracy is impressive given that the exploration set was designed to be distant from any sequence the model had seen during training. The model’s performance on this exploration set indicates its ability to predict the properties of sequences distant from the training set.
The data from the exploration set were then used to better inform our models about highly diverse sequences that localize. To characterize the classification model’s performance, we calculated the area under the receiver operating characteristic (ROC) curve (AUC). A poorly performing model would not do better than random chance, resulting in an AUC of 0.5, while a model that perfectly separates the two classes will have an AUC of 1.0. The revised models achieved AUC up to 0.87 for “leave-one-out” (LOO) cross-validation, indicating that there is a high probability that the classifiers will accurately separate ‘high’ and ‘low’ performing sequences for the properties measured. The AUC is 0.83 for localization, 0.77 for localization efficiency and 0.87 for expression for LOO cross-validation predictions (S5 Fig).
To further test the models, we then built a verification set of eleven chimeras, designed using the localization model. This verification set was composed of four chimeras predicted to be highly likely to localize, six chimeras predicted to be very unlikely to localize, and one chimera with a moderate predicted probability of localizing (S4 Fig). The measured localization (Fig 2E) and localization efficiency (S3B Fig) of the chimeras in the verification set show clear differences, ‘high’ vs ‘low’, consistent with the model predictions (Table 1 and S1 Data). The verification sets consist exclusively of chimeras with ‘high’ measured expression, which is consistent with the model’s predictions (Fig 2B). The model perfectly classifies the eleven chimeras as either ‘high’ or ‘low’ for each property (expression, localization, or localization efficiency) as shown in plots of predicted vs measured properties (Fig 2B and 2E and S3B Fig) and by perfect separation in ROC curves i.e. AUC = 1.0 (S5 Fig). These models are powerful tools that can confidently predict whether a chimera will have 'high' or 'low' expression (Fig 2C), localization (Fig 2F), and localization efficiency (S3C Fig). Of the 118,098 chimeras in the recombination library, 6,631 (5.6%) are predicted to have a probability > 0.5 of 'high' localization, whereas the vast majority of chimeras (99%) are predicted to have a probability > 0.5 of 'high' expression.
The classification model predicts the probability that a sequence falls into the ‘high’ localizer category, but does not give a quantitative prediction as to how well it localizes. Our next goal was to design chimera sequences with optimal localization. Localization is considered optimal if it is at or above the level of CsChrimR, the best localizing parent, which is more than adequate for in vivo applications using ChR functionality to control neuronal activity [28]. A regression model for ChR plasma membrane localization is required to predict sequences that have optimal levels of localization. We used the localization data from the training and exploration sets to train a GP regression model (Fig 4A). The diversity of sequences in the training data allows the model to generalize well to the remainder of the recombination library. For this regression model, we do not use all of the features from the combined sequence and structure information; instead, we used L1 linear regression to select a subset of these features. The L1 linear regression identifies the sequence and structural features that most strongly influence ChR localization. Using this subset of features instead of all of the features improved the quality of the predictions (as determined by cross-validation). This indicates that not all of the residues and residue-residue contacts have a large influence on localization of ChR. We then used a kernel based on these chosen features (specific contacts and residues) for GP regression. The regression model for localization showed strong predictive ability as indicated by the strong correlation between predicted and measured localization for LOO cross-validation (correlation coefficient, R > 0.76) (Fig 4A). This was further verified by the strong correlation between predicted and measured values for the previously-discussed verification set (R > 0.9) (Fig 4A). These cross-validation results suggest that the regression model can be used to predict chimeras with optimal localization.
We used the localization regression model to predict ChR chimeras with optimal localization using the Lower Confidence Bound (LCB) algorithm, in which the predicted mean minus the predicted standard deviation (LB1) is maximized [37]. The LCB algorithm maximally exploits the information learned from the training set by finding sequences the model is most certain will be good localizers. The regression model was used to predict the localization level and standard deviation for all chimeras in the library, and from this the LB1 was calculated for all chimeras (Fig 4B). We selected four chimeras whose LB1 predictions for localization were ranked in the top 0.1% of the library (S4 Fig). These were constructed and tested (Fig 3 and S6 Fig and S1 Data). Measurements showed that they all localize as well as or better than CsChrimR (Fig 3 and Fig 4A and Table 1). Cell population distributions of the optimal set show properties similar to the CsChrimR parent, with one chimera showing a clear shift in the peak of the distribution towards higher levels of localization (S7 Fig). These four sequences differ from CsChrimR at 30 to 50 amino acids (S4 Fig).
We were interested in how predictive the GP localization models could be with fewer training examples. To assess the predictive ability of the GP models as a function of training set size, we sampled random sets of training sequences from the dataset, trained models on these random sets, then evaluated the model’s performance on a selected test set (S8 Fig). As few as 100 training examples are sufficient for accurate predictions for both the localization regression and classification models. This analysis shows that the models would have been predictive with even fewer training examples than we chose to use.
In developing the GP regression model for localization, we used L1-regularized linear regression to identify a limited set of sequence and structural features that strongly influence ChR localization (Fig 4). These features include both inter-residue contacts and individual residues and offer insight into the structural determinants of ChR localization. To better gauge the relative importance of these features, L2-regularized linear regression was used to calculate the positive and negative feature weights, which are proportional to each feature’s inferred contribution to localization. While not as predictive as the GP regression model because it cannot account for higher-order interactions between features, this linear model has the advantage of being interpretable.
When mapped onto the C1C2 structure, these features highlight parts of the ChR sequence and structural contacts that are important for ChR localization to the plasma membrane (Fig 5). Both beneficial and deleterious features are distributed throughout the protein, with no single feature dictating localization properties (Fig 5). Clusters of heavily weighted positive contacts suggest that having structurally proximal CsChrimR-residue pairs are important in the N-terminal domain (NTD), between the NTD and TM4, between TM1 and TM7, and between TM3 and TM7. CsChrimR residues at the extracellular side of TM5 also appear to aid localization, although they are weighted less than CheRiff residues in the same area. Beneficial CheRiff contacts and residues are found in the C-terminal domain (CTD), the interface between the CTD and TM5-6, and in TM1. C1C2 residues at the extracellular side of TM6 are also positively weighted for localization, as are C1C2 contacts between the CTD and TM3-4 loop. From the negatively weighted contacts, it is clear that total localization is harmed when CheRiff contributes to the NTD or the intracellular half of TM4 and when CsChrimR contributes to the CTD. Interestingly, positive contacts were formed between TM6 from C1C2 and TM7 from CheRiff, but when the contributions were reversed (TM6 from CheRiff TM7 from C1C2) or if CsChrimR contributed TM6, strong negative weights were observed. Not surprisingly, the sequence and structure of optimal localizers predicted by GP regression (Fig 4) largely agree with the L2 weights (S9 Fig).
Using this strategy for model interpretation (L1 regression for feature selection followed by L2 regression), we can also weight the contributions of residues and contacts for ChR expression (S10 Fig and S11 Fig). There is some overlap between the heavily weighted features for ChR expression and the features for localization, which is expected because more protein expressed means more protein available for localization. For example, both expression and localization models seem to prefer the NTD from CsChrimR and the extracellular half of TM6 from C1C2, and both disfavor the NTD and the intra-cellular half of TM4 from CheRiff. While the heavily-weighted expression features are limited to these isolated sequence regions, localization features are distributed throughout the protein. Moreover, the majority of heavily-weighted features identified for expression are residues rather than contacts. This is in contrast to those weighted features identified for localization, which include heavily-weighted residues and structural contacts. This suggests that sequence is more important in determining expression properties, which is consistent with the largely sequence-dependent mechanisms associated with successful translation and insertion into the ER membrane. In contrast, both sequence and specific structural contacts contribute significantly to whether a ChR will localize to the plasma membrane. Our results demonstrate that the model can ‘learn’ the features that contribute to localization from the data and make accurate predictions on that property.
We next tested the ChR localization regression model for its ability to predict plasma-membrane localization for ChR sequences outside the recombination library. For this, we chose a natural ChR variant, CbChR1, that expresses in HEK cells and neurons but does not localize to the plasma membrane and thus is non-functional [28]. CbChR1 is distant from the three parental sequences, with 60% identity to CsChrimR and 40% identity to CheRiff and C1C2. We optimized CbChR1 by introducing minor amino acid changes predicted by the localization regression model to be beneficial for membrane localization. To enable measurement of CbChR1 localization with the SpyTag-based labeling method, we substituted the N-terminus of CbChR1 with the CsChrimR N-terminus containing the SpyTag sequence downstream of the signal peptide to make the chimera CsCbChR1 [36]. This block swap did not change the membrane localization properties of CbChR1 (Fig 6C). Using the regression model, we predicted localization levels for all the possible single-block swaps from the three library parents (CsChrimR, C1C2 and CheRiff) into CsCbChR1 and selected the four chimeras with the highest Upper Confidence Bound (UCB). These chimeras have between 4 and 21 mutations when compared with CsCbChR1. Unlike the LCB algorithm, which seeks to find the safest optimal choices, the UCB algorithm balances exploration and exploitation by maximizing the sum of the predicted mean and standard deviation.
The selected chimeras were assayed for expression, localization, and localization efficiency (S1 Data). One of the four sequences did not express; the other three chimeras expressed and had higher localization levels than CsCbChR1 (Fig 6B). Two of the three had localization properties similar to the CheRiff parent (Fig 6B). Images of the two best localizing chimeras illustrate the enhancement in localization when compared with CbChR1 and CsCbChR1 (Fig 6C and S12 Fig). This improvement in localization was achieved through single-block swaps from CsChrimR (17 and 21 amino acid mutations) (Fig 6A). These results suggest that this regression model can accurately predict minor sequence changes that will improve the membrane localization of natural ChRs.
The ability to differentiate the functional properties of closely related sequences is extremely powerful for protein design and engineering. This is of particular interest for protein types that have proven to be more recalcitrant to traditional protein design methods, e.g. MPs. We show here that integral membrane protein expression and plasma membrane localization can be predicted for novel, homologous sequences using moderate-throughput data collection and advanced statistical modeling. We have used the models in four ways: 1) to accurately predict which diverse, chimeric ChRs are likely to express and localize at least as well as a moderately-performing native ChR; 2) to design ChR chimeras with optimized membrane localization that matched or exceeded the performance of a very well-localizing ChR (CsChrimR); 3) to identify the structural interactions (contacts) and sequence elements most important for predicting ChR localization; and 4) to identify limited sequence changes that transform a native ChR from a non-localizer to a localizer.
Whereas 99% of the chimeras in the recombination library are predicted to express in HEK cells, only 5.6% are predicted to localize to the membrane at levels equal to or above the lowest parent (CheRiff). This result shows that expression is robust to recombination-based sequence alterations, whereas correct plasma-membrane localization is much more sensitive. The model enables accurate selection of the rare, localization-capable, proteins from the nearly 120,000 possible chimeric library variants. In future work we will show that this diverse set of several thousand variants predicted to localize serves as a highly enriched source of functional ChRs with novel properties.
Although statistical models generalize poorly as one attempts to make predictions on sequences distant from the sequences used in model training, we show that it is possible to train a model that accurately distinguishes between closely related proteins. The tradeoff between making accurate predictions on subtle sequence changes vs generalized predictions for significantly different sequences is one we made intentionally in order to achieve accurate predictions for an important and interesting class of proteins. Accurate statistical models, like the ones described in this paper, could aid in building more expressive physics-based models.
This work details the steps in building machine-learning models and highlights their power in predicting desirable protein properties that arise from the intersection of multiple cellular processes. Combining recombination-based library design with statistical modeling methods, we have scanned a highly functional portion of protein sequence space by training on only 218 sequences. Model development through iterative training, exploration, and verification has yielded a tool that not only predicts optimally performing chimeric proteins, but can also be applied to improve related ChR proteins outside the library. As large-scale gene synthesis and DNA sequencing become more affordable, machine-learning methods such as those described here will become ever more powerful tools for protein engineering offering an alternative to high-throughput assay systems.
The design, construction, and characterization of recombination library chimeras is described in Bedbrook et al. [5]. Briefly, HEK 293T cells were transfected with purified ChR variant DNA using Fugene6 reagent according to the manufacturer’s recommendations. Cells were given 48 hours to express before expression and localization were measured. To assay localization level, transfected cells were subjected to the SpyCatcher-GFP labeling assay, as described in Bedbrook et al. [36]. Transfected HEK cells were then imaged for mKate and GFP fluorescence using a Leica DMI 6000 microscope (for cell populations) or a Zeiss LSM 780 confocal microscope (for single cells: S12 Fig). Images were processed using custom image processing scripts for expression (mean mKate fluorescence intensity) and localization (mean GFP fluorescence intensity). All chimeras were assayed under identical conditions.
For each chimera, net hydrophobicity was calculated by summing the hydrophobicity of all residues in the TM domains. The C1C2 crystal structure was used to identify residues within TM domains (S2B Fig), and the Kyte & Doolittle amino acid hydropathicity scale [38] was used to score residue hydrophobicity.
Both the GP regression and classification modeling methods applied in this paper are based on work detailed in [26]. Romero et al. applied GP models to predict protein functions and also defined protein distance using a contact map. We have expanded on this previous work. Regression and classification were performed using open-source packages in the SciPy ecosystem [39–41]. Below are specifics of the GP regression and classification methods used in this paper. The hyperparameters and the form of the kernel were optimized using the Bayesian method of maximizing the marginal likelihood of the resulting model.
|
10.1371/journal.pntd.0006088 | Turning poop into profit: Cost-effectiveness and soil transmitted helminth infection risk associated with human excreta reuse in Vietnam | Human excreta is a low cost source of nutrients vital to plant growth, but also a source of pathogens transmissible to people and animals. We investigated the cost-savings and infection risk of soil transmitted helminths (STHs) in four scenarios where farmers used either inorganic fertilizer or fresh/composted human excreta supplemented by inorganic fertilizer to meet the nutrient requirements of rice paddies in the Red River Delta, Vietnam. Our study included two main components: 1) a risk estimate of STH infection for farmers who handle fresh excreta, determined by systematic review and meta-analysis; and 2) a cost estimate of fertilizing rice paddies, determined by nutrient assessment of excreta, a retailer survey of inorganic fertilizer costs, and a literature review to identify region-specific inputs. Our findings suggest that farmers who reuse fresh excreta are 1.24 (95% CI: 1.13–1.37, p-value<0.001) times more likely to be infected with any STH than those who do not handle excreta or who compost appropriately, and that risk varies by STH type (Ascaris lumbricoides RR = 1.17, 95% CI = 0.87–1.58, p-value = 0.29; Hookworm RR = 1.02, 95% CI = 0.50–2.06, p-value = 0.96; Trichuris trichiura RR = 1.38, 95% CI = 0.79–2.42, p-value = 0.26). Average cost-savings were highest for farmers using fresh excreta (847,000 VND) followed by those who composted for 6 months as recommended by the WHO (312,000 VND) and those who composted for a shorter time (5 months) with lime supplementation (37,000 VND/yr); however, this study did not assess healthcare costs of treating acute or chronic STH infections in the target group. Our study provides evidence that farmers in the Red River Delta are able to use a renewable and locally available resource to their economic advantage, while minimizing the risk of STH infection.
| Each year, hundreds of millions of people worldwide are infected with intestinal worms spread by contaminated soil, also known as soil transmitted helminths (STHs). These worms are most common in tropical climates in areas lacking good hygiene and sanitation, and negatively impact child development, quality of life, and economic wellbeing. Reuse of human excreta for fertilizer is a common practice in many low to middle income countries because farmers require a low cost source of nutrients to grow food crops eaten by people and animals. Excreta can contain microbes, such as STHs, that cause disease in people; however, composting is a known method of killing STHs. Therefore, our goal was to determine if Vietnamese rice farmers involved in this practice are at higher risk of STH infection, and to calculate the amount of money saved by farmers composting for different lengths of time, and supplementing with various commercial fertilizers. We suggest that farmers compost excreta for six months to reduce disease exposure and optimize household savings. Optimizing practices to improve food production and protect farmer health is critical for poverty alleviation in low to middle income countries.
| Application of human excreta onto rice paddies as fertilizer is a common practice in northern Vietnam, where many farmers use single or double vault latrines, lack access to wastewater infrastructure, and have variable access to commercial inorganic fertilizers [1]. Using organic waste to fertilize fields has clear benefits for crop yield [2]; however, this practice increases certain health risks for farmers and consumers, such as infection by soil transmitted helminths (STHs)[3,4]. The STH group includes Ascaris lumbricoides, Trichuris trichiura, and hookworm spp., which are intestinal parasites that spread between people when sanitation is inadequate or when good hygiene is not practiced [4]. People are infected when they accidentally ingest infective eggs or when their skin contacts infective larvae in contaminated soil. These parasites are particularly prevalent in regions with warm, moist climates, and are included in the category of tropical neglected diseases associated with poverty.
World Health Organization (WHO) guidelines recommend that farmers compost human excreta for six months prior to application in order to inactivate STH eggs and larvae, and thereby reduce spread between people [5]. This practice is not feasible for all Vietnamese farmers, in particular those who harvest multiple crops per year or have single vault latrines that lack a chamber for long-term excreta storage. Current evidence suggests that only one-third of farmers who use human excreta follow the six-month recommendation [6], and that STH infection remains an occupational hazard associated with handling human excreta [3]. It is common practice for household members to add a handful of kitchen ash after using a latrine, as this reduces smell. A recent study characterizing A. lumbricoides egg die-off during excreta composting suggests that adding lime reliably accelerates egg inactivation so that WHO criteria for safe handling (<1 viable egg/g total solids) are met by 153 days [7]. Ascarid eggs can survive longer periods in adverse environmental conditions than other STHs, and for that reason we chose A. lumbricoides die-off as a proxy for overall STH die-off [8].
Rice farmers in some agricultural regions of Vietnam have shifted their source of fertilizer from human excreta to commercial inorganic products, either wholly or in part. It is unclear whether this trend will become universal as not all farmers are able to afford or access commercial fertilizer, and others consider human excreta a superior source of long-term nutrition for plants and soil [9]. Inorganic fertilizers are primarily imported, and their costs are influenced by a wide range of factors, including energy prices [10]. Using human waste to fertilize crops is recognized as a way to decrease household expenditures; however, it is unclear how costs and health risks associated with STH infection interact. The goal of this study was to compare the costs and STH risk associated with fertilizing rice paddies in the Red River Delta (RRD).
The RRD encompasses eight provinces and two major urban municipalities (Hanoi and Haiphong) in northern Vietnam. The RRD is an agriculturally intense area that produces approximately 15% of the national annual rice output [11]. Throughout the region, farmers use various combinations of human excreta, inorganic fertilizers, and animal manure to replenish soil nutrients and maximize rice yield. To generate cost estimates, we chose four fertilization scenarios: (A) Fresh human excreta (≤ 139 day storage without lime); (B) Composted human excreta (153 day storage with 10% lime as per [7]); (C) Composted human excreta (181 day storage without lime; WHO standard [5]); (D) Inorganic fertilizer. Although three scenarios (A-C) involved human excreta, we assumed that only farmers who handled fresh excreta (A) would experience STH infection risk, as the composting scenarios (B and C) met WHO standards for helminth inactivation. Risk of STH infection for Vietnamese farmers handling fresh excreta was evaluated by systematic review and meta-analysis. Our economic analysis of the four scenarios included the direct costs incurred for composting human excreta (i.e. lime) and supplementing excreta with inorganic fertilizers. Capital costs (i.e. cost to build a double vault latrine) were not included because differences in factors such as materials and design cause costs to vary substantially in the RRD, and would add a high level of uncertainty to our analysis. To estimate the direct costs, we determined nutrient content of organic fertilizer scenarios, conducted a retailer survey of inorganic fertilizers in the study area, and collected economic inputs from published sources specific to the RRD (e.g. household size, excreta production per household, annual harvest frequency, average paddy size).
This study adds to current knowledge about the opportunities and risks associated with reusing human excreta to fertilize rice plants in one region of Vietnam. Our finding that handling fresh excreta increases the risk of STH infection in farmers (RR = 1.24, 95% CI: 1.13–1.37) emphasizes the importance of adequately treating excreta to inactivate STH life stages. Furthermore, the risk is not limited to farmers as fresh excreta reuse facilitates STH spread to other commune residents, and ultimately to consumers, through food, water, and environmental transmission routes. This practice is one factor contributing to the high prevalence of A. lumbricoides (44.4%; N = 34 million), T. trichiura (23.1%; N = 17.6 million) and hookworm (28.6%; N = 21.8 million) infections in Vietnam [23]. Our study did not assess healthcare costs associated with STH prevention, treatment, or chronic disability. Individuals with low intensity infections are often asymptomatic; however, those with high intensity infections can experience a variety of acute or chronic conditions (e.g. diarrhea, abdominal discomfort, anemia and rectal prolapse) that reduce quality of life and may require costly medical interventions to treat [4]. Long-term sequelae of chronic infections, such as impaired cognitive development and growth faltering, can negatively impact lifelong earnings and contribute to the cycle of poverty in low resource communities. Although our study showed that farmers using fresh excreta benefitted from the largest cost-savings in fertilizer expenditure, the direct and indirect societal costs incurred due to prolonged STH infection would likely outweigh these savings.
Despite laws that prohibit use of human excreta for agriculture in Vietnam, this practice remains common among certain farming groups [24]. Human excreta is perceived as more valuable than animal manure due to differences in dietary protein content, and it is believed to improve soil structure more sustainably than inorganic fertilizers [9]. Although many farmers compost excreta, WHO recommendations for hygienic composting are not commonly followed, as farmers harvest multiple crops per year and are unwilling or unable to store excreta for six months prior to use [5,6]. Some misperceptions about the reasons for safe composting might influence farmer willingness to use fresh excreta. For example, focus group participants in the RRD emphasized ease of application and benefits to soil structure, rather than the benefits of composting to protect human health [1]. Our alternative to the WHO standard, composting for 153 days with 10% lime to accelerate STH inactivation, was not a reasonable alternative for farmers prioritizing cost savings. However, for farmers less concerned about cost savings, the 10% lime compost strategy could be further accelerated to 111 days by adding aeration to latrines, which would allow excreta to be safely handled at more frequent intervals [7].
Human excreta use in crop agriculture was previously estimated to represent 83 million USD in fertilizer import savings to the Vietnamese economy [6], which is one-fifth of the 2014 net expenditure on inorganic fertilizer importation (384 million USD) [25]. Our cost analysis indicated that farmers could save 37,000–847,000 VND/yr (1.48–37.28 USD, $2017) [26] by using human excreta. While these savings might appear low, they represent 1–22% of a farmer’s average annual income in the RRD [27]. Furthermore, the savings could represent a higher percent of annual income in regions that are less fertile, where rice yields are lower, or in remote locations where transportation challenges result in higher commercial fertilizer costs. Another report, suggesting that household excreta traded on the domestic market could contribute up to 15% of household income for those in the lowest income quintile, is in line with our analysis [6]. Therefore, it is unlikely that low-income rural farmers would be willing to universally replace organic fertilizers of human origin with inorganic commercial fertilizers. This was previously demonstrated by farmers who were given non-composting latrines and who ultimately broke the seals open to access the excreta [24].
Our nutrient analysis of human excreta originating from the RRD and composted over time demonstrated excreta to be an adequate organic source of phosphorus and potassium, but not nitrogen, for plant growth. Therefore, all of the scenarios using human excreta (A-C) required additional inorganic fertilizer in order to meet the recommendations for optimizing rice yield. It is not clear how far outside Vietnam these results should be extrapolated as differences in dietary intake directly influence NPK excretion, and soil supplementation requirements vary regionally. Furthermore, our analysis was based on total excreta collected in a double vault latrine, rather than waste separated into liquid and solid components, as occurs in some other regions that use excreta. However, beyond the immediate economic and agricultural gains to reusing excreta, there are global benefits to nutrient recycling. It is estimated that the demand of phosphate rock will outweigh supply by the mid-21st century, which has important consequences for food security as phosphorus is essential for plant growth [28]. As access to a hygienic toilet (flush, pour flush, sulabh or double vault latrine) is still regionally variable in Vietnam (61.6–96.7% of homes containing a latrine), an opportunity currently exists to optimize nutrient recovery infrastructure in homes requiring sanitation upgrades [27].
Our systematic review and meta-analysis found a statistically significant higher risk of infection with any STH among Vietnamese farmers who use human excreta, and highlighted the limited volume of evidence to describe this association. Only four studies met our inclusion criteria, despite searching academic and grey-literature sources in Vietnamese and English. Of these, only three were included in meta-analysis due to poor reporting quality. Out of a possible score of 13, two studies achieved a quality score of 50% or lower. Our findings showed that studies often did not report descriptions of appropriate sample size determinations, confounders controlled for, sample size for positive exposure and/or outcomes, as well as risk estimates or associated p-values. However, aside from one study, all studies were from the RRD study area, included participants of similar age and gender, and estimated exposure and outcomes using similar methods. Each study had slightly different definitions for agricultural use of human excreta, and therefore our meta-analysis included both individuals whose primary occupation was rice farming, but also those who worked with human excreta in other ways aside from direct field application. Inclusion of Yajima et al., 2009 in the meta-analyses of both T. trichiura and A. lumbricoides produced significant heterogeneity in the estimates of infection risk. This study had a small sample size and very low prevalence of STHs (i.e. one case of A. lumbricoides detected), leading to low risk ratios and wide confidence intervals. The study did not provide information on approaches used to measure or control for potential confounders, further adding to the difficulty in interpretation of protective properties of human excreta use in STH. Pooled results by STH type revealed the lowest risk for hookworm infection, and significant variation in studies combined in meta-analysis for this outcome. It was not possible to examine potential factors contributing to heterogeneity in hookworm risk estimates due inclusion of only two studies in meta-analysis. Therefore, in order to better understand factors influencing infection and to substantiate the limited body of evidence on STH risk in the RRD of Vietnam, additional research employing high methodological rigour is warranted.
Although our study attempts to represent the typical situation in the RRD, much of our data comes from Ha Nam province exclusively which may differ in relevant ways from other RRD provinces. The estimate of risk was limited by the number of estimates reported in the literature and we assumed that STH risk was equal in scenarios B-C. Human excreta was collected from various households and mixed before analysis. Thus, results may not accurately reflect the nutrient and moisture content of unmixed excreta if collected and analysed from time of defecation. Costs related to the removal of excess human excreta, latrine construction and maintenance, and personal safety equipment were not explored.
Our study confirmed that human excreta is a significant and sustainable source of nutrients needed for crop fertilization. Its use as agricultural fertilizer, a common practice in Vietnam, offers direct benefits to rice farmers. Human health, agricultural productivity, household earnings are optimized when farmers follow WHO standards for excreta use and government standards for crop fertilization; however current policies prohibit excreta use altogether and therefore may need to be revisited. Furthermore, our results suggest that farmers and the Vietnamese economy would benefit by forward thinking public health messaging promoting STH prevention, such as safe excreta handling strategies, personal protective equipment (e.g. gloves and boots) and regular anthelmintic prophylaxis, rather than an outright ban on excreta use. This study highlights agricultural policies needing further attention, and demonstrates the value of promoting research that provides innovative solutions for safely and economically extracting nutrients from human excreta.
|
10.1371/journal.pgen.1005354 | Silencing of DNase Colicin E8 Gene Expression by a Complex Nucleoprotein Assembly Ensures Timely Colicin Induction | Colicins are plasmid-encoded narrow spectrum antibiotics that are synthesized by strains of Escherichia coli and govern intraspecies competition. In a previous report, we demonstrated that the global transcriptional factor IscR, co dependently with the master regulator of the DNA damage response, LexA, delays induction of the pore forming colicin genes after SOS induction. Here we show that IscR is not involved in the regulation of nuclease colicins, but that the AsnC protein is. We report that AsnC, in concert with LexA, is the key controller of the temporal induction of the DNA degrading colicin E8 gene (cea8), after DNA damage. We demonstrate that a large AsnC nucleosome-like structure, in conjunction with two LexA molecules, prevent cea8 transcription initiation and that AsnC binding activity is directly modulated by L asparagine. We show that L-asparagine is an environmental factor that has a marked impact on cea8 promoter regulation. Our results show that AsnC also modulates the expression of several other DNase and RNase colicin genes but does not substantially affect pore-forming colicin K gene expression. We propose that selection pressure has “chosen” highly conserved regulators to control colicin expression in E. coli strains, enabling similar colicin gene silencing among bacteria upon exchange of colicinogenic plasmids.
| Colicins are considered model proteins for studying bacterial toxins. These narrow spectrum antibiotics can kill by a variety of mechanisms, e.g. by forming pores in the membranes of susceptible cells or by degrading their nucleic acids. Colicin genes are plasmid-encoded and repressed by the master regulator of the DNA damage response, LexA. Induction of several pore-forming colicin genes is also repressed by IscR, which ensures that colicin genes are switched on as a last resort in DNA damaged cells, when nutrients are depleted. Here we show that nuclease colicin genes are not controlled by IscR but that the AsnC protein, in concert with LexA, is directly responsible for uncoupling the immediate expression of the DNase colicin E8 from the main induction of the SOS response. AsnC wraps the DNA of the colicin E8 promoter into a complex nucleoprotein assembly and the architecture of this complex is altered by the presence of the amino acid L-asparagine. Thus, repression by metabolite-responsive and DNA-damage responsive regulators operates at the regulatory regions of different colicins. Hence, the response to several environmental signals have been integrated to ensure that, following DNA damage, colicin synthesis is tightly repressed and induced only in terminally damaged cells.
| Colicins are high-molecular-weight toxic proteins that are produced by and specifically target Escherichia coli and its close relatives [1]. These narrow-spectrum antibiotics kill by either targeting the DNA, RNA or cell membranes of susceptible cells. Cytoplasmic colicins are released upon the synthesis of a lysis protein, the expression of which is independent of intracellular colicin accumulation [2]. This causes the stochastic lysis of producing cells and is suggested to assist surviving sister cells by killing potential competing sensitive cells [3]. Colicin-mediated competition has been suggested to have functions in modulating population dynamics and maintaining diversity of microbial communities [4–7]. Nutrient limitation and DNA damage seem to be the major signals that control colicin production, enabling interference competition among strains [8].
Colicins are plasmid-encoded and are expressed from strong promoters whose activity is tightly repressed by the LexA transcription factor, the master regulator for the SOS DNA damage repair response in bacteria [1,9]. Most of the SOS genes involved in DNA repair and cell division arrest are expressed immediately after DNA damage, but induction of colicin genes is delayed. This presumably provides cells time to repair DNA in order to preserve the integrity of their genome, before the induction of colicin production [10]. In previous work, we established that the global transcriptional repressor IscR delays the induction of the pore-forming colicin K gene (cka) [11]. We showed that IscR participates in a double-locking mechanism, in concert with LexA, by stabilizing the LexA SOS repressor at the promoter and this links colicin expression to the nutritional status of the cell. Thus, the IscR protein uncouples the induction of colicin expression from the temporal induction of the SOS response that deals with repairable DNA damage. This mechanism also operates at other promoters, which control the expression of bactericidal pore-forming colicins [11], however, it is not known if a similar fail-safe double-lock system has also evolved for the nuclease colicin genes.
Here we report that IscR does not modulate the expression of nuclease colicin genes. Hence, we studied the regulation of the DNA degrading colicin E8 gene (cea8) in more depth and identified the AsnC transcription factor as directly responsible for the delay in cea8 expression. AsnC is a member of Lrp/AsnC family of transcriptional regulators that modulate cellular metabolism in both archaea and bacteria [12,13]. In E. coli, AsnC is required to activate the expression of the L-asparagine synthetase A gene (asnA) and this stimulation is abolished in the presence of the amino acid L-asparagine [14]. In addition to this, expression of asnC is negatively autoregulated by AsnC and also repressed by the nitrogen assimilation control (Nac) protein, under nitrogen-limiting conditions [15], however, this regulation is not modulated by the presence of L-asparagine [14]. Functional E. coli AsnC is an octamer, whose structure was resolved by X-ray crystallography [13]. We show that AsnC binds to the cea8 regulatory region at multiple sites, likely wrapping the DNA into a nucleoprotein assembly and that its binding is affected by L-asparagine. In the LexA-AsnC-cea8 complex, two LexA dimers are flanked by multiple AsnC octamers, and the presence of L-asparagine influences AsnC modulated promoter region geometry. Data presented here shows that double locking by LexA and AsnC operates at the cea8 promoter region to delay induction of the colicin E8 gene, thereby linking cea8 expression to DNA damage and L-asparagine availability. Thus, AsnC provides colicinogenic cells with time for DNA damage repair and limits colicin E8 induction to terminally damaged cells.
To study the induction of various nuclease colicins after DNA damage, we assayed the activity of colicin promoters in E. coli K-12 strain BW25113. In this experiment, different DNA fragments, carrying colicin promoters, were cloned into the lac expression vector, pRW50, to give colicin promoter::lac fusions. After inducing the DNA damage response with a sub-inhibitory concentration of nalidixic acid, we observed that the promoters of the DNA degrading colicins E2, E7 and E8, of colicins E5 and D, targeting tRNA, and of the rRNA cleaving colicin E6, are only induced after a prolonged delay (Fig 1A). Note that there is little expression from nuclease colicin promoters in the absence of DNA damage (S1 Fig).
We previously established that the delayed expression of several pore-forming colicins, is due to co-repression by the global transcriptional repressors LexA and IscR [11]. At the colicin K promoter, the LexA repressor was shown to bind to the tandem operators just downstream of the -10 promoter element and prevented RNA polymerase binding. The IscR protein was suggested to increase stability of LexA at these targets. To determine if a similar mechanism controls the expression of the nuclease colicins, we investigated the regulation of the DNA degrading colicin E8 by electrophoretic mobility-shift assays (EMSA) and DNase I footprinting. Our results reveal that at the promoter region of cea8, the LexA repressor binds to two overlapping targets and blocks the access of RNA polymerase to the promoter (Fig 1B and 1C). To investigate the possible binding of IscR to the cea8 promoter, an EMSA assay was again used. Results in Fig 1D show that IscR binds specifically to the cka promoter region but not to that of cea8. In addition, we tested whether the IscR protein is directly responsible for the delayed production of colicin E8 and other nuclease colicins, by comparing the colicin production in our wild-type and ΔiscR strains. Results, illustrated in Fig 1E, show that IscR has a negligible effect on the synthesis of many nuclease colicins. This contrasts with the pore-forming colicin K, where the ΔiscR allele causes a 100 increase after the first hour of induction (S3 Fig) [11]. This indicates that IscR does not regulate any of the nuclease colicin genes and prompted us to search for other transcription factors, involved in controlling the timing of their expression.
To investigate the delay in cea8 induction in SOS-induced cells we used a pull-down assay [11], using a cleared cell extract from mid-logarithmic grown, SOS induced, E. coli cells, and a biotinylated 179 bp cea8 promoter fragment as a bait. Eluted proteins were separated by SDS-polyacrylamide gel electrophoresis and nine bands were analysed by mass spectroscopy (Fig 2A). We identified 30 transcription regulators and nucleoid associated factors that had associated with the bait (S1 Table). To screen for their ability to regulate cea8 expression after DNA damage induction with nalidixic acid, we measured cea8::lac activity in deletion mutants from the Keio collection [16] and we selected strains in which a 3-fold increase in cea8 promoter activity, in comparison to the wild-type strain, was observed (S2 Table). Thus, we focused on the AsnC, StpA, OmpR, YbjK, YihW, YegW and MngR proteins and measured cea8 promoter activity following SOS induction using pRW50 cea8::lac fusion in the corresponding deletion mutant strains throughout the bacterial growth curve. Results presented in Fig 2B show that disruption of asnC resulted in the biggest effects on cea8 promoter induction after DNA damage. An intermediate increase in promoter activity was observed in the strain deficient for stpA, whilst the other deletions had a minimal effect, with our data confirming that IscR does not regulate colicin E8 expression. The StpA protein, a paralogue of the nucleoid-associated protein H-NS, forms a rigid filament along DNA, and can cause DNA bridging [17]. Furthermore, StpA can act as an RNA chaperone [18] and a transcriptional repressor [19,20], thus, it may be involved in colicin gene expression. However, here we focused on AsnC, and assayed its binding to cea8 promoter region and its effect on colicin E8 synthesis. To do this, we introduced the ΔasnC allele into a strain that harbours a cea8-encoding plasmid. After treatment of cells with a subinhibitory concentration of nalidixic acid that induced DNA damage, cell growth and colicin production was compared in the wild-type and the ΔasnC mutant. Our results show that AsnC enhances viability of the strain harbouring the colicin E8-encoding plasmid (Fig 2C). Bioassays were also used to follow colicin levels in crude cell extracts prepared from cells before and after SOS induction. The results show that, in the ΔasnC strain, colicin E8 was produced an hour earlier in comparison to the delayed synthesis in the wild-type strain (Fig 2D). This suggests that AsnC directly represses cea8 promoter activity and, in concert with LexA, ensures regulated and delayed expression of the cea8 gene.
The AsnC protein is a member of the Lrp/AsnC family of regulators, which often assemble to form wheel-like octamers and whose DNA binding activity can be modified by small molecules, such as amino acids [13]. AsnC regulates the expression of its own gene, asnC, and the asnA gene, encoding for a synthetase that catalyses the ammonia-dependent conversion of aspartate to asparagine [14]. To investigate the binding of AsnC to the cea8 promoter (Fig 3A), we over-expressed and purified the AsnC protein and performed in vitro experiments in the presence or absence of the amino acid L-asparagine. EMSA experiments show that several AsnC molecules can interact with cea8 (Fig 3) and that in the presence of L-asparagine a number of distinct complexes can be observed (Fig 3B). In the absence of L-asparagine, at higher AsnC concentrations, DNA remained in the wells of the gel, indicating that high molecular weight nucleoprotein complexes had formed.
DNase I footprinting was also used to study the location of AsnC binding to the cea8 promoter sequence, again in the presence or absence of L-asparagine. Results in Fig 3C show that AsnC interacts along the entire length of the 179 bp cea8 promoter region. Inspection of the cea8 region, interacting with AsnC, revealed 27 DNase I hypersensitive sites, which is indicative of local bending and distortion of the DNA helix. This results in a widening of the minor groove and makes the DNA more susceptible to DNase I attack, leading to the production of hypersensitive bands [21]. In several locations the presence of L- asparagine altered the binding of AsnC to the cea8 promoter (see red boxes in Fig 3C). The position of the red boxes in Fig 3C was determined by comparing the AsnC footprint gels in the presence or absence of L-asparagine. Our in vitro analysis indicates that AsnC binds to the cea8 promoter region at multiple sites, likely wrapping the DNA into a complex nucleoprotein assembly, and that the architecture of this complex is altered by the presence of L-asparagine (Fig 3).
Since our data show that AsnC binds at multiple locations and alters the architecture of colicin E8 regulatory region, as well as binding within LexA target sites (Fig 3), we tested if AsnC and LexA can simultaneously bind to the cea8 promoter region. To investigate this, we performed EMSA analysis on the cea8 promoter fragment. Using purified LexA and AsnC, in the presence of L-asparagine, we observed a large nucleoprotein complex composed of at least two AsnC functional oligomers, presumably octamers [13], and two LexA dimers interacting at cea8 (Fig 4A). Note, that LexA was used at a concentration of 400 nM at which LexA repressor occupies both LexA binding sites within cea8 (Fig 1B). To determine whether occupancy of the DNA by AsnC affects the binding of LexA at the cea8 promoter region, we performed DNase I footprint analysis and compared signatures of LexA and AsnC in the presence or absence of L-asparagine. In both conditions, LexA repressors bound to tandem targets just downstream of the -10 promoter element (Fig 4B). As observed for AsnC binding in Fig 3C, the addition of L-asparagine also modulated the binding of AsnC in the LexA-AsnC nucleoprotein complex (Fig 4B and 4C). In the AsnC-LexA-cea8 complex, specific hypersensitive sites were apparent (determined by stars in Fig 4B), suggesting that the binding of both proteins subtly alters the structure or trajectory of the DNA around -10 element (Fig 4D–4F). Thus, we conclude that concurrent binding of LexA and AsnC to the cea8 regulatory region ensures delayed induction of the DNase E8 synthesis after DNA damage.
Our data suggest that tight repression of DNase colicin E8 might be affected by the availability of amino acid L-asparagine and this signal is relayed via AsnC. To test this hypothesis we measured cea8 promoter activity in SOS-induced wild-type cells grown in the M9 minimal medium containing either 10 mM NH4Cl or 20 mM L-asparagine as the sole source of nitrogen. Fig 5 shows that in the L-asparagine containing medium, the expression from cea8 remains low, whilst it increases in medium containing higher levels of NH4Cl. This is in agreement with our in vitro data (Figs 3 and 4) and suggests that L-asparagine is needed to stabilize a specific AsnC assembly at the cea8 promoter region. Hence, we suggest that depletion of L-asparagine is the signal for AsnC de-repression at the cea8 promoter.
To determine whether AsnC modulates the expression of other colicins, we assayed DNase colicin E2 (cea2) and rRNase colicin E6 (cea6) promoter activities following SOS induction with nalidixic acid using cea6::lacZ and cea2::lacZ promoter fusions in the wild-type, ΔiscR and ΔasnC strains. Results illustrated in Fig 6A and 6B show that the disruption of asnC resulted in elevated cea2 and cea6 promoter activity immediately after DNA damage induction, when compared to the wild-type and the ΔiscR strains. This indicates that AsnC, rather than IscR, is a key transcriptional repressor of the cea2 and cea6 promoters. In addition, we transferred the colicinogenic plasmids for these DNase and RNase colicins into the ΔasnC and wild-type strains. Following the induction of DNA damage, cell growth (Fig 6C) and colicin production (Fig 6D) was monitored in both strains. In the absence of asnC, cells failed to reach as high optical density as in the wild-type strain, suggesting that elevated colicin expression and cell lysis had taken place (Fig 6C). Cell extracts were prepared from cultures before and after DNA damage induction and colicin levels compared by a colicin production bioassay (Fig 6D). After SOS induction, nuclease colicin synthesis was induced earlier for colicins E2, E5 and E6 in the ΔasnC strain, in comparison to the wild-type strain, indicating that AsnC directly modulates the expression of a number of other colicin genes. Alignment of the cea8 promoter region sequence with corresponding sequences from colicin E2, E5 and E6 indicated that these promoters are very similar (S4 Fig) and, thus, similar co-ordinated regulation is perhaps to be expected.
It is clear from S4 Fig that regions of the cka promoter are similar to that of cea8, particularly around the two LexA SOS boxes. As this region was bound differentially by AsnC at cea8 in the presence of L-asparagine, we examined whether purified AsnC could bind in vitro to a radiolabelled cka promoter fragment using EMSA. Results in S5A Fig indicated that AsnC does bind to cka and its DNA binding was modulated by L-asparagine. As this raises the possibility that AsnC could regulate colicin K production in vivo, we measured cka promoter activity after SOS-induction in wild-type, ΔasnC and ΔiscR cells, carrying a cka::lacZ fusion cloned into pRW50. Results in Fig 6E indicate that AsnC had little effect on cka expression, but confirmed that IscR is a major repressor of the cka promoter. In addition, colicin K expression was also examined in our wild-type, ΔasnC and ΔiscR strains, whilst carrying a colicin K-encoding plasmid (S5B and S5C Fig). These experiments again showed that IscR is the major regulator of colicin K expression and that AsnC has little effect on the expression of this pore-forming colicin.
E. coli harbours many promoters that are regulated by multiple transcription factors, each of which ensures that different intra- or extracellular signals are integrated into gene expression [22]. At the DNase colicin E8 regulatory region we identified a large nucleoprotein complex composed of two LexA repressors flanked by at least two AsnC octamers, that likely wrap DNA in a nucleosome-like structure to firmly prevent transcription initiation (Fig 4D–4F).
The AsnC protein belongs to the Lrp/AsnC family of transcriptional regulators that are widely distributed among prokaryotes and affect cellular metabolism, often in response to exogenous amino acids [12,13]. In contrast to other members of the family, which are global regulators and affect a variety of bacterial functions [23], the AsnC protein was thought to be a gene specific regulator, controlling only two genes in E. coli (asnA and asnC) [14]. Here we report a novel role for AsnC, in which the promoters of the nuclease colicins have “recruited” this protein, enabling regulation in response to L-asparagine levels.
Our data show that an amino acid effector modulates AsnC interaction at the colicin E8 promoter, which influences regulation of cea8 expression. We predict that, as for the Neisseria meningitidis AsnC ortholog, [24], L-asparagine binding modulates the stability of a certain protein oligomeric state and also the mode of binding in the E. coli AsnC-cea8-LexA nucleoprotein complex. Furthermore, as expression of asnC is negatively autoregulated and dependent on the Nac protein under nitrogen-limiting conditions [15], nutrient conditions, specifically nitrogen levels, and nitrogen metabolism might coordinate the cea8 expression through altering the amount of AsnC within the cell. Note that nutrients were recently reported to modulate the release of the DNase colicins by modulating the translation efficiency of the colicin E2 lysis gene transcript [25]. Therefore, AsnC appears to couple metabolic signals to the induction of colicin operon components, in order to synchronise accumulation with the release of the colicin.
In our previous work, we showed that the global transcriptional factor, IscR, in response to the nutritional status of the cell, and, co-dependently with LexA, delays induction of the pore-forming colicin genes after SOS induction [11]. This was a surprising finding as E. coli IscR had been thought to be primarily involved with controlling housekeeping iron sulphur cluster biogenesis, anaerobic respiration enzymes and biofilm formation [26,27]. Here our data strongly suggests that temporal induction of DNA and RNA targeting colicins is IscR independent, and show that the key regulator is the AsnC repressor. At the cea8 promoter, AsnC repression seems to reflect L-asparagine levels and presumably serves as an indicator of general amino acid abundance and availability. In contrast, AsnC does not affect the expression of pore-forming colicin K gene expression. Thus, our data imply that the promoters of the nuclease and pore-forming colicins have adopted different transcription regulators to co-ordinately regulate transcription in conjunction with the LexA repressor and distinct metabolic inputs are integrated at these promoters, which both affect the timing and level of colicin induction.
It is clear that colicinogenic plasmids seem to have evolved to exploit transcriptional factors that are of the host origin. We suggest that ubiquitous regulators, present in most E. coli strains were “picked” in order that the colicinogenic plasmids can be swapped between strains [1], with the colicin promoters being silenced in the same manner. In addition, colicin production and subsequent lysis protein driven colicin release causing death of the producing bacteria, may enable eradication of strains that lack or synthesize a non-functional regulator and cannot efficiently respond or adapt to different environmental signals.
The bacterial strains, plasmids, promoter fragments and oligodeoxynucleotide primers used in the present study are listed in S3 Table. The E. coli Keio collection wild-type strain, BW25113, and its derivatives were used throughout the study [16]. To verify the Keio collection deletion strains, a transposon-specific primer Keio1Kn [16] and the gene specific primer (named as gene pre) was used (S3 Table) in 30 cycles of PCR reactions (30 sec 94°C, 30 sec 55°C, 60 sec 72°C). PCR products were analysed on 1.2% agarose gels and stained with ethidium bromide. The colicin D (cda), E2 (cea2), E5 (cea5), E6 (cea6), E7 (cea7) and E8 (cea8) promoter fragments were amplified by PCR from natural colicin encoding plasmids using primers colX_beta_F and colX_beta_R (X denotes the relevant colicin), which introduce flanking EcoRI and HindIII sites (S3 Table). For testing colicin promoter activities, each promoter fragment was cloned into the lac expression vector, pRW50. Plasmid constructs were named as pRW50cxay (x and y denotes each colicin). As a source of DNA fragments for in vitro analysis, EcoRI-HindIII colicin E8 and colicin K promoter fragments were cloned into pSR. To assay colicin synthesis and ensure plasmid selection, the transposon Tn3 (ApR) was inserted into the naturally occurring colicinogenic plasmids of the Pugsley colicin collection [28], harbouring operons for either colicin D, E2, E5, E6, E7 or E8. Strain CL127 carrying Tn3 on the conjugative plasmid pHly152-T8 was used as a donor strain. To generate plasmid pAsnC, for the overexpression of the N-terminal His6 AsnC fusion protein, primers asnC_u and asnC_d were used to PCR amplify the asnC open reading frame and introduce flanking BamHI and MluI restriction sites. Purified PCR product was subsequently cloned into expression vector pET8c (Novagen) to generate pAsnC.
E. coli RNA polymerase holoenzyme harbouring σ70 (RNAP) was obtained from Epicentre Technologies (Madison). The His6-LexA protein was overexpressed and purified as described in [29] and stored in 20 mM Tris (pH 7.3), 200 mM NaCl at -80°C. The His6-IscR protein was overexpressed, purified and its concentration determined as described in [11]. To induce the synthesis of AsnC protein, an overnight culture of E. coli BL21 (DE3)pLysE strain grown on an agar plate, containing ampicilin (100 μg ml-1) and chloramphenicol (25 μg ml-1), harbouring pAsnC was grown to an optical density at 600 nm (OD600) of 0.6 when 0.8 mM isopropyl β-D-1-thiogalactopyranoside (IPTG) was added to the culture. After 4 h of growth the cells were harvested and the N-terminally His6-tagged AsnC was affinity purified by Ni-chelate chromatography (Quiagen) and stored at 4°C in 50 mM NaH2PO4 (pH 8), 300 mM NaCl, 250 mM imidazole. The concentrations of the LexA and AsnC proteins were determined using a NanoDrop 1000 (Thermo Scientific) using the extinction coefficients at 280 nm of 6990 M-1 cm-1 and of 10555 M-1 cm-1, respectively.
The low-copy number lac expression vector, pRW50 [30], was used to measure the activity of the colicin promoters. Plasmids harbouring colicin promoter fragments (S3 Table) were transformed into the relevant strains. Cells were grown aerobically (180 r.p.m.) at 37°C in Lysogeny Broth (LB) supplemented with tetracycline (12.5 μg ml-1). To induce the SOS response, a sub-inhibitory concentration [31], 37 μM, of nalidixic acid (Sigma-Aldrich) was added to the culture when the OD600 reached 0.3. Culture samples were assayed for β-galactosidase activity according to the Miller method [32]. The presented values are the averages of at least three independent experiments and are shown with standard deviations. To measure the L-asparagine effect on cea8 promoter activity, the relevant wild type strain carrying pRW50cea8 was grown to an OD600 ~0.2 in M9 medium [33], containing a low concentration of NH4Cl (0.5 mM) to sustain growth. The bacterial culture was then split and supplemented with either 10 mM NH4Cl or 20 mM L-asparagine. After the addition of 37 μM nalidixic acid, as indicated, samples were taken and analysed as described above.
A biotinylated 179 bp colicin E8 promoter fragment from position -169 to position +10 from the translation start site (TSS) was generated by PCR using primers Pull_FE8 and Pull_RE8 with pColE8-Tn3 as a template. The DNA fragment was purified by GeneJET PCR purification kit (Thermo Scientific). Immobilisation of the biotinylated DNA (50 μg) to 5 mg of M-280 streptavidin Dynabeads (Invitrogen) was carried out in 15 minutes at room temperature as described [11]. An overnight culture of the E. coli BW25113, harbouring the pRW50cea8, was diluted 1: 200 into 0.5 l LB broth supplemented with tetracycline (12.5 μg ml-1) and induced with nalidixic acid (37 μM) once the OD600 had reached 0.3. After 45 min, cells were harvested and cell extract prepared as described [31]. Cleared lysates (~20 ml) were mixed with streptavidin beads with or without cross-linked biotinylated cea8 promoter fragment in 50 ml centrifuge tubes (Costar) and incubated for 10 min with gentle mixing on ice. Dynabeads were collected using a magnet and washed four times in 20 mM Hepes-Na (pH 7.4), 100 mM NaCl, 0.1% (v/v) Tween 20. Proteins were eluted with 500 μl of buffer (20 mM Hepes-Na, 800 mM NaCl, 0.1% (v/v) Tween 20) and concentrated by TCA precipitation. Proteins were resolved on a 12% SDS-PAGE gel (Invitrogen), and visualized by Coomassie blue staining. To identify proteins, nine 1 mm gel slices were excised and analysed by the Functional Genomics, Proteomics and Metabolomics Facility at the University of Birmingham using a Thermo-Finnigan LTW Orbitrap mass spectrometer. Candidate proteins that exhibited DNA binding properties were analysed further.
Colicin synthesis was monitored in the wild-type E. coli BW25113 strain or its ΔasnC derivative JW3721 [16], harbouring one of the colicinogenic plasmids, and was grown aerobically at 37°C in LB broth supplemented with ampicillin. Nalidixic acid (Sigma-Aldrich) was added to the culture at a final concentration of 37 μM, when the OD600 reached 0.3. Samples were taken before induction and at 1, 2 and 3 h after. Cells were diluted to obtain 1 ml samples with an OD600 of 0.3 and the crude cell extracts were prepared by sonication and the cell debris cleared by centrifugation for 1 min at 17000 x g. 100 μl of each extract was injected into wells in an agar plate containing tetracycline (12.5 μg ml-1) overlaid with the lawn of the indicator strain (DH5α harbouring pBR322) as described in [11]. As an alternative approach for colicin determination in the crude cell extracts, 5 μl of a ten-fold or five-fold dilution series of extracts were applied to an agar plate overlaid with the indicator strain as above. Indicator strains were grown at 37°C and the plates photographed using a G:Box (Syngene).
EMSA analysis, using purified LexA, IscR, AsnC and RNAP, with the cea8 and cka promoter regions, was performed as described in [34]. DNA fragments were excised from pSRcea8 or pSRcka using EcoRI and HindIII restriction enzymes and purified promotror fragments were labelled at the HindIII end with [γ-32P]-ATP using polynucleotide kinase (NEB). Approximately 0.5 ng of DNA fragment was incubated with varying amounts of purified proteins, as indicated. The reaction buffer contained 20 mM Hepes (pH 8), 5 mM MgCl2, 50 mM potassium glutamate, 1 mM DTT, 5% (v/v) glycerol and 0.5 mg/ml BSA and the final reaction volume was 10 μl. Where AsnC was used, the EMSA buffer contained 5 mM L-asparagine (L-asn), where indicated. Samples were incubated at 37°C for 15 min before electrophoresis. For competitive EMSA experiments, DNA fragments were first incubated with various concentrations of LexA (for 15 min at 37°C) followed by the addition of RNAP and incubated for another 15 min at 37°C. Herring sperm DNA was included at a concentration of 6.5 μg ml-1 for these experiments. After incubation, all samples were immediately run on a 5% polyacrylamide gel at 12 V cm-1 in 0.25 x TBE, running under tension, and were visualised using a Bio-Rad Molecular Imager FX and Quantity One Software (Bio-Rad).
DNase I footprinting of AsnC and LexA at the cea8 promoter region was performed as described [35], using purified proteins in the presence or absence of 5 mM L-asparagine and a purified EcoRI-HindIII cea8 fragment that had been 32P-end labelled at the HindIII site using polynucleotide kinase and [γ-32P]ATP.
|
10.1371/journal.pgen.1006231 | Linking Core Promoter Classes to Circadian Transcription | Circadian rhythms in transcription are generated by rhythmic abundances and DNA binding activities of transcription factors. Propagation of rhythms to transcriptional initiation involves the core promoter, its chromatin state, and the basal transcription machinery. Here, I characterize core promoters and chromatin states of genes transcribed in a circadian manner in mouse liver and in Drosophila. It is shown that the core promoter is a critical determinant of circadian mRNA expression in both species. A distinct core promoter class, strong circadian promoters (SCPs), is identified in mouse liver but not Drosophila. SCPs are defined by specific core promoter features, and are shown to drive circadian transcriptional activities with both high averages and high amplitudes. Data analysis and mathematical modeling further provided evidence for rhythmic regulation of both polymerase II recruitment and pause release at SCPs. The analysis provides a comprehensive and systematic view of core promoters and their link to circadian mRNA expression in mouse and Drosophila, and thus reveals a crucial role for the core promoter in regulated, dynamic transcription.
| Circadian rhythms switch gene expression on and off with a daily rhythm in most tissues in mammals and other animals. Typically, thousands of genes are affected, and the functions of these rhythms include preparing and adjusting various physiological functions in tissues to meet time-of-day dependent requirements optimally. The controllers of the rhythms are often transcription factors (proteins which regulate transcription), which are relatively well known. However, there is a layer between transcription factor action and transcriptional activity whose role in circadian transcription has not previously been characterized: the core promoter. The core promoter acts as a template for the assembly of the intricate machinery that governs initiation of transcription. There are different types of core promoters that are typically used for different types of genes. It is not known which types of core promoters are used for the rhythmically induced genes. Here, it is shown that there are specific characteristics of core promoters driving rhythmic transcription in mice and flies. Furthermore, it is shown that there is a class of strong circadian promoters in mice that drive strong rhythms with very intense average transcriptional rates. These results help understanding the regulatory systems governing circadian transcription, and ultimately, aid the understanding of many diseases resulting from disturbed circadian rhythms.
| In many metazoans, transcription of numerous genes in most cell types occurs in a rhythmic fashion with a period of ~24 hours, also if the organism is held under constant conditions; these rhythms are termed circadian transcriptional rhythms [1]. They are to a large extent orchestrated by the cellular circadian clock, which consists of connected feedback loops of clock genes that code for clock proteins. In turn, clock proteins repress or activate transcription of themselves or other clock genes, in this way generating the rhythms [2]. When organisms are held under rhythmic conditions, such as 12hr/12hr light-dark cycles, the circadian clocks in their cells become synchronized to these external so-called zeitgebers. Strictly, one then speaks of diurnal rhythms in gene expression, although here we use the term circadian rhythms also for this case.
The cellular circadian clock becomes manifest in an output of rhythmic mRNA expression of typically hundreds or thousands of clock-controlled genes (CCGs) [3]. The CCGs are thought to be controlled to a large extent by transcription factors (TFs) or coregulators with rhythmic abundances that are part of the cellular circadian clock. In mammals, prominent examples are the CLOCK/BMAL1 heterodimer, which binds to E-boxes in promoters of its target genes [4], the nuclear receptors REV-ERB α and β, which bind to ROR elements [5], and DBP and E4BP4, which bind to D-boxes [6,7]. In the fly Drosophila melanogaster, the CLOCK homolog CLK has a similar function, also binding to E-boxes [8]. Rhythmic binding of these core circadian clock TFs (CTFs) and their coregulators to promoters are thought to induce rhythms in transcriptional activities, which ultimately lead to rhythms in mRNA and protein abundances.
Circadian transcriptional rhythms have been investigated by combined analysis of transcript abundances and CTF binding to corresponding promoters. Mouse liver and Drosophila have served as useful model systems for investigating transcriptional rhythms, and by for instance comparing nascent transcript to mature mRNA abundances, the propagation of rhythms from transcriptional activities to transcript abundances can be monitored [9–12]. Analysis of such data has revealed a contribution of post-transcriptional regulation to the generation of rhythms in mature mRNA abundances, but transcriptional rhythms remain the dominant determinant of the rhythmic transcriptome [13,14]. Circadian rhythm generation is correlated with rhythmic binding of CTFs to binding sites in the promoters of genes with circadian mRNA expression: for instance, BMAL1 binding phases (peak time), measured at the promoter level in mouse liver, correlate well with phases of the corresponding transcripts [4]. On the bioinformatics and data analysis side, position weight matrix (PWM) based prediction of TF binding sites at the promoter population level paired with quantitative modeling has led to additional insights into the combinatorial regulation of rhythmic transcription by CTFs and also other circadian TFs [15–17]. However, the layer of mechanistic rhythm propagation between CTF binding and transcriptional rhythms has not been characterized systematically. This layer consists of the core promoter and its chromatin environment, the latter which entails specific nucleosome arrangements and histone tail modifications. There are different classes of core promoters, and a question that has not yet been addressed is whether certain types of core promoters are more suitable than others to propagate rhythmic TF binding to rhythmic transcriptional activities.
Core promoters often contain specific binding sites (core promoter elements) for general transcription factors (GTFs) such as TFIID and TFIIB [18,19]. GTFs together with RNA polymerase II (Pol II) nucleate the pre-initiation complex (PIC), which assembles at the transcription start site (TSS) before transcription can initiate. Different GTFs play different roles in PIC nucleation and initiation of transcription [20]. The TFIID subunit TATA-binding protein (TBP) binds to the TATA box, a core promoter element conserved from archaea to mammals, situated ~30 bp upstream of the TSS. The TFIIB subunit binds to BRE elements, located closely upstream (BREu) or downstream (BREd) of the TATA box, and helps recruiting Pol II. Other GTFs help opening DNA at the promoter to form a so-called transcription bubble, and also phosphorylate Pol II at its C-terminal domain and prime it for transcription. Various combinations of core promoter elements are thought to modulate the propagation of TF binding to transcriptional activation, in this manner helping to specify developmentally regulated or tissue specific transcription [21].
Eukaryotic core promoters are often thought to have an inactive ground state, assured by a tendency for nucleosomes to cover promoter DNA. In this ground state, the TSS is assumed to be covered by a nucleosome which constitutes a certain barrier for Pol II to penetrate [22–24]. Transcriptional activators recruit various chromatin remodeling factors, which facilitate PIC assembly and transcriptional initiation by loosening the nucleosomal barrier. This can happen through e.g. histone acetylation and ATP-dependent nucleosomal displacement, as shown for the human IFN-β promoter [25]. One mode of circadian regulation of transcription could then be rhythmic binding and unbinding of transcription factors, resulting in rhythmic recruitment of histone modifiers and nucleosome remodelers that in turn free up the TSS for PIC formation and transcription initiation in a rhythmic fashion. Such a mode is indeed consistent with several hallmarks of the strongly circadian Dbp transcript, including observed rhythmic CLOCK/BMAL1 binding, histone H3 Lys4 trimethylation (H3K4me3) and Lys9 acetylation modifications, as well as rhythmic gross H3 abundance at the TSS [26]. However, a subset of metazoan promoters appear to have an active ground state, in the sense that they are depleted of nucleosomes and bound by Pol II even when there is no active transcription. A signature of such promoters is a high level of Pol II immediately downstream of the TSS (as detected by ChIP-Seq), compared to the Pol II level in the gene body [27,28]. The TSS to gene body Pol II ratio is referred to as pausing index, since Pol II has often already engaged in initial transcription at these promoters but sits paused ~50 bp downstream of the TSS. Paused Pol II is specifically detectable by techniques such as global run-on sequencing (GRO-Seq) as sharp peaks of nascent mRNA fragments aligning just downstream of the TSS [29]. There are regulated processes promoting such pausing, which involve pausing factors such as DSIF and NELF. Induction of transcription at these TSSs involves pause release factors such as P-TEFb [30–32].
There are several correlated hallmarks of promoters with active ground state, including nucleosome-depleted regions immediately upstream of the TSSs, absence of TATA box, high levels of Pol II and the H3K4me3 mark immediately downstream of the TSS, as well as high levels of the H2A.Z histone variant at the first nucleosome downstream of the TSS [28,33–39]. In mammalian promoters with active ground state, there is the additionally associated hallmark of high observed to expected ratios of CpG dinucleotides around the TSS (hereafter: CpG ratios). The distribution of this ratio computed for promoters is bimodal, motivating a classification of promoters as either having low or high CpG ratios [40,41]. The causal relationships between these hallmarks are debated and not easily teased apart. Probably, causality is cyclical in some cases: for instance, Pol II recruits the H3K4 methylases SET1 and MLL1 at some promoters [42], while H3K4me3 in turn is able to actively recruit the TAF3 subunit of TFIID, a part of the Pol II pre-initiation complex [43]. There are also several other mechanisms that help explain the correlated hallmarks of promoters with active ground state. Non-methylated CpG dinucleotides may recruit the protein CFP1, which in turn is associated with SET1, resulting in increased H3K4me3 levels [44,45]. The H3K4me3 mark may help recruit ATP-dependent chromatin remodelers to influence nucleosome positioning [46]. Certain TF coregulators, nucleosome-depleted regions, and ATP-dependent chromatin remodelers may elevate levels of the H2A.Z variant at the +1 nucleosome [47,48]. Although a debated issue, in vitro and in vivo evidence collected to date suggest CpG dinucleotides help instruct the formation of nucleosome-depleted regions of mammalian promoters [49–51]. Besides this DNA sequence feature, proteins associated with CpG-rich DNA as well as ATP-dependent chromatin remodelers probably help to establish nucleosome-depleted regions in vivo [46,50,52]. In particular, TSS-bound Pol II, perhaps in a paused state, could contribute to establishing nucleosome-depleted regions by sterically hindering nucleosome formation in a competitive fashion. This effect is probably more pronounced at nucleosome-depleted regions at CpG-poor promoters, with their higher propensity for nucleosome coverage [50,53].
Based on these myriad features and mechanisms, two main metazoan promoter classes, type I and type II promoters, have been proposed [54]. Type I promoters drive regulated, tissue-specific mRNA expression, have an inactive ground state with a less pronounced nucleosome-depleted region, are enriched for TATA boxes, and are depleted of CpG dinucleotides in vertebrates [55,56]. These promoters are often focused, which means that they have a well defined TSS, as measured with the CAGE assay [57]. Type II promoters drive ubiquitously expressed transcripts, have more pronounced nucleosome-depleted regions, are enriched for the H3K4me3 mark, CpG dinucleotides, and are depleted of TATA boxes. These promoters are often dispersed, meaning that they have multiple TSSs scattered over regions of tens or hundreds of bp [55,56]. They are also enriched for constitutively transcribed ("housekeeping") genes [41], presumably because their nucleosome-depleted regions are conducive to constitutive transcription.
Both type I and type II promoters may govern inducible transcription, however implemented in different manners, as suggested by two studies of LPS-induced transcription of ~50 genes in mouse macrophages [58,59]. Primary response genes, which are quickly induced without requirement for additional protein synthesis, tend to have type II promoters, and induction tends to occur by induced release of paused Pol II. Secondary response genes, which are more slowly induced with requirement for additional protein synthesis, tend to have type I promoters, require ATP-dependent chromatin remodeling, and to have an inactive basal state without bound Pol II with a nucleosome blocking PIC assembly. Open questions include whether circadian transcription–a particular mode of inducible expression–preferably employs type I or type II promoters, to which degree these rhythms are accompanied by oscillations on the associated chromatin hallmarks at the core promoters, and which mechanisms might be involved in circadian regulation of transcription at the core promoter.
This report describes the core promoters of genes transcribed in a circadian manner (hereafter: circadian promoters) in mouse liver and in Drosophila, by integrating promoter sequence features (CpG ratios and core promoter elements) with a wealth of genome-wide data measuring chromatin state, transcriptional activities, and CTF binding (Fig 1A). It is shown that the core promoter is of greatest importance for circadian transcription. In particular, a core promoter class in mouse combining hallmarks of type I and type II promoters is uncovered. This promoter class, here termed strong circadian promoter, drives circadian transcription with both high amplitudes and high average transcriptional rates.
To investigate whether there is a preferred core promoter architecture for circadian promoters in mammals, a collection of mouse transcripts assignable to unique TSSs and promoters was established (RefSeq transcript annotation, Methods). The transcriptional activities producing these transcripts were estimated by reanalyzing Nascent-Seq data [11]. These data were collected from mouse livers sampled every 4 hrs under 12 hr/12 hr light-dark (LD) conditions (time points hereafter referred to as zeitgeber time, ZT). To avoid confounding effects from lighting conditions, only data obtained from mice held under LD were used in the present study. A group of 1895 promoters with very clear rhythmic transcriptional activities (Methods) were classified as circadian promoters. In contrast, a background group of 5829 promoters with significant transcriptional activities yet without any signs of circadian rhythms were classified as constitutive promoters. A group of 4892 silent promoters without detectable or with very low transcriptional activities was also identified. For rhythmic activities, relative amplitudes (absolute amplitudes divided by means, hereafter: amplitudes) were estimated by harmonic regression (Methods) and used for the further analysis.
To characterize core promoters, areas in the vicinities of the TSSs were scanned for the presence of TATA boxes, BREu, and BREd motifs (Methods and S1A Fig). In addition, CpG ratios were quantified (Methods). Furthermore, MNase-Seq data [60], which quantify nucleosome occupancies, were reanalyzed. A summary statistic for nucleosome occupancy for each promoter was formed by averaging nucleosome coverages between −101 and −1 bp from the TSS. Finally, mouse liver CAGE data, which measure precise TSS usage by transcripts [57], were used to classify promoters as focused or dispersed (Methods). Focused promoters have a well-defined TSS: Most transcripts start within a narrow range of a few bp. Dispersed promoters give rise to transcripts with starting positions varying by tens of bp. These three characteristics–core promoter sequence features and motifs, nucleosome occupancy, and TSS variation–have previously been used to characterize different promoter classes, and were used as a starting point for the present study.
Circadian promoters turned out to be enriched for TATA boxes when compared to constitutive promoters (Fig 2A, Fisher's exact test, p < 5×10−7, odds ratio 1.54), although not as strongly enriched as silent promoters. However, circadian promoters driving transcription with high amplitudes and high average activity were exceptionally strongly enriched for TATA boxes, with around 35% TATA box containing promoters (Fig 2A). Circadian promoters had a slightly less pronounced nucleosome-depleted region immediately upstream of the TSS than constitutive promoters (rank sum test, p < 10−15, median ratio 1.24), but were significantly more nucleosome-depleted than silent promoters (Fig 2B). There was a positive correlation between nucleosome occupancy immediately upstream of the TSS (hereafter: nucleosome occupancy) and transcriptional amplitude (Spearman's rho = 0.28, p < 10−15, Fig 2C). CpG ratios were negatively correlated with transcriptional amplitudes (S1B and S1C Fig). Circadian promoters had overall slightly lower CpG ratios than constitutive promoters (rank sum test, p < 10−13, median ratio 0.93), but silent promoters had much lower CpG ratios than circadian promoters (S1D Fig). In agreement with these correlations and with previous studies [51,59], nucleosome occupancy was negatively correlated with CpG ratio (Spearman's rho = −0.45, p < 10−15 S1E Fig).
Consistent with their enrichment for TATA boxes, circadian promoters also tended to be more focused than constitutive promoters according to the CAGE data (Fisher's exact test, p < 0.002, odds ratio 1.24). Circadian TATA box promoters were furthermore depleted of the BREd motif compared to constitutive TATA box promoters (Fisher's exact test, p = 0.02, odds ratio 0.68), or when comparing circadian to constitutive promoters overall (Fisher's exact test, p < 0.0033, odds ratio 0.85). No enrichment or depletion was detected for the BREu motif in circadian promoters compared with constitutive promoters (Fisher's exact test, p = 0.27, odds ratio 0.93). Thus, the BREd motif appears to favor constitutive mRNA expression.
Taken together, these findings seem to indicate that circadian promoters, especially those that drive transcriptional rhythms with high amplitude, tend to be of the type I class. However, the distribution of nucleosome occupancies for circadian promoters with TATA box was broad and bimodal (Fig 2D). Clearly, a large population of circadian promoters with TATA box had a nucleosome-depleted region upstream of the TSS, a hallmark thought to be more typical of type II promoters, while a group of other circadian promoters had much higher nucleosome occupancies (Fig 2D, arrow). As is the case in human cells [51,61], there was a general negative correlation between nucleosome occupancy and transcriptional activity among all promoters corresponding to expressed transcripts (Spearman's rho = −0.34, p < 10−15, S1F Fig). Consistent with this and the positive correlation between amplitude and nucleosome occupancy, there was among circadian promoters also a negative correlation between mean transcriptional activity and amplitude (Spearman's rho = −0.28, p < 10−15). However, although circadian promoters had higher nucleosome occupancy on average compared to constitutive promoters (Fig 2B), they did not have lower median transcriptional activity (rank sum test, p = 0.13, median ratio 1.03). Thus, circadian promoters might harbor separate promoter classes that are not detectable by averaging measurements over the entire promoter group. In particular, there might be a hidden group of circadian promoters with atypically high transcriptional activities, increasing the population average transcriptional activity of circadian promoters.
To stratify circadian promoters further and uncover possible hidden groups, a reanalysis was performed of mouse liver ChIP-Seq data measuring the genome-wide binding of core circadian clock transcription factors (CTFs) known to potently induce rhythmicity in transcription of CCGs. CTFs considered were REV-ERB α and REV-ERB β (which bind to ROR elements), E4BP4 (which binds to D-boxes) and BMAL1 (which binds to E-boxes), and promoters were scanned ± 3000 bp of the TSS for CTF binding events (Methods). The reason for this stratification was that CTFs levels exhibit strong circadian rhythms in most tissues. Promoters not binding CTFs but still driving circadian transcriptional rhythms in mouse liver are probably controlled by other rhythmic TFs or cofactors that could be induced partly by rhythmic external cues in an inducible, tissue-specific manner [16]. Thus, core promoter architecture could differ between CTF and non-CTF binding promoters. Finally, based on their bimodal CpG ratio distribution, promoters were classified as high CpG ratio (HCpG) or low CpG ratio (LCpG) promoters (Fig 2D and Methods).
As expected, circadian promoters were enriched for CTFs compared to constitutive promoters (Fisher's exact test, p < 10−10, odds ratio 1.44). Further, CTF-binding circadian promoters (n = 900) were enriched for TATA boxes, compared to circadian promoters not binding CTFs (n = 995, Fisher's exact test, p = 0.0017, odds ratio 1.56). Unexpectedly, however, CTF binding alone was not associated with high transcriptional amplitude (Fig 3A, left, rank sum test, p = 0.57, median ratio 0.99). Rather, TATA boxes and LCpG were associated with high transcriptional amplitudes, regardless of CTF binding (rank sum test, p < 10−15, median ratio 1.4). In contrast, CTF binding was strongly associated with high average transcriptional activities (Fig 3A, right, rank sum test, p < 10−15, median ratio 2.4), independently of TATA boxes or LCpG. These effects were additive: Promoters with both CTF binding and a TATA box or LCpG were associated with very strong transcription, with high amplitudes as well. Note that the scale is logarithmic and that the fold differences in mean transcriptional activities are quite large. The same effects were observed when stratifying for TATA boxes or CpG content separately (S2A Fig).
CTF-binding circadian promoters with either LCpG or TATA box are hereafter referred to as strong circadian promoters (SCPs, listed in S1 Table), in light of the marked association of these sequence features with both strongly rhythmic and high transcriptional activities.
To investigate and characterize SCPs further, a search for distinguishing characteristics of these promoters in terms of nucleosome occupancy and sequence features was carried out. By stratifying all circadian promoters with LCpG or TATA box into either SCPs or non-CTF binding promoters (Fig 1B), the broad bimodal nucleosome occupancy distribution (Fig 2F) was resolved into the population of SCPs, which had low nucleosome occupancy, and the population of non-CTF binding LCpG or TATA box containing circadian promoters, which had high nucleosome occupancy (Fig 3B). In stark contrast to the general anti-correlation between promoter nucleosome occupancy and CpG ratio, SCPs turned out combine low average CpG ratios with low nucleosome occupancies (S2B Fig). SCPs had, in fact, almost as low nucleosome occupancy as constitutive TATA-less HCpG promoters without CTF peaks (hereafter: constitutive type II; Figs 1B and 3C). Thus, SCPs appeared to escape both the negative correlation between amplitude and mean transcriptional activity, as well as between amplitude and nucleosome depletion (Fig 2C). The group of circadian promoters that do not bind CTFs, but with LCpG or TATA box, on the other hand, had characteristics of canonical type I promoters [54]: low median transcriptional activities (Fig 3A) and high nucleosome coverage (Fig 3C). Promoters in this group are hereafter referred to as circadian type I promoters (Fig 1B).
Circadian CTF binding HCpG promoters without TATA box, on the other hand, had the low nucleosome occupancies typical of type II promoters (Fig 3C), which given the initial analysis here were consistent with their low oscillation amplitudes (Fig 3A). These promoters are hereafter called circadian type II promoters (Fig 1B).
For comparison to earlier results on correlations between core promoter properties and nucleosome occupancies [55,56], constitutive promoters with LCpG or TATA box (but without CTF peaks, to exclude residual circadian promoters whose transcriptional rhythms might not have been detected), were also investigated. These promoters (hereafter: constitutive type I promoters, Fig 1B) had the expected high nucleosome occupancies upstream of the TSS, whereas constitutive type II promoters tended to have the familiar pronounced nucleosome-depleted region upstream of the TSS (Fig 3C).
One possibility would be that SCPs simply constitute regular type I promoters that happen to be very strongly transcribed. To exclude this, constitutive type I promoters were sampled (with replacement) to match the distribution of average nascent mRNA read counts of SCPs. This resulted in an artificial group of constitutive TATA box promoters with the same statistical distribution of transcriptional activities as SCPs. A bootstrapping test procedure was then employed to probe whether nucleosome occupancies of the matched constitutive type I promoter population are similar or different to those of SCPs (Methods). This analysis (S2 Table) showed that SCPs indeed have significantly lower nucleosome occupancy than constitutive type I promoters with matched mean transcriptional levels.
Could the low nucleosome occupancy of SCPs be due to BMAL1 binding, since this CTF is known to act as a pioneering factor and induce eviction of nucleosomes [60]? A reanalysis of MNase-Seq data from mouse livers of Bmal1−/− animals [60] yielded nucleosome profiles very similar to those of livers from wild type mice (S2C Fig). This points to other reasons than BMAL1 for the nucleosome depletion immediately upstream to the TSS of SCPs.
Does daily up- and downregulation of transcription require mechanisms associated with focused transcription initiation? As mentioned, regulated and CpG-poor TATA box containing promoters are often focused. However, promoters with an ordered nucleosome pattern and a pronounced nucleosome-depleted region upstream of the TSS are generally more dispersed [54]. This relationship between nucleosome coverage and TSS width could be reproduced with the present data set: dispersed promoters had significantly lower nucleosome occupancy (rank sum test, p < 10−15, median ratio 0.81). Thus, since SCPs are CpG-poor and enriched for TATA boxes, but on the other hand have low nucleosome occupancies, it is not obvious if SCPs should tend to be focused or dispersed: the general rules are in conflict. An analysis showed that SCPs with TATA box were mainly focused with a smaller dispersed subpopulation, but that TATA-less SCPs (consequently with LCpG) tended to be much more dispersed (S2D Fig). Thus, strong circadian transcriptional rhythms appear compatible both with focused and dispersed transcriptional initiation.
Since differentiated cells often do not employ TFIID for the majority of transcripts, it is of interest to assess whether SCPs, circadian promoters, or even expressed transcripts in general are associated with this PIC subunit in the adult mouse liver. To do this, the lists of expressed promoters, circadian promoters, and SCPs were cross-referenced with a list of transcripts that are down-regulated in mice with a liver-specific deletion of the TAF10 subunit of TFIID, without which TFIID dissociates [62].
There was a remarkable enrichment for such TFIID-dependent transcripts among transcripts with promoters of expressed transcripts binding CTFs, especially CTF-binding promoters with TATA box. In fact, around 89% of all 130 TFIID-dependent expressed transcripts had CTF binding promoters (S2E Fig), compared to 41% for non-TFIID-dependent expressed transcripts (neither up-regulated nor down-regulated upon TAF10 depletion, Fisher's exact test, p < 10−15, odds ratio 12.2). Conversely, focusing on circadian promoters, CTF binding in conjunction with LCpG was associated with TFIID-dependence. 12% of LCpG SCPs were strongly TFIID-dependent, compared to 0.03% for circadian promoters with neither CTF binding nor TATA box or LCpG (Fig 3D). A possible caveat might be that the observed down-regulation thought to be due to TFIID dissociation in fact was due to circadian transcription: samples were not necessarily collected at the same times of day. However, then an overrepresentation of transcripts up-regulated upon TAF10 depletion would also be expected among CTF binding promoters if all phases are roughly equally probable. Such an overrepresentation was not found among the 119 expressed transcripts up-regulated upon TAF10 depletion (48% vs. 41%, Fisher's exact test, p = 0.11, odds ratio 1.35). This suggests that the down-regulation likely indeed is due to TFIID dissociation. These results suggest that CTF binding promoters, in particular those with LCpG, tend to employ TFIID also in fully differentiated hepatocytes to a higher degree than other promoters.
The SCPs appeared to drive extraordinarily high transcriptional activities and amplitudes (Fig 3A). Is this reflected in mature transcript abundances as well? By analyzing mouse liver mature mRNA abundances from RNA-Seq measurements over 2 circadian cycles [11], matching these to the Nascent-Seq transcripts, and estimating mean mRNA expression levels and oscillation amplitudes (Methods), this question could be answered. As shown in Fig 3E, the extraordinarily high mean expression levels and amplitudes of SCPs were indeed carried over to the mature transcript levels. These results were confirmed using an alternative microarray-based data set to quantify mature mRNA abundances [63] (Methods and S2F Fig).
Short mean half lives are an absolute requirement for high circadian rhythm amplitudes in any molecular species [13]. Short half lives tend to decrease mean abundances, which could be counteracted by high transcriptional activities in the case of mRNAs. Cross-referencing the data on mRNA amplitudes with data on mRNA half lives [64,65] showed that mRNAs driven by SCPs with TATA box have extraordinarily short half lives compared to any other class of mRNAs, including constitutively expressed mRNAs driven by either TATA box or TATA-less promoters (Fig 3E). The high mRNA abundances associated with TATA box SCPs are thus reached in spite of short transcript half lives, presumably through their exceptionally high transcriptional activities.
In summary, the results suggest that SCPs might constitute a separate core promoter class. SCPs are characterized by being much more strongly transcribed and by having much lower nucleosome occupancies than standard type I promoters, yet they retain low CpG ratios.
What may cause the nucleosome depletion upstream of the TSSs of SCPs? It might be the case that promoter-proximal paused Pol II, or associated factors, hinder nucleosome assembly at these promoters [53]. To follow up on this hypothesis here, ChIP-Seq data on Pol II in mouse liver sampled at 7 time points over 24 hours [10] were reanalyzed. Sequencing reads were compiled both for the promoter-proximal pausing region (20–100 bp downstream of the TSS) as well as for the gene body (300–1300 bp downstream of the TSS), which yielded normalized RPKM values for these regions, reflecting the amount of bound Pol II, or Pol II coverage (Methods). This also allowed calculating the pausing index (PI), which is the Pol II coverage in the promoter-proximal pausing region normalized by the gene body Pol II coverage. The PI reflects the tendency of Pol II to remain associated with the TSS for a given level of transcriptional activity. To verify that this is accompanied by truly paused Pol II, the analysis was complemented with GRO-Seq data sampled at 8 time points over 24 hours in duplicates [6], in order to detect peaks of short transcripts produced by paused Pol II.
Indeed, SCPs had higher promoter-proximal Pol II coverage than any other promoter group, including circadian or constitutive type II promoters, even though the latter two promoter groups had slightly lower nucleosome occupancies as SCPs (Fig 4A). These high Pol II levels at SCPs were accompanied by an even stronger peak of GRO-Seq reads (S3A Fig), suggesting that Pol II sits paused but engaged in transcription at these promoters. The bootstrapping test procedure for comparing SCPs to transcriptional activity-matched constitutive type I promoters (above and Methods) showed that this high Pol II coverage is indeed much more common among SCPs as compared to highly expressed type I promoters (p < 10−6).
Although there also was a general negative correlation between nucleosome occupancy and paused Pol II, SCPs consistently displayed atypically high levels of paused Pol II for a given nucleosome occupancy (S3B Fig), but not atypically high PIs (S3C Fig). The lower PIs may be attributed to SCPs being associated with higher Pol II gene body levels when compared to other circadian promoters (rank sum test, p < 10−15, median ratio 2.3, Fig 4A). This is consistent with the idea that especially SCPs might employ Pol II to maintain a low nucleosome occupancy. It also suggests that Pol II sits mainly in a paused state at SCPs, ready to initiate transcription at appropriate time windows.
One could thus ask whether there are widespread oscillations in the competition between a nucleosome-occupied state and a Pol II-bound nucleosome-depleted state at circadian promoters, and at SCPs in particular. If so, rhythms in nucleosome densities and Pol II levels should be observable in MNase-Seq and Pol II ChIP-Seq data. Since both the Pol II ChIP-Seq data [10] and the MNase-Seq data [60] analyzed here were sampled over the course of one day at 7 and 6 time points, respectively, these data allow a glimpse into the dynamics of nucleosome occupancies and Pol II levels. The sequencing depth of the MNase-Seq data is too low to allow reliable quantification at the single-promoter level; they are, however, readily analyzed at the promoter population level. For this, promoters were binned according to their transcriptional phase as determined by analysis of the Nascent-Seq data (Methods), plus/minus 2 hrs. The MNase-Seq data were reordered temporally so that for each promoter, time point 0 corresponded to the transcriptional phase of the promoter's bin. The same exercise was performed with the Pol II ChIP-Seq data, in order to compare the results properly. Visualizing the resulting 6 phase bins revealed a distinctly static nucleosome landscape around the TSSs of SCPs (Fig 4B, p = 0.52, harmonic regression test for rhythmicity). On the other hand, Pol II levels, analyzed in the same manner, revealed a clear rhythmicity that coincided with the transcriptional phase (Fig 4B, p = 0.01, harmonic regression test for rhythmicity). The lack of nucleosome occupancy rhythms was pervasive also at circadian promoters other than SCPs, and the promoter-proximal Pol II levels had weaker rhythms at these promoters (S3D Fig). Hence, the present analysis showed no evidence that nucleosomal dislocation is commonly employed for generating transcriptional circadian rhythms. However, at SCPs, rhythmic promoter-proximal Pol II levels may reflect an involvement in rhythmic transcriptional regulation. These Pol II rhythms had a very high base line, which might serve to keep these promoters permanently nucleosome-depleted, enabling a high average rate of transcriptional initiation.
How may the strong tendency for promoter-proximal Pol II accumulation be achieved and combined with generation of transcriptional rhythms? To investigate this from a biochemical kinetics perspective, a simple but general mathematical model formulated previously [10] was considered. The model describes Pol II recruitment, transcriptional initiation, and release of Pol II into productive elongation, allowing rhythmicity both in recruitment and release rates (Fig 4C). A rigorous analytical and parameter-independent reanalysis of the model (S1 and S2 Texts, S1 Interactive Text) showed that two scenarios–a higher average Pol II recruitment rate, or a lowered release rate–could both explain the increased Pol II promoter-proximal pausing at SCPs compared to e.g. circadian type II promoters. The lower PI could in both these cases be explained by additionally assuming decreased transcriptional elongation speeds causing higher Pol II density (note that not transcriptional speeds, but rather transcriptional initiation rates determine how many transcripts are produced per time unit [66]). The first scenario, higher average Pol II recruitment rates, in theory generally comes with weaker combined net propagation of rhythms in Pol II recruitment and Pol II release rates to transcriptional activities (Fig 4C, S2 Text, S1 Interactive Text). Checking this prediction against data indeed revealed a negative correlation between PI and transcriptional amplitude (Spearman's rho = −0.30, p < 10−15). However, the assumption of only higher average Pol II recruitment rates for SCPs is not an attractive explanation for their high Pol II signals in the promoter-proximal regions, since SCPs drive strong rhythms. Instead, the situation for SCPs is better explained by invoking the alternative scenario (Fig 4C): Additionally assuming lowered average Pol II release rates also help explain the higher promoter-proximal Pol II signals of SCPs. This would be consistent with a tendency at SCPs for shifting of rhythmic control from Pol II recruitment to Pol II release, with retained overall rhythm propagation, that could be inferred from the data at hand (S1 and S2 Texts, S1 Interactive Text, S3E and S3F Fig).
If this is true, additionally a higher Pol II initiation rate (from recruited to paused Pol II) must be assumed to explain the high transcriptional activities of SCPs (S1 and S2 Texts, S1 Interactive Text). Such high initiation rates have indeed been found to be prevalent for TATA box promoters, partly due to a transcription initiation scaffold being constitutively bound to TATA box promoters of active genes [67]. A combination of much higher initiation rates and lower release rates would be consistent with all data: SCPs having higher promoter-proximal Pol II signal, lower PI, higher transcriptional rates, and stronger rhythms.
How might the predicted lower Pol II release rate be achieved? A possible mechanism could be depletion of the histone variant H2A.Z from the first (+1) nucleosome downstream of the TSS. Incorporation of the histone variant H2A.Z is recognized as a general route to transcriptional activation via various mechanisms [48]. In general, nucleosomes with the H2A.Z variant are more unstable, and the variant is associated with nucleosome-depleted regions [68,69], although the causal relationships behind this association are unclear [48]. Certain studies have also established a negative correlation between paused Pol II occupancy and H2A.Z. Possibly, H2A.Z at the first (+1) nucleosome downstream of the TSS facilitates the elongation of Pol II through the +1 nucleosome position [70]. If this is true, and if the average Pol II release rates of SCPs are indeed lower than for circadian type II promoters, then SCPs would have lower H2A.Z levels. If H2A.Z is associated with transcriptional activation, periodic up- and down-regulation might be incompatible with high H2A.Z levels. Hence, another question of more general nature is whether circadian transcriptional regulation is associated with H2A.Z depletion.
To address this, mouse liver H2A.Z ChIP-Seq data were reanalyzed. For all mouse promoters, the H2A.Z signal around the +1 nucleosome (determined from the MNase-Seq data with a peak detection algorithm, Methods) was computed, and normalized to the MNase-Seq based nucleosome signals. Indeed, there was a negative correlation between H2A.Z signal and levels of promoter-proximal paused Pol II for promoters with marked levels of paused Pol II (S3G Fig). In particular, SCPs were characterized by low H2A.Z levels and correspondingly high promoter-proximal paused Pol II levels. As was the case for Pol II coverage, the bootstrapping test procedure (Methods) showed that SCPs also have lower H2A.Z levels than highly expressed type I promoters (p < 10−6).
This is consistent with the model prediction that SCPs have lower Pol II release rates, given the notion that H2A.Z facilitates transcription through the +1 nucleosome. Depletion of H2A.Z might thus enable tighter regulation of Pol II release, exploited at promoters of genes transcribed in a strongly circadian manner. In line with the idea that H2A.Z might be more associated with constitutive transcription, there was a negative correlation in general between circadian transcriptional amplitudes and H2A.Z levels (Fig 4D, Spearman's rho = −0.28, p < 10−15).
The H3K4me3 modification is strongly associated with active transcription and has been implicated in generation of circadian rhythms in certain transcripts [26,71]. H3K4me3 is furthermore associated with high levels of Pol II in the promoter-proximal region [28], and might thus be a hallmark of SCPs. Inconsistent with this, however, is an observed positive correlation between H3K4me3 and H2A.Z levels [68]. Thus, there are reasons both to expect high as well as moderate H3K4me3 levels at SCPs. H3K4me3 levels of mouse liver promoters at different ZT have been analyzed by ChIP-Seq [10]. Although H3K4me3 levels oscillate at the promoters of many genes transcribed in a circadian fashion, the oscillations generally lag the transcriptional rhythms by several hours. Thus, rhythms in H3K4me3 levels near promoters might be consequences rather than causes of rhythmic transcription. This situation could, however, vary between promoter classes, although this possibility has yet not been investigated.
In order to clarify the role of H3K4me3 at SCPs and other promoter classes, the H3K4me3 ChIP-Seq data [10] were reanalyzed and quantified at different locations of the mouse promoters. This analysis showed that SCPs had significantly lower levels of H3K4me3 just downstream of the TSSs than circadian type II CTF binding promoters (Fig 4E, rank sum test, p = 3.8×10−7, median ratio 0.67). The lower H3K4me3 levels of SCPs apparently occur despite their higher promoter-proximal Pol II coverages: Normalizing each promoter-proximal Pol II level to its TSS H3K4me3 level (hereafter: P/M ratio) likewise resulted in significantly higher P/M ratios for SCPs compared to circadian type II CTF binding promoters (rank sum test, p < 10−15, median ratio 2.9). Again, the bootstrapping test procedure (Methods) verified that the TSS H3K4me3 levels and P/M ratios of SCPs indeed are significantly lower and higher, respectively, than for highly expressed type I promoters (p < 10−6 in both cases). This relationship apparently runs counter to the previously observed positive association between the H3K4me3 mark and promoter-proximal Pol II levels [28], which was reproduced with the present data set (S3H Fig). For a given promoter-proximal Pol II level, SCPs had atypically low H3K4me3 levels at the TSS. This does not mean that SCPs had very low H3K4me3 levels; in fact they were intermediate to high (Figs 4E and S3H), but significantly lower than expected given their extraordinarily strong promoter-proximal Pol II signals. Particularly notable, however, was the marked extension of the H3K4me3 mark far into the gene bodies of SCPs (Fig 4E), so that from a position of +500 bp of the TSS and onwards, genes with SCPs had the highest gene body H3K4me3 level of any group of genes, which was also verified by the bootstrapping test against expression-matched constitutive type I promoters (p < 10−6). There was also a general negative correlation between average TSS H3K4me3 levels and circadian transcriptional amplitudes (Spearman's rho = −0.37, p < 10−15), and consistent with this and the results regarding H2A.Z and amplitudes outlined above, the previously observed correlation between H2A.Z and H3K4me3 levels [68] was notable also in the present data set (S3I Fig). This suggests that although SCPs had intermediate to relatively high H3K4me3 levels, extremely high levels of the H3K4me3 mark do not go hand in hand with strong circadian rhythms. Previously, high H3K4me3 levels have been suggested to co-occur almost exclusively with high CpG ratios [45,54]. Although this correlation was evident also in the present data, SCPs constituted a notable exception, combining intermediate to relatively high H3K4me3 levels with low CpG ratios (S3J Fig).
A possible interpretation of these observations is that the observed intermediate H3K4me3 levels of SCPs reflect Pol II-mediated H3K4 trimethylation, pushing H3K4me3 levels up to intermediate rather than low levels. This intermediate level might be low enough to avoid a constitutively permissive chromatin state that would weaken rhythmic regulation. A non-causal role of H3K4me3 in circadian transcriptional activities has also been proposed to account for the later phases of rhythmic H3K4me3 levels with respect to transcriptional activities [10]. This would be consistent too with the high H3K4me3 levels in the gene body, since transcriptional elongation speed in the present analysis was predicted to be slower for genes with SCP promoters (see above), which could lead to the observed higher gene body Pol II signals. Indeed, the later phases of H3K4me3 signal compared to transcriptional activities were readily reproducible (S3K Fig) for CTF-binding promoters (including SCPs). However, this analysis revealed that rhythmic H3K4me3 levels of non-CTF binding circadian promoters tended to have earlier phases than those of transcriptional activities (S3K Fig). This opens up the possibility that H3K4 trimethylation plays a causal role in rhythmic transcriptional activation mediated by other TFs than CTFs.
The results reported so far established that mouse liver circadian promoters are strongly enriched for TATA box promoters, and further uncovered a special class of promoters: SCPs. These circadian promoter traits could be specific for mammals, or they could represent more universal constraints on core promoters for circadian regulation of transcription. A comparison with an evolutionary distant animal may throw light on the universality of circadian core promoters.
Drosophila is besides mouse the model organism where the mRNA expression output pathways of the circadian clock have been best studied. Data from various studies of the Drosophila clock were compiled, resulting in a data set that, bar tissue-specificity, almost rivaled the data compiled for mouse liver. Time-resolved nascent and mature RNA-sequencing [12], and ChIP-Seq measurements for CLK binding sites [8] of fly head mRNA samples, nucleosome occupancies (MNase-Seq as well as H2A.Z ChIP-Seq), and transcribing Pol II positions in S2 cells (3' end nascent transcripts-sequencing, 3'NT-Seq) [70] as well as CAGE data on TSS usage in fly embryos [72] were compiled, using the RefSeq Drosophila transcript annotation (Methods). As was the case for mouse liver, Drosophila circadian transcripts turned out to be enriched for TATA boxes, compared to constitutive transcripts, especially transcripts with high amplitudes (Fig 5A, p = 0.035 and 0.00061, respectively, Fisher's exact test). Moreover, the presence of a TATA box positively correlated with amplitude but not mean mRNA expression levels, whereas CLK binding to the promoter region correlated with mean mRNA expression level but not amplitude (Fig 5B). These relationships are analogous to what was found for TATA box and CTF binding, respectively, in mouse liver. Reinforcing the analogy to mouse TATA box SCPs, circadian promoters were overall depleted of broad TSSs (Fisher's exact test, p = 0.0011, odds ratio 0.56). As for mouse liver, there was also a positive correlation between circadian amplitude and nucleosome occupancy immediately upstream of the TSS, as well as a negative correlation between transcriptional amplitude and H2A.Z content at the first nucleosome downstream of the TSS (Fig 5C, Spearman's rho = 0.44, p < 10−12 for nucleosome occupancy, Spearman's rho = −0.33, p = 1.2×10−7 for H2A.Z).
However, there were only 2 promoters both containing a TATA box as well as binding CLK. In fact, there was no evidence of a promoter class corresponding to mouse SCPs. Circadian promoters in Drosophila generally had high nucleosome occupancies (except CLK-binding promoters, Fig 5D), even circadian TATA box promoters driving strong average transcriptional rates (S4A Fig). Further, Drosophila circadian promoters driving high transcriptional amplitudes were depleted of paused Pol II compared to circadian promoters driving low transcriptional amplitudes (rank sum test, p = 8.8×10−9, median ratio 0.25). There was a negative correlation between paused Pol II and nucleosome occupancy immediately upstream of the TSSs of expressed transcripts (Spearman's rho = −0.20, p < 10−15), as previously described [30], which implicates competition between Pol II and nucleosome occupancy. However, there was no apparent SCP-like subpopulation of promoters with high amounts of paused Pol II and low nucleosome occupancy, at the same time also driving transcriptional rhythms with both high amplitudes and mean levels (S4A–S4D Fig).
These results suggest that the following traits: 1. high nucleosome occupancies, 2. low Pol II pausing levels, 3. depletion of the H2A.Z variant downstream of the TSS, and 4. a TATA box-based promoter architecture, are universally correlated to circadian regulation of transcription, albeit with moderate average mRNA expression levels. The evolution of SCPs–TATA box promoters with high Pol II pausing levels and low nucleosome occupancies that drive rhythms with both high averages and amplitudes–seems to not have occurred in Drosophila.
The results presented here show that the core promoter and its chromatin state are fundamental determinants of circadian transcriptional rhythms. This should motivate a broadening of focus for circadian transcriptional regulation studies to routinely cover aspects of the core promoter. Such efforts could also initially provide validation of the results of the present study. Further, when studying clock output mechanisms in cells or tissues, deletions of CTF binding sites, or over-expression of core clock genes, may abolish rhythmicity in expression of clock controlled genes, but also strongly influence average levels of gene product [26]. This makes it is difficult to discern whether observed phenotypes are due to loss of rhythmicity, or due altered average levels. By targeting the core promoter, e.g. the TATA box, by gene editing techniques [73,74], it might be possible to more precisely attenuate rhythmicity while keeping average levels constant.
Two major classes of circadian promoters emerged in the analysis. The most peculiar of these represented strong circadian promoters (SCPs), driving circadian transcription with both high amplitudes and high average rates. SCPs appear to represent a class of promoters that has not previously been characterized, combining traits of type I and type II core promoter classes as commonly defined [54], and combining circadian regulation of Pol II recruitment and pause release. With regard to chromatin state, a defining signature of SCPs was a high paused Pol II level combined with a low H3K4me3 level relative to the high Pol II level. On the other hand, SCPs had H3K4me3 levels extending further into the gene bodies than other promoters. Finally, although SCPs had low CpG ratios, they had marked nucleosome-depleted regions upstream of the TSSs. The second major circadian promoter class consists of type I promoters, which exhibited moderate average transcriptional activities and high nucleosome occupancies upstream of the TSSs.
The important role played by the core promoter may help explain why only some CTF binding promoters drive circadian transcription. The E-box is strongly associated with the generation of circadian rhythms in mammals and Drosophila since it provides a binding platform for circadian TFs CLOCK/BMAL1 and CLK/CYC, respectively. However, the E-box was originally discovered in the context of constitutive or transient activation of transcription, affecting processes involved in differentiation and development [75]. Thus, the E-box is presumably also implicated in non-rhythmic activation of target gene transcription. In the same way, ROR elements, which are the DNA recognition sites for nuclear receptors REV-ERB α and β, probably are able to bind other, non-rhythmic TFs of the nuclear receptor family, directing transient or constitutive rather than rhythmic transcription [76,77]. What decides whether a given CTF recognition site and promoter/transcript are associated with actual CTF binding and rhythmic transcription? The present results suggest that sequence features of the core promoter, such as TATA boxes, provide one piece of this puzzle. In particular, among circadian promoters, TATA boxes and low CpG content were more predictive of strong rhythms than CTF binding per se. Interestingly, a recent study found rather limited effects of BMAL1 deletion on CCG transcriptional rhythmicity in mouse liver [78].
What properties of the TATA box make it suitable for circadian regulation of transcription? TATA boxes can facilitate the assembly of a transcription reinitiation scaffold, consisting of TFIIA, TFIID, TFIIE, TFIIH, and Mediator, which remains on the promoter after Pol II promoter escape and makes rapid reassembly of the PIC possible [67,79]. Such reinitiation may not be unconditional, since the scaffold does not contain TFIIB and TFIIF, and since it is stabilized by TFs [79,80]. This scaffold mechanism involves TFIID, and although differentiated tissues in mammals do not always rely on TFIID for PIC formation [20,81], there is a subgroup of promoters in mouse liver that does show a strong dependence on TFIID for transcription [62]. Remarkably, CTF binding promoters and SCPs in particular accounted for around 89% of these TFIID-dependent promoters (S2E Fig). Thus, TATA box promoters and PIC scaffolds may constitute a suitable platform for strong but rhythmically gated transcription: in this scenario, circadian activators or repressors recruit TFIIB and Pol II in a rhythmic fashion to permanently assembled scaffolds.
The TFIIB-binding element BREd was underrepresented among circadian promoters. This element might thus rather be conducive to constitutive basal levels of TFIIB binding to promoters, especially in the scaffold scenario. TFIIB has been implicated in stimulating release of paused Pol II in via interactions with TFIIF [82–84], and constitutive TFIIB presence at the promoter may thus also dampen rhythms in release of paused Pol II. TFIIB binding might, in fact, be linked to high H3K4me3 levels [85]. Moderate H3K4me3 levels in the vicinity of the TSSs were indeed another hallmark of circadian promoters with high amplitudes (SCPs and circadian type I promoters). For SCPs, the levels were intermediate on an absolute scale but low given the high Pol II levels: high promoter-proximal Pol II levels and the H3K4me3 are generally associated [28,86]. There is considerable uncertainty as to the mechanistic causality relationships between the H3K4me3 mark and transcriptional activities [87]. Pol II can attract histone methylases leading to H3K4 trimethylation [42], but clearly, since SCPs had higher promoter-proximal Pol II than any other promoter class but lower H3K4me3 levels than circadian or constitutive type II promoters, other mechanisms must be at work. SCPs are CpG-poor, which may be a reason for their relative H3K4me3 depletion, since CpG dinucleotides may induce H3K4 trimethylation via CFP1 [45].
Why are promoters driving high amplitude circadian transcription (SCPs and circadian type I promoters) depleted of H3K4me3, relatively seen? Perhaps high levels of these nucleosome features should be viewed as hallmarks of high constitutive (or perhaps poised) transcription [32]. Circadian promoters need to not only rhythmically induce transcription, but also repress it, and possibly, slow kinetics of H3K4me3 demethylation precludes timely clearance of this histone mark at some promoters, so that they cannot drive circadian rhythms [10,58]. On the other hand, this might not be the case universally, since circadian type I promoters exhibited rhythmic H3K4me3 levels with earlier phases than transcription (S3K Fig), possibly indicating a causal role for this histone mark, as has also been demonstrated for certain genes [71]. Such a causal role may involve H3K4me3-mediated TFIID recruitment [43]. Assuming this scenario, a prediction is that H3K4me3 demethylation is only slow or inefficient enough to preclude circadian regulation when Pol II levels are constitutively high, perhaps persistently inducing H3K4 (re-)methylation. Finally, the extension of the H3K4me3 mark into the gene bodies of SCPs could be a by-product of high Pol II densities due to high rates of transcription initiation. This would also be in line with recent observations made by other investigators [88].
As was the case for the H3K4me3 mark, high levels of the H2A.Z nucleosome variant were also associated with lower amplitudes. These two observations are probably related, since the degree of correlation between the H3K4me3 mark and H2A.Z levels was notable, which corroborates earlier findings [38,68]. The association of H2A.Z with low amplitudes entails the prediction that regulation of H2A.Z levels at the time scale of circadian rhythms in general is not fast or reliable enough to be exploited for rhythm generation.
This may be the case for circadian nucleosomal occupancy dynamics in general. There are certainly cases where induced transcription involves nucleosome rearrangement [26,59,89]. However, the MNase-Seq data analyzed here provided no evidence at the population level of widespread rhythmic nucleosome occupancies at the promoters of circadian transcripts (sequencing depth was not high enough to reliably assess rhythms in nucleosome occupancy at single promoters). Absence of evidence is not evidence of absence, but notable is that rhythmic nucleosome occupancies at the population level was readily detectable at BMAL1 binding sites [60]. Thus, the data suggest that circadian core promoters have a relatively fixed nucleosome occupancy determined by factors not varying at the circadian time scale. This does not mean that nucleosome occupancy has nothing to do with transcriptional activities: nucleosome occupancies were negatively correlated to average transcriptional activities in a continuous fashion (S1F Fig). Such correlations are also present in yeast [90] and mouse and human embryonic stem cells [91].
Importantly, observed averages in tissue samples translate to probabilities of nucleosome occupancy in the single cell. There, nucleosomal occupancy is a dynamic phenomenon, where nucleosome sliding, binding and unbinding continuously takes place [92–94]. It thus makes sense that transcriptional activity varies continuously with nucleosome occupancy as observed at the tissue level, rather than being an on/off phenomenon. The average nucleosome occupancies may be set by a combination of sequence-determined nucleosome forming potential and auxiliary proteins [46,52]. However, these features do apparently not represent activities that vary in a circadian manner.
SCPs and circadian type I promoters have, by definition, low CpG ratios on average. This sequence feature is often considered to increase the nucleosome forming potential [50]. However, only circadian type I promoters had high nucleosome occupancies, those of SCPs were considerably lower (Fig 3C). The high average Pol II levels in the promoter-proximal regions may instead explain the low nucleosome occupancies, assuming competition between Pol II and nucleosome occupancy [46,53]. It is also possible that stable PIC scaffolds at the TATA boxes may compete with nucleosome formation. Thus, even though the low nucleosome occupancies of SCPs may appear to run contrary to the general association between high amplitudes and nucleosome occupancy (Fig 2C), their nucleosome forming potential is probably still high. It is notable that a high nucleosome forming potential and absence of permissive nucleosome features (H3K4me3, H2A.Z) were all associated with high circadian amplitudes. It appears that promoters with an active ground state are not generally employed to drive circadian transcription.
The establishment of a nucleosome-depleted chromatin state is apparently not widespread for Drosophila circadian promoters, where no counterpart of mouse SCPs was found. This could mean that SCPs are a relatively late evolutionary invention, and that CCGs originally employed the type I promoter architecture thought to be common for heavily regulated genes.
The short nascent mRNA produced by Pol II prior to entering the paused state might act as a recruitment platform for regulators of transcription [95]. Thus, future investigations of circadian recruitment regulators of pause release at SCPs may also be directed to nascent mRNA interacting proteins. More broadly, investigators of clock-controlled genes may find the specific signatures of SCPs or type I circadian promoters useful to direct experimental designs when studying mechanisms of transcriptional regulation.
Mouse promoters were compiled from the UCSC mm9 RefSeq annotation [96,97]. Then, CAGE clusters were computed based on the "FANTOM3and4"/"FANTOMtimecourseCAGEmouse"/"liver_under_constant_darkness" data set as provided by the CAGEr/FANTOM3and4CAGE R packages [57], using the recommended procedures in the package. CAGE peak widths were computed for locations ±300 bp of the TSSs provided by the UCSC RefSeq annotation, using the tagClusters function; the widths were defined as the distance between the 10% and 90% quantiles. The CAGE clusters represent mouse liver TSSs; for the cases that a CAGE cluster was measured within 300 bp of a RefSeq-annotated TSS, the CAGE location was used as a TSS instead, resulting in slight adjustments of 14585 out of 33333 RefSeq transcript TSSs. The distribution of CAGE peak widths was bimodal (S2D Fig), prompting a classification of promoters with a CAGE peak width smaller than 10 as focused, otherwise as dispersed.
Then, transcript sets with identical start and end coordinates were reduced to all but one transcript, leaving 27274 transcripts. Further, for TSSs corresponding to more than 1 resulting transcripts, only the shortest transcript was kept, 26251 transcripts now remained. Finally, a set of transcripts unambiguously assignable to a TSS was needed to conduct the present study, since it combines core promoter and transcript properties. For this, any overlapping transcripts with different TSSs needed to be discarded. Finally, transcripts shorter than 500 bp and non-protein coding transcripts (RefSeq annotation starting with "NR") were excluded, leaving a set of 17686 promoters and transcripts used for the following analysis.
Drosophila promoters were compiled from the UCSC dm3 refGene table, which corresponds to the FlyBase transcript annotation [98]. The same procedure as for the mouse genome (above) resulted in 11096 Drosophila promoters.
Drosophila CAGE peaks classified as "peaked" (= focused) or "broad" (= dispersed) were obtained from a study of fly embryos [72]. The nearest CAGE peak within 300 bp of the UCSC refGene TSS annotation (if there was such a peak) was used to classify TSSs.
Mouse liver Nascent-Seq and RNA-Seq data from a study of mice kept under 12 hr/12 hr light-dark cycles [11] were obtained from the NCBI sequence read archive, accession numbers SRP011984 and SRP011981, respectively. Reads were aligned to the UCSC mm9 assembly using the Bowtie2/TopHat2 pipeline [99,100], allowing only uniquely mapped reads. Duplicate reads were removed. Transcripts were quantified against the USCS mm9 RefSeq annotation by counting reads mapping anywhere within the exons for Nascent-Seq, or reads compatible to splicing annotation for RNA-Seq; the R package GenomicAlignments [101] was used to design this workflow. Transcript abundances were reported as standard normalized RPKM values, with a further correction step using the calcNormFactors function of the R package edgeR [102].
Microarray data from samples of livers of mice kept under similar light-dark conditions [63] were obtained from NCBI GEO, accession number GSE33726, RMA normalized and summarized according to the RefSeq annotation using Brainarray v. 18 CDF files [103].
Drosophila head Nascent-Seq and RNA-Seq data [12] were obtained from the NCBI sequence read archive, accession number SRP012175. Transcript abundances were quantified using the same workflow outlined above.
Circadian rhythms in nascent or mature transcript abundances were detected using the RAIN algorithm and R package [104]. For both mouse liver and Drosophila heads, data spanned two days, with samples taken every 4 hrs, yielding 12 samples for each Nascent-Seq or poly(A)+ RNA-Seq data set. Averages and relative amplitudes were quantified by harmonic regression using the HarmonicRegression R package [13]. Expressed transcripts were defined as transcripts having mean nascent RPKM values > 0.1 (mouse) or > 1 (Drosophila). Silent transcripts were defined as having mean nascent RPKM values of < 0.01 in both organisms. Mouse circadian transcripts were defined as expressed transcripts having Benjamini-Hochberg corrected RAIN p values < 0.2 and circadian amplitudes > 0.1 as estimated by harmonic regression. Drosophila circadian transcripts were defined as expressed transcripts having Benjamini-Hochberg corrected RAIN p values < 0.25 (the laxer constraints on Drosophila transcripts were necessary in order to obtain a number of circadian transcripts large enough for reliable statistics). Highly expressed transcripts were defined as the upper 25% quantile of mean transcriptional activities of expressed transcripts. High amplitude transcripts were defined as the upper 25% quantile of transcriptional amplitudes of circadian transcripts.
Position count matrices (PCM) for the TATA box, BREu, BREd DNA sequences were obtained from JASPAR [105]. To discover matches for these matrices, standard logarithmic odds scores for each matrix were computed [106], corresponding to the GC content of each promoter. A rigorous method [107] was used to set score thresholds for each matrix and promoter search region. This method computes the false discovery (FDR) and false negative rates (FNR) of discovery for a PCM and a DNA sequence, given the sequence's GC content; here, the threshold that brings the FDR and FNR as close as possible was used to determine hits. Further, a rigorous method for PCM regularization [107] was used on the JASPAR matrices as a preprocessing step. These algorithms were implemented as the accompanying R/Bioconductor package "profileScoreDist" (https://bioconductor.org/packages/profileScoreDist). The package is generally applicable to any position count matrices.
Scans were made for TATA box PCM matches between −50 (start position) and −10 bp (end position) of mouse and Drosophila TSSs, and when plotting the positions of all mouse matches (not just the one closest to the a priori consensus position [108]), the familiar TATA box peak at around position −30 of the TSS [109] was recovered (S1A Fig), which validated the present approach. The BREu and BREd PCMs were scanned for between the −75/−26 and the −30/−1 positions, respectively. Additionally filtering for evolutionary conservation improves Drosophila DNA binding site predictions, but this effect is less clear for mouse [110]. Hence, only Drosophila TATA box hits with a median phastCons score > = 0.75 were retained. The PhastCons scores [111] were obtained from UCSC for the Drosophila melanogaster dm3 assembly, median phastCons scores were computed for each TATA box hit. For the promoters with more than one TATA box hit (these were often overlapping hits), the maximal median phastCons score was reported.
CpG ratios (observed/expected CG dinucleotides) were determined for mouse promoters using the standard formula CG×N/(G×C) [112], where N is the width of the DNA sequence considered. Here, an interval between −100 and +100 of the TSS was scanned for the CpG ratio determination, so that N = 200 (by convention, there is no "0" position).
Mouse liver ChIP-Seq peaks for BMAL1 (E-box binding) [4] were obtained from the Supplementary file doi:10.1371/journal.pbio.1000595.s019. Normalized tag counts for each time point as given in this file were used to estimate BMAL1 binding phases (ZT) using the HarmonicRegression package.
Mouse liver ChIP-Seq peaks for REV-ERB α and β (ROR element binding) [5] were obtained from GEO (accession numbers GSM840528 and GSM840529, respectively).
Mouse liver ChIP-Seq peaks for E4BP4 (D-box binding) [6] were obtained from GEO (accession number GSM1437733). All mouse liver ChIP-Seq peaks were from reads aligned to the mouse mm9 assembly.
Drosophila ChIP-chip calls for CLK (E-box binding) [8] were obtained from the Supplementary Table 1 of that article, cycling peaks were retained. The chip used was based on the Drosophila dm3 (BDGP5) assembly.
ChIP-Seq and ChIP-chip peaks were narrowed to their center coordinates, which were matched to regions ±3000 of the set of TSSs, except for Drosophila CLK peaks, which were matched to ±2000 bp regions around the TSSs, following the original study. Promoter were classified as CTF binding if they had one or more CTF peak center within these regions.
ChIP-Seq reads were aligned to the mouse mm9 genome assembly using the BWA aligner [113]. Duplicate reads were removed, and aligned reads with a phred alignment quality of 30 or greater were retained.
Reads were mapped to the promoter regions using the GenomicAlignments package [101]; the "coverage" function was used to compute pileups for each promoter: reads per base per million reads (e.g. Fig 4A). Reads were the shifted equal amounts for top and bottom strands to maximize the correlation between coverages at both strands for regions spanning −500 to +500 of all TSSs, resulting in shifts of 35–40 bp for each strand, as in the previous study [10]. The bimodal peaks evident in the Pol II pileups were similar to those observed earlier [32]. Pileups were averaged over the 7 time points of the study. For the final pileup averaging across promoter groups (e.g. Fig 4A), top and bottom 1% quantiles were left out of the averaging, due to a few outliers.
For statistical tests and calculation of circadian phases (ZT), mouse liver promoter-proximal pausing region Pol II levels were quantified for each of the 7 time points as RPKM counts of ChIP-Seq reads overlapping the region +21 to +100 bp of the TSS. Gene body Pol II was quantified as RPKM counts of reads overlapping the positions +301 to +1300 bp downstream of the TSS for transcripts 1300 bp or longer, otherwise between positions +301 and +500 for the few short transcripts. PIs were computed as proximal/gene body Pol II signals.
Promoter-proximal (TSS) H3K4me3 (ChIP-Seq data spanning the same 7 time points as the Pol II data) was quantified in the same way using the +1 to +200 bp interval, downstream gene body H3K4me3 was quantified using the +801 to +1000 bp interval.
As in the original study [10], Pol II and H3K4me3 values were first quantile normalized over the set of promoters, then the R package HarmonicRegression [13] was used to compute phases (ZT) and means based on the 7 time points. This yielded ZT estimations with reasonable confidence intervals, and also p values against the null hypothesis of random signals without rhythms. For 7 time points, no rigorous false discovery rate cutoffs can be applied for thousands of promoters. Rather, to compute phase differences, promoters with HarmonicRegression rhythm p values < 0.1 for all thee of promoter-proximal Pol II, gene body Pol II, and PI, were compiled, and phase differences computed for S3E Fig.
Aligned and normalized mouse liver MNase-Seq and H2A.Z Chip-Seq reads for livers of WT and BMAL1−/− knockout mice [60] were obtained as bigWig files from GEO, accession numbers GSE47142 and GSE47143, respectively. Mouse liver GRO-Seq data were obtained as bigWig files from GEO, accession number GSE59486 [6]. Pileups (reads per bp per ten million reads) were computed with the GenomicAlignments R package for each time point, and then averaged over time points (mean) for all analyses except the nucleosome rhythmicity analysis. For the final pileup averaging across promoter groups (e.g. Fig 4B), top and bottom 1% quantiles were left out of the averaging, due to a few outliers.
For statistical tests, mean nucleosome pileups for the regions between −101 and −1 bp of the TSS were averaged to represent nucleosomal coverages.
The locations of the first (+1) nucleosome peak downstream of the TSSs in the pileups were detected exactly as described in the original report [70]. The procedure was implemented in the accompanying R package peakPick (https://cran.r-project.org/web/packages/peakPick/index.html).
H2A.Z and nucleosomal pileups were averaged over intervals between −80 and +80 bp of the +1 peaks, then the H2A.Z signals were normalized to the nucleosomal signals. This value was used as the H2A.Z signal for downstream analysis.
Drosophila S2 cell MNase-Seq, H2A.Z ChIP-Seq, and 3'NT-Seq data [70] were obtained in the wiggle format from GEO, accession number GSE49106. Pileups for each promoter were created exactly as described for the mouse liver MNase-Seq and ChIP-Seq data, except that the normalization as given in the wiggle files was retained.
For statistical tests, averaged nucleosome pileups for the regions between −100 and −1 bp of the Drosophila TSSs were averaged to represent nucleosomal coverages.
The locations of the first (+1) nucleosome peaks and the H2A.Z signals normalized to the MNase-Seq signal for these +1 peaks were computed exactly as for the mouse liver TSSs described above.
For statistical tests, Drosophila Pol II promoter-proximal pausing region signals were quantified as 3'NT-Seq normalized pileups averaged over the region +1 to +100 bp of the TSS. Gene body Pol II signals were quantified as normalized pileups averaged over the region +250 –+650 bp downstream of the +1 nucleosome. Due to a few outliers, top and bottom 1% quantiles were left out of the averaging. Finally, Drosophila PIs were computed as the ratios between these two signals. The original study [70] analyzed the specific phenomenon of Pol II stalling at the +1 nucleosome position, a phenomenon distinct from regulated pausing. These results were reproduced and related to the PI estimation used here (S5 Fig).
Mean transcriptional activities were binned into 55 bins spanning the logarithmic scale. Non-CTF binding constitutive promoters with TATA box (constitutive type I) had probabilities assigned to each bin. These probabilities were weighted according to the counts of SCPs with transcriptional activities falling within the corresponding bin, then normalized. This enabled sampling of non-CTF binding constitutive promoters to obtain populations with the same mean transcriptional activity distributions as SCPs. Properties of SCPs compared to those of the sampled population of "non-SCPs with SCP-like transcriptional activities" could then be analyzed to distinguish SCP features that are not merely epiphenomena of high transcriptional activities.
For this, standard bootstrapping procedure was employed: The SCP properties and constitutive type I promoter properties to compare (such as CpG ratio) and their probability distributions (uniform for SCPs, weighted probabilities as outlined above for constitutive type I promoters) were pooled, samples of the same size as the SCP group and the constitutive type I promoters group, respectively, were drawn randomly with replacement from the pool, and a rank sum test statistic was computed each time. This was done 1,000,000 times. The resulting empirical test statistic probability distribution was then used when repeatedly drawing samples from each population separately, each time with sample sizes corresponding to the population sizes, rank sum test statistics and median location differences were computed. Then, two-sided p values could be estimated; 1000 such p values were computed for each property, median p value was reported (S2 Table).
Gene symbols for up- (at least 2-fold) and down-regulated (at least 0.5-fold) transcripts at Postnatal day 30 (P30) in livers of Taf10lox/lox-AlbCre mice were obtained from Supplementary Table 1 of the original article [62], then matched to the promoter collection used in the present work. These mice experience a liver-specific deletion of the Taf10 gene postnatally between days P15 and P22.
|
10.1371/journal.pbio.1000095 | Reawakening Retrocyclins: Ancestral Human Defensins Active Against HIV-1 | Human alpha and beta defensins contribute substantially to innate immune defenses against microbial and viral infections. Certain nonhuman primates also produce theta-defensins—18 residue cyclic peptides that act as HIV-1 entry inhibitors. Multiple human theta-defensin genes exist, but they harbor a premature termination codon that blocks translation. Consequently, the theta-defensins (retrocyclins) encoded within the human genome are not expressed as peptides. In vivo production of theta-defensins in rhesus macaques involves the post-translational ligation of two nonapeptides, each derived from a 12-residue “demidefensin” precursor. Neither the mechanism of this unique process nor its existence in human cells is known. To ascertain if human cells retained the ability to process demidefensins, we transfected human promyelocytic cells with plasmids containing repaired retrocyclin-like genes. The expected peptides were isolated, their sequences were verified by mass spectrometric analyses, and their anti-HIV-1 activity was confirmed in vitro. Our study reveals for the first time, to our knowledge, that human cells have the ability to make cyclic theta-defensins. Given this evidence that human cells could make theta-defensins, we attempted to restore endogenous expression of retrocyclin peptides. Since human theta-defensin genes are transcribed, we used aminoglycosides to read-through the premature termination codon found in the mRNA transcripts. This treatment induced the production of intact, bioactive retrocyclin-1 peptide by human epithelial cells and cervicovaginal tissues. The ability to reawaken retrocyclin genes from their 7 million years of slumber using aminoglycosides could provide a novel way to secure enhanced resistance to HIV-1 infection.
| Defensins are a large family of small antimicrobial peptides that contribute to host defense against a broad spectrum of pathogens. In primates, defensins are divided into three subfamilies—alpha, beta, and theta—on the basis of their disulfide bonding pattern. Theta-defensins were the most recently identified defensin subfamily, isolated initially from white blood cells and bone marrow of rhesus monkeys. They are the only known cyclic peptides in mammals and act primarily by preventing viruses such as HIV-1 from entering cells. Whereas theta-defensin genes are intact in Old World monkeys, in humans they have a premature stop codon that prevents their expression; they thus exist as pseudogenes. In this work, we reveal that, upon correction of the premature termination codon in theta-defensin pseudogenes, human myeloid cells produce cyclic, antiviral peptides (which we have termed “retrocyclins”), indicating that the cells retain the intact machinery to make cyclic peptides. Furthermore, we exploited the ability of aminoglycoside antibiotics to read-through the premature termination codon within retrocyclin transcripts to produce functional peptides that are active against HIV-1. Given that the endogenous production of retrocyclins could also be restored in human cervicovaginal tissues, we propose that aminoglycoside-based topical microbicides might be useful in preventing sexual transmission of HIV-1.
| Nearly 33 million people are infected with HIV worldwide [1,2], and despite extensive efforts there are no effective vaccines or other countermeasures to protect against HIV transmission [3]. In our attempts to find effective anti-HIV agents, our group determined that certain synthetic θ-defensins called “retrocyclins” are potent inhibitors of HIV-1 infection [4–8]. Retrocyclins belong to a large family of antimicrobial peptides known as defensins, all of which are cationic, tri-disulfide bonded peptides that have important roles in innate host defense. On the basis of the position of the cysteines and the disulfide bonding pattern, defensins are grouped into three subfamilies: α-defensins, β-defensins, and θ-defensins [9,10].
θ-Defensins such as retrocyclin have a cyclic peptide backbone, derived from the head-to-tail-ligation of two peptides that each contributes nine amino acids to form the 18-residue mature peptide [11]. θ-Defensins are the only known cyclic peptides in mammals and were originally isolated from rhesus macaque leukocytes and bone marrow [11–13]. While θ-defensin peptides are produced in old world monkeys and orangutans, in humans they exist only as expressed pseudogenes [14]. A premature termination codon in the signal peptide portion of human retrocyclin mRNA prevents its translation. The retrocyclin gene is otherwise remarkably intact, showing 89.4% identity with rhesus θ-defensins. Its genetic information was utilized to recreate retrocyclins synthetically and confirm their activity against both X4 and R5 strains of HIV-1 [4–7].
Retrocyclins inhibit the fusion of HIV-1 Env by selectively binding to the C-terminal heptad repeat region on gp41 blocking 6-helix bundle formation [15,16]. RC-101 is a congener of retrocyclin with a single arginine to lysine substitution that retains structural and functional similarity to retrocyclin [4]. RC-101 exhibited enhanced anti-HIV-1 activity against over two dozen primary isolates from several clades [7,8], and did not induce inflammation or toxicity in organotypic models of human cervicovaginal tissue [17]. Continuous passaging of HIV-1 BaL in the presence of subinhibitory concentrations of RC-101 for 100 days induced only minimal viral resistance [18]. Given these beneficial attributes, we envisioned that restoring the endogenous expression of retrocyclins in humans would provide an effective and natural way of combating HIV-1 infection.
In the current study we restored the translation of this evolutionarily lost retrocyclin peptide by ablating the premature termination codon using site-directed mutagenesis, and analyzed whether human cells can synthesize biologically active retrocyclins. We found that promyelocytic HL60 cells stably transfected with retrocyclin constructs in which the premature termination codon was corrected could express retrocyclins. Application of the expressed retrocyclins to TZM-bl cells, PM1 cells, and peripheral blood mononuclear cells (PBMCs) conferred protection against HIV-1 infection. Moreover, mass spectrometric techniques confirmed the presence of correctly folded mature retrocyclin peptides. We also explored methods to read-through the premature termination codon within the retrocyclin pseudogene. Previous reports revealed that aminoglycoside antibiotics could suppress the termination codon of pseudogenes and disease-associated nonsense mutations [19–25]. In bacteria, aminoglycosides bind strongly to the decoding site on the 16S rRNA, thereby hindering protein synthesis [26]. However, in eukaryotes, aminoglycosides bind to the eukaryotic decoding site with low affinity and induce a low level of translational misreading, which suppresses the termination codon through the incorporation of an amino acid in its place [27]. Herein, we utilized aminoglycosides to induce translational read-through of the θ-defensin pseudogene, which restored the expression of functional anti-HIV-1 retrocyclin peptides in human cervicovaginal tissue models. Topical application of aminoglycosides to produce endogenous retrocyclins in the vaginal mucosa might soon be an effective preventative to combat sexual transmission of HIV-1.
θ-Defensins are formed by post-translational modification of two 12-residue gene products, each of which is processed to give a nonapeptide that contains three cysteines. The N-terminus of one nonapeptide forms a peptide bond with the C-terminus of another nonapeptide, resulting in a cyclic 18 residue peptide with three intramolecular disulfide bonds [11,14]. To determine if human cells have retained the ability to process θ-defensins, we transfected promyelocytic HL60 cells with retrocyclin constructs each encoding a nonapeptide in which the premature termination codon was replaced with a glutamine (⊗17Q).
Four types of constructs were produced: R1, R3, A1, and A3 (Figure 1). Aside from the corrected premature termination codon (⊗17Q), all constructs were engineered to contain two termination codons at the end of the gene to ensure read-fidelity. Constructs with an “R” designation terminate after the retrocyclin portion of the gene, while constructs with an “A” designation contain the retrocyclin portion with additional downstream residues that might be critical for translation and/or processing [14,28]. Constructs with a “1” designation do not have any additional residues mutated, while constructs with a “3” designation have the additional Arg → Lys mutation (R70K) encoding the RC-101 nonapeptide. HL60 cells were cotransfected by electroporation with either R1 and R3, or A1 and A3, and propagated in the presence of G418 (300 μg/ml) to create stably transfected cell lines. Stable transfection was verified by analyzing genomic DNA and mRNA (Figure S1). Since two different constructs were cotransfected for each condition, combinatorially it would be possible to generate three different retrocyclin peptides as illustrated in Figure 1B. For example, if cells were cotransfected with the R1 and R3 constructs, they could theoretically generate a heterodimer (HL60 cells containing retrocyclin constructs R1 and R3 [R1R3]) or homodimers (R1R1 or R3R3).
We next analyzed if correcting the termination codon in the retrocyclin constructs could restore the translation of biologically active retrocyclin peptides. The infection of TZM-bl cells with HIV-1 BaL was significantly reduced when cells were treated with cellular acid extracts of R1R3 cells (p < 0.004) and HL60 cells containing retrocyclin constructs A1 and A3 (A1A3) (p < 0.002) (Figure 2A). A standard tetrazolium MTT assay revealed that the extracts did not affect cellular metabolism at the concentrations used in the experiment (Figure 2E). Addition of A1A3 cell extracts to HIV-1 infected PM1 cells (Figure 2B) and PBMCs (Figure 2C) showed significant (p < 0.002 and p < 0.004, respectively) decrease in the viral titer as compared to cells treated with control HL60 cell extract. A trypan blue exclusion assay was performed in PBMCs to monitor cell viability (Figure 2F). We next affinity purified R1R3 and A1A3 cell extracts using anti-RC-101 antibody and confirmed the antiviral activity in a luciferase-based assay system (Figure 2D). Interestingly, A1A3 cell extracts were found to be consistently more active than equivalent amounts of R1R3 cell extract, which suggests a role for the downstream residues in retrocyclin processing. These results indicate that biologically active recombinant retrocyclin peptides can be synthesized in human promyelocytic cells. As a next step we tested the presence of retrocyclin in promyelocytic cells using immunostaining.
Immuno-dotblot analyses revealed that our anti-RC-101 antibody specifically recognized lysine-containing human retrocyclin analogs (synthetic RC-101 and RC-101_2K) and RC-100 (i.e., wild-type form) to a lesser extent (Figure 3A) but not human neutrophil peptides 1–3, or peptides with very similar tertiary structure including rhesus theta defensin-1 (RTD-1) and protegrin-1 (PG-1) (Figure 3B). This antibody was used to visualize the expressed retrocyclin peptides in the stably transfected HL60 cells by immunofluorescence staining, which revealed that R1R3 cells and A1A3 cells were brightly stained as compared to vector control (VC) cells (Figure 3C). Slides treated with preimmune serum showed no staining (unpublished data). Note that the staining of A1A3 was brighter than R1R3 and the morphology of A1A3 cells was smaller than VC cells. Experiments were next designed to purify and confirm the identity of the expressed retrocyclin peptides from the cell extracts.
Reverse-phase high-performance liquid chromatography (RP-HPLC) was utilized to purify the recombinant retrocyclin peptides from stably transfected HL60 cell extracts. Figure 4A shows the RP-HPLC trace of A1A3 and synthetic RC-101. Synthetic RC-101 was recovered in fractions collected at 26–28 min. A1A3 HPLC Fractions collected from 23–30 min were analyzed on a 16% Tricine-SDS-gel. Control samples did not contain any protein bands at the expected size, whereas fractions from R1R3 cell extracts revealed protein bands of about 6-kDa size (unpublished data). Interestingly, A1A3 HPLC fractions revealed multiple protein bands, which we further analyzed by western blot (Figure 4B). The western blot analysis revealed bands at sizes corresponding to a monomer, dimer, and trimer of retrocyclin. Interestingly, the presence of multimeric forms of retrocyclin has been independently observed by Daly and colleagues [29]. Furthermore, the RP-HPLC purified A1A3 fractions inhibited entry of HIV-1 BaL in TZM-bl cells (Figure 4C). The IC50 of retrocyclin peptides expressed by A1A3 cells (2 μg/ml) was similar to that of synthetic RC-101 (1.25 μg/ml) [8].
To determine the identity of the retrocyclin peptide expressed by A1A3 cells, HPLC fraction 26 was analyzed by mass spectrometric analysis (MALDI-TOF-MS) at the Microchemical and Proteomics Facility, Emory University (Atlanta, Georgia, US). Analysis of A1A3 Fraction 26 revealed peaks with masses 1,889.775 Da (oxidized) and 1895.890 Da (reduced), which is nearly identical to the expected mass of synthetic cyclic RC-101 (1,889.85 Da and 1,895.96 Da, respectively; unpublished data) and is in agreement with reduction of the three disulfide bridges in the molecule. Furthermore, treatment with iodoacetamide yielded mass species of 2,238.081 Da for the A1A3 fraction 26 and 2,238.071 Da for RC-101 corresponding to the predicted 6-fold–alkylated form of RC-101 (expected mass = 2,238.097 Da). Comparison of spectrum of the Lys-C digest of reduced/alkylated synthetic RC-101_2K (peak at 1,123.577 Da; peptide cleaved at two Lys-Gly bonds; Figure 4D), synthetic RC-101 (peak at 2,256.097 Da; peptide cleaved at a single Lys-Gly bond; N-terminal sequence determined as: Gly-Ile-Cys-Arg-; Figure 4E), and A1A3 fraction 26 (peak at 2,256.010 Da) suggests that the A1A3 cells are expressing RC-101 (Figure 4F). These data confirmed that correctly folded mature retrocyclin peptides can be expressed by human cells. In the following experiments we explored alternative methods to express the peptide endogenously. Of particular interest was the effect of aminoglycosides in mediating varying degrees of termination codon read-through as previously described [19–25].
We tested the ability of three commonly used aminoglycosides (gentamicin, amikacin, and tobramycin) to induce termination codon read-through of retrocyclin cDNA. The native retrocyclin gene was fused with a luciferase reporter at the C terminus to create two constructs: unrescued RC-101 and rescued RC-101 (positive control) as shown in Figure 5A. These constructs were transfected into HOS-CD4-CCR5 cells, grown in the presence of varying concentrations of aminoglycosides, and the degree of read-through quantified by measuring luciferase. Application of tobramycin (10 μg/ml) was the most effective, producing a 26-fold increase in read-through (p < 0.0007; Figure 5B).
Having thus established the optimal aminoglycoside concentration required to achieve read-through of retrocylin cDNA, we next determined if aminoglycosides could restore the translation and anti-HIV-1 activity of native retrocyclin peptides. HeLa-derived cells lines such as TZM-bl cells can natively express retrocyclin mRNA (unpublished data). We applied aminoglycosides to TZM-bl cells and challenged them with HIV-1 BaL. We found that cells treated with gentamicin and tobramycin significantly (p < 0.0005 and p < 0.0001, respectively) inhibited HIV-1 infection as compared to untreated cells (Figure 5C). The effect was modest when compared to inhibition by synthetic peptides. Cell viability, determined by a tetrazolium-based MTT assay, was not affected by the application of aminoglycosides at the mentioned concentrations (Figure 5E).
In order to visualize the retrocyclins expressed by application of aminoglycosides, we performed immunostaining. TZM-bl cells were treated with PBS control or 10 μg/ml tobramycin and stained with anti-retrocyclin antibody or preimmune serum. Control cells showed no staining while cells treated with tobramycin revealed brightly stained cells suggesting that aminoglycosides can induce the expression of retrocyclin peptides (Figure 5D).
We next incubated TZM-bl cells with tobramycin (10 μg/ml) for 24 h, and then treated the cells with preimmune or anti-retrocyclin serum followed by infection with HIV-1. Figure 5F reveals that cells treated with preimmune serum showed a modest yet significant reduction in infection as compared to cells treated with anti-retrocyclin antibodies (p < 0.018), suggesting that the antibody inhibited the endogenous retrocyclins. These data confirm that the anti-HIV-1 activity observed is due to the endogenous retrocyclin peptides expressed when tobramycin was applied to cells.
We next analyzed the ability of aminoglycosides to induce the expression of retrocyclin peptides in an organotypic model cervicovaginal tissue. Tissues were treated apically with tobramycin or control (PBS) for 24 h and anti-retrocyclin immunohistochemical analysis was performed. Interestingly, tissues treated with tobramycin alone and stained with anti-retrocyclin antibody revealed brightly stained cells (Figure 6A) suggesting that production of retrocyclin peptides is induced upon application of aminoglycosides. Lactate dehydrogenase (LDH) activity in the medium underlying the tissues was performed to determine tissue cytotoxicity. The LDH assay revealed that application of 10 μg/ml tobramycin was not cytotoxic to the tissues (Figure 6B). In addition, treatment of tobramycin did not affect the metabolic activity adversely, which was determined by an MTT assay performed on one tissue (unpublished data).
In order to purify endogenous retrocyclins expressed in the tissues, we utilized RP-HPLC. Figure 6C shows an HPLC trace of control, tobramycin-treated tissue extracts as compared to synthetic RC-100 peptide. Synthetic RC-100 peptide was recovered in fractions collected at 27–29 min. Corresponding fractions from control and tobramycin-treated tissues were analyzed by immuno-dotblot analysis using the anti-RC-101 antibody. Figure 6D shows that retrocyclin peptides were recovered in fractions 27–29 min in tobramycin-treated tissue samples but not in control tissue samples. The amount of retrocyclin (RC-100) expressed in tobramycin-treated cervicovaginal tissues was estimated by densitometry to be approximately 1.6 μg/tissue. Together these studies show that aminoglycosides are promising molecules to suppress the premature termination codon of retrocyclin transcripts and restore the ability of cervicovaginal tissues to protect cells from HIV-1.
Identifying effective drugs to prevent HIV-1 infection and other viral infections is essential for countering the spread of these diseases. Exogenous (synthetic) retrocyclins exhibit full activity in the complex environment of vaginal fluid and the peptide is very well tolerated in organotypic human cervicovaginal tissue models [17]. Moreover, HIV-1 evolves little resistance during continued passaging in the presence of the peptide [18]. For these and other reasons, retrocyclins have emerged as potential topical microbicides to protect against sexually transmitted HIV-1 infections.
In this study we have taken a different path towards developing θ-defensin therapeutics. The human pseudogenes that encode the demidefensin precursors whose post-translational processing gives rise to mature retrocyclin are expressed at the mRNA level in multiple organs, including the spleen, bone marrow, thymus, testis, and skeletal muscle [14], and cervicovaginal epithelia (A. M. Cole, unpublished data). By transfecting human myeloid cells with plasmids containing retrocyclin genes without a premature termination codon, we demonstrated that the “machinery” needed to process, trim, splice, and oxidize retrocyclin precursors was available in human myeloid cells. Two sets of expression constructs were transfected into cells: a shorter form (R1R3) that terminates at the end of the retrocyclin gene and a longer form that contains (A1A3) additional 3′ untranslated residues (UTR). Interestingly, A1A3 cells expressed higher levels of retrocyclin peptides as compared to R1R3 cells indicating a role for additional residues in the translational efficiency of these peptides. This was not altogether surprising as other studies have shown that the length of the 3′-UTR regulates translation efficiency [28,30]. Finally, we showed that aminoglycoside-treated cells and cervicovaginal tissues could produce retrocyclins endogenously by suppressing the premature termination codon in their endogenous mRNA transcript.
Since approximately 30% of inherited disorders may result from premature termination codon mutations, there has been tremendous interest and some progress in developing and applying agents that can read-through premature UAA, UAG, or UGA termination codons [25]. Although aminoglycosides, as used in this study, have been most widely investigated, exciting new agents such as PTC-124, have also appeared [31,32]. In a sense, human retrocyclin-deficiency is also an inherited disorder, albeit one with an incidence of 100%. It is caused by a premature termination codon mutation that occurred after human lineage diverged from the lineage we share with orangutans, lesser apes, and old world monkeys. Since HIV-1 and other viruses that currently infect humans have evolved in the absence of selective pressure exerted by retrocyclins, the ability to reawaken this ancestral molecule could be used to strengthen the innate immune system's ability to prevent or limit the infections they now induce.
HL60 cells [33,34] obtained from ATCC were cultured in Iscoves's DMEM with 20% FBS, 100 U/ml penicillin, and 100 μg/ml streptomycin (I20). TZM-bl cells [35] stably expressing CD4, CCR5, and CXCR4, has firefly luciferase gene under the control of HIV-1 promoter (from J. C. Kappes, X. Wu, and Tranzyme Inc). TZM-bl, HOS-CD4-CCR5 [36,37] (from N. R. Landau), PM1 cells [38], (from M. Reitz), and HIV-1 BaL, an R5 tropic strain, were all procured through the National Institutes of Health (NIH) AIDS Research and Reference Reagent program. HIV-1 BaL viral stocks were prepared by infecting PM1 cells [18]. PBMCs were isolated from blood drawn from a healthy HIV-1 seronegative donor as per the guidelines of the institutional review board of University of Central Florida. PBMCs were isolated using Lymphosep (MP biomedicals LLC), and cultured in RPMI-1640 medium with 10% FBS (R10) supplemented with 50 units of IL-2 (R10-50U) and 5 μg/ml of phytohemagglutinin (PHA) for 3 d. The cells were then resuspended in R10-50U at a density of 0.8 × 106 cells/ml and grown for 5–6 d.
Cervicovaginal tissues (EpiVaginal) were obtained from MatTek Corporation and maintained in proprietary growth medium as per the company's guidelines. The tissues were composed of a full-thickness, stratified vaginal-ectocervical layer intermixed with Langehans cells and underlying lamina propria. The tissues were allowed to grow on transwell cell culture inserts at the air-liquid interface.
|
10.1371/journal.pntd.0002169 | Uncertainty Surrounding Projections of the Long-Term Impact of Ivermectin Treatment on Human Onchocerciasis | Recent studies in Mali, Nigeria, and Senegal have indicated that annual (or biannual) ivermectin distribution may lead to local elimination of human onchocerciasis in certain African foci. Modelling-based projections have been used to estimate the required duration of ivermectin distribution to reach elimination. A crucial assumption has been that microfilarial production by Onchocerca volvulus is reduced irreversibly by 30–35% with each (annual) ivermectin round. However, other modelling-based analyses suggest that ivermectin may not have such a cumulative effect. Uncertainty in this (biological) and other (programmatic) assumptions would affect projected outcomes of long-term ivermectin treatment.
We modify a deterministic age- and sex-structured onchocerciasis transmission model, parameterised for savannah O. volvulus–Simulium damnosum, to explore the impact of assumptions regarding the effect of ivermectin on worm fertility and the patterns of treatment coverage compliance, and frequency on projections of parasitological outcomes due to long-term, mass ivermectin administration in hyperendemic areas. The projected impact of ivermectin distribution on onchocerciasis and the benefits of switching from annual to biannual distribution are strongly dependent on assumptions regarding the drug's effect on worm fertility and on treatment compliance. If ivermectin does not have a cumulative impact on microfilarial production, elimination of onchocerciasis in hyperendemic areas may not be feasible with annual ivermectin distribution.
There is substantial (biological and programmatic) uncertainty surrounding modelling projections of onchocerciasis elimination. These uncertainties need to be acknowledged for mathematical models to inform control policy reliably. Further research is needed to elucidate the effect of ivermectin on O. volvulus reproductive biology and quantify the patterns of coverage and compliance in treated communities.
| Studies in Mali, Nigeria, and Senegal suggest that, in some settings, it is possible to eliminate onchocerciasis after 15–17 years of ivermectin distribution. Computer models have been used to estimate the required duration of ivermectin distribution to reach elimination. Some models assume that annual ivermectin treatment reduces the fertility of the causing parasite, Onchocerca volvulus, by 30–35% each time the drug is taken. Other analyses suggest that ivermectin may not have such an effect. We explore how assumptions regarding: a) treatment effects on microfilarial production by female worms (fertility), b) proportion of people who receive the drug (coverage), c) proportion of people who adhere to treatment (compliance), and d) whether people are treated once or twice per year (frequency) affect temporal projections of infection load and prevalence in highly endemic African savannah settings. We find that if treatment does not affect parasite fertility cumulatively, elimination of onchocerciasis in highly endemic areas of Africa may not be feasible with annual ivermectin distribution alone. If two areas have equal coverage but dissimilar compliance, they may experience very different infection load, prevalence and persistence trends. Projections such as these are crucial to help onchocerciasis control programmes to plan elimination strategies effectively.
| Human onchocerciasis, caused by Onchocerca volvulus and transmitted by Simulium blackflies, is a parasitic disease leading to ocular (vision loss, blindness) and cutaneous (itching, dermatitis, depigmentation) pathology [1], [2], as well as to increased host mortality [3], [4], [5].
The Onchocerciasis Control Programme in West Africa (OCP) started in 1974. The programme was initially based on vector control until, in 1987, ivermectin was registered for human use against onchocerciasis. Thereupon, Merck & Co. Inc. took the unprecedented decision to donate ivermectin for as long as needed to eliminate onchocerciasis as a public health problem [6]. Mass drug administration (MDA) of ivermectin began in some OCP regions in 1988–1989, particularly in extension areas [7]. In some areas of the OCP both antivectorial and antiparasitic measures were combined, whilst in others (mainly in the western extension) ivermectin distribution alone, annually or biannually, was implemented [7], [8]. The African Programme for Onchocerciasis Control (APOC) was launched in 1995 to target the 19 onchocerciasis endemic countries in Africa not covered by the OCP [8], [9]. APOC's strategy involved the establishment of effective and sustainable, community-directed, annual mass ivermectin treatment for all those aged five years and older [10], [11]. The programme, initially conceived to end in 2007 [8], and subsequently in 2015 [12], has recently been extended until 2025 with the new goal and commitment for the elimination of onchocerciasis [13].
In addition to OCP western extension areas that were treated twice-yearly (e.g. Senegal [7]), some countries such as Ghana (in the former OCP), and Uganda (in APOC), have adopted a biannual treatment strategy in selected foci; the former because of suspected suboptimal responses to ivermectin treatment [14], and the latter because, in combination with vector control, elimination may be accelerated [15], [16].
Ivermectin is a potent microfilaricide, causing a greater than 90% reduction in skin microfilarial load within a few days, and a maximum reduction of 98–99% two months after treatment [17]. Ivermectin also has an embryostatic effect on adult female worms, temporarily blocking the release of microfilariae (mf) [18]. The efficacy of the embryostatic effect is approximately 70%, with the maximum reduction of microfilarial production reached one to two months after treatment [17]. Recuperation of adult worms' fertility occurs slowly from three to four months after treatment onwards [17], [18] but may not regain its original level up to 18 months after treatment. (The term fertility is used here to refer to worms producing live, stretched mf, by contrast with females producing oocytes or embryos, which would correspond to worm fecundity [17].)
Recent epidemiological and entomological evaluations conducted in Mali and Senegal suggest that 15–17 years of annual (or biannual) ivermectin distribution (in the absence of vector control) may be sufficient to lead to local onchocerciasis elimination in certain foci [19]. In addition, local elimination may have been achieved with 15–17 years of ivermectin distribution in 26 villages in Kaduna state, Nigeria (the first report of such evidence for the operational area of APOC) [20]. These studies have provided proof of principle that elimination with annual ivermectin distribution may be feasible in some African foci. In 2009, an international expert group convened to discuss the implications of these results [21]. Based on experiences with cessation of onchocerciasis control in West Africa and predictions from mathematical models, the group developed an operational framework for elimination and provisionally defined transmission thresholds, namely, a microfilarial prevalence below 5% in all surveyed villages (and below 1% in 90% of the villages), and a proportion of local simuliid vectors harbouring <0.5 L3 larvae per 1,000 flies [19], [21].
Mathematical models such as [22], have been used to assess the feasibility of, and predict the duration of ivermectin distribution required for elimination [23]. In these modelling projections, overall (therapeutic) treatment coverage was varied as part of the sensitivity analysis, and those not taking treatment included a (correlated but unreported) fraction of systematic non-compliers. However, the effect of systematic non-compliers (i.e. the proportion of the population aged five years and older who never take treatment) on the feasibility of elimination was not investigated independently from that of coverage. A crucial conjecture of these projections (based on analysis of a 5-year community ivermectin trial in Asubende, Ghana [24]), was that adult female worms, after temporarily ceasing microfilarial production due to the embryostatic effect of ivermectin, gradually reach a new production level which is reduced irreversibly by an average of 30–35% after each treatment round [25], effectively assuming a cumulative effect of ivermectin on female worm fertility (equivalent to an increasing proportion of worms not contributing to transmission; a sort of ‘macrofilaricidal’ effect [23], [25]). However, another modelling study, using data from a community trial with five biannual treatment rounds in Guatemala [26], did not find evidence for a cumulative effect on microfilarial production [27].
Whether or not ivermectin has a cumulative effect on female worm fertility [28], [29] will have important implications for the optimal design of MDA programmes, and given the sparse data that exist, this issue represents an area of considerable uncertainty which needs to be taken into account in modelling studies estimating the long-term impact of ivermectin treatment on parasite populations in humans and vectors.
In this paper, we modify our current onchocerciasis transmission model [30] to explore the uncertainty in modelling projections of the long-term impact of ivermectin on O. volvulus populations due to assumptions concerning: a) the effect of ivermectin on mf production by female worms (biological variables), and b) treatment coverage and compliance (programmatic variables). We also explore how these affect the benefit of annual vs. biannual treatment frequency.
We modified our sex- and age-structured deterministic onchocerciasis transmission model [30], [31], which describes the rate of change with respect to time and host age of the mean number of fertile and non-fertile female adult worms per host, the mean number of microfilariae per milligram (mg) of skin (mf/mg), and the mean number of infective (L3) larvae per fly. To obtain infection prevalence from infection intensity in humans, we assumed that the distribution of mf among hosts is negative binomial as described in [32]. A detailed description of the model equations is given in Supporting Information Text S1: Protocol S1, Onchocerciasis Population Dynamics Model. Parameter definitions and values can be found in Supporting Information Text S2: Supplementary Tables, Table S1.
After each dose of ivermectin there is a microfilaricidal effect with 99% efficacy, and a reduction in microfilarial production (embryostatic effect) by fertile female worms [17]. The ivermectin-exposed adult worms are then assumed either to: a) reach a new microfilarial production rate which is reduced by 30% ten months after each treatment round (representing a cumulative effect, depicted in Figure 1A), or b) resume microfilarial production, which ten months after each treatment would reach 70% of its baseline value, i.e. is also reduced by 30% from baseline, but the reduction is not additive (representing a non-cumulative effect, as concluded in [27], and illustrated in Figure 1B). The equations modelling the effect of ivermectin in female worm fertility are described in Supporting Information Text S1: Protocol S2, Modelling the Cumulative Effect of Ivermectin. Parameter definitions and values can be found in Supporting Information Text S2: Supplementary Tables, Table S2.
Although the cumulative reduction proposed in [25] was estimated from data corresponding to annual ivermectin distribution [24], it was assumed that in the case of biannual treatments, each 6-monthly treatment causes the same proportional reduction. An analysis of the sensitivity of model outputs to this assumption was conducted following [23]. Ivermectin was assumed to have no macrofilaricidal action (i.e. not to reduce adult worm life-expectancy) at the standard dose used for MDA [17], [33], [34], and to have intact efficacy, i.e., no sub-optimal response [14] or drug resistance [35] were included.
The model is stratified into four treatment compliance classes: a first group of individuals who take treatment every round; two groups who take treatment every other round alternately, and a fourth group who never take treatment. The latter class represents individuals in the community who are systematic non-compliers, as opposed to a situation in which a proportion of individuals miss some treatment rounds (e.g. because they are absent or pregnant at the time of treatment). The proportion of systematic non-compliers was set at 0.1%, 2%, and 5% to investigate its effect on model outputs. These values were chosen to explore potential variability in this parameter. A recent ivermectin compliance study reported that 6% had never taken the drug over the course of eight consecutive treatment rounds [36]. The four compliance groups were assumed not to differ in exposure to vectors (which depends on age and sex according to [30]). Children under five years were not treated in the model as they are not eligible to receive ivermectin.
Human age- and sex-structure reflects the demography in savannah areas of northern Cameroon [37], [38], as it is in savannah areas of Africa that the prevailing O. volvulus–S. damnosum combinations are responsible for the most severe sequelae of onchocerciasis [1], [2]. Parameters for vector competence, survival, and host choice were those for savannah species of the Simulium damnosum complex (S. damnosum sensu stricto and S. sirbanum) [30], [39], responsible for onchocerciasis transmission in the region [40], [41].
The overdispersion parameter for the distribution of adult worms among hosts was as estimated in [27] (see Supporting Information Text S1: Protocol S3, Mating Probability and Supporting Information Text S2: Supplementary Tables, Table S3). The parameterisation of the relationship between microfilarial prevalence and load was that for West African savannah areas [32] (see Supporting Information Text S1: Protocol S4, Microfilarial Prevalence and Supporting Information Text S2: Supplementary Tables, Table S3). The annual biting rate (ABR) by blackfly vectors was set to 19,000 bites per person per year (well within the range of values recorded in savannah areas [32], [40], [41]), to achieve a baseline mean microfilarial load of 27 mf/mg (all ages), and of 44 mf/mg of skin in those aged 20 years and above. This resulted in an overall microfilarial prevalence (all ages) of 70%, representing an area of high baseline endemicity. In onchocerciasis, hyperendemic areas are those with overall infection prevalence above 60% [42], but this class can encompass a wide range of transmission and infection intensities. (Note that the mean microfilarial load per mg of skin in those aged ≥20 years here is an arithmetic mean, not a geometric mean of the number of microfilariae per skin snip (ss) (mf/ss) in the same age group, known as the community microfilarial load (CMFL) [43].) Understanding the long-term impact of ivermectin in highly hyperendemic areas is particularly important, as such areas will be those in which controlling the disease has the highest priority (morbidity will be more severe), elimination of the infection reservoir is likely to be more difficult or take longer [23], and from which the infection could reinvade controlled areas.
The model was used to explore the effect of 15 years of (annual or biannual) mass ivermectin distribution on: a) infection intensity defined as mean microfilarial load per mg of skin in those aged ≥20 years, and b) prevalence of microfilaridermia in the overall population. We choose 15 years as a suitable timescale to investigate the impact of long-term treatment of onchocerciasis with ivermectin, motivated by the epidemiological studies described in [19], [20]. Since the model is deterministic, the probability of reaching elimination was not investigated.
The sensitivity of the above model outputs was explored regarding the following assumptions: 1) cumulative effect of ivermectin on female worm fertility (present vs. absent); 2) overall therapeutic coverage (proportion of the total population receiving ivermectin at each round: 60%, 70%, 80%); 3) proportion of systematic non-compliers (those who never take treatment: 0.1%, 2%, 5%); and 4) treatment frequency (annual vs. biannual). In order to explore the extent to which our results were sensitive to the assumption that biannual treatments each caused the same reduction in fertility of 30% per treatment; we also explored model outputs with a more conservative reduction of 16.5% per 6-monthly treatment (which gives an overall annual reduction of 30%).
Model outputs indicate that the assumption of a cumulative impact of ivermectin on microfilarial production by female O. volvulus has a substantial effect on projections of long-term ivermectin treatment (Figure 2). Regarding infection intensity in adults aged 20 years and older, there would be a very pronounced decrease partly due to little repopulation of the skin by mf, and partly due to the ensuing suppressed transmission. This is because, under this conjecture, the model assumes that the number of mf produced per female worm per unit time would progressively be reduced to a very low level. By contrast, under the assumption of ivermectin not exerting a cumulative effect on microfilarial production, there is a substantial amount of repopulation of the skin by mf in-between annual treatments, leading to more transmission and a smaller impact on infection intensity.
Assumptions regarding the operation or absence of a cumulative effect of ivermectin on parasite fertility can also influence the expected relative benefits of annual vs. biannual treatment frequency regarding reductions in infection intensity, prevalence, and transmission. In the presence of a cumulative reduction with each treatment round, there is initially a very marked benefit of the biannual distribution on the reduction of parasitological indicators (as the rate of microfilarial production is rapidly reduced). However, after repeated treatments, there would be much less difference in the long-term impact of ivermectin treatment on microfilarial prevalence compared to an annual treatment strategy (Figure 3A). In the absence of a cumulative effect, biannual treatments are more beneficial both in the short and long terms in reducing microfilarial prevalence than annual treatments (Figure 3B). With the more conservative 16.5% reduction in female fertility per 6-monthly treatment, the initial benefit of microfilarial prevalence reduction is less pronounced than in the previous scenario, but again, there is relatively little difference in the long-term impact of biannual compared to annual ivermectin treatments (Supporting Information Text S3: Supplementary Figures, Figure S1).
Varying the therapeutic coverage in the overall population, and the proportion of systematic non-compliers had a large influence on the infection intensity achieved at the end of the 15th year of ivermectin distribution. An increased overall coverage, or a decreased proportion of systematic non-compliers lead to lower microfilarial loads 12 months after the 15th year of intervention (Figure 4). Under annual treatment, overall coverage had a larger effect on projected infection intensity (Figure 4A) and microfilarial prevalence (Supporting Information Text S3: Supplementary Figures, Figure S2A) than under biannual treatment (Supporting Information Text S3: Supplementary Figures, Figure 4B and Figure S2B). (Because of the nonlinear relationship between infection prevalence and intensity, the proportional reductions in prevalence are smaller.) For instance, under the assumption of a cumulative effect of ivermectin, and for a 5% proportion of non-compliers, increasing therapeutic coverage from 60% to 80% decreased microfilarial load by ∼50% for annual frequency compared to 16% for biannual frequency. The corresponding values when no cumulative effect was assumed were ∼37% and ∼30%. By contrast, the assumed proportion of systematic non-compliers had a more pronounced effect on the impact of biannual treatment delivery. Under the assumption of a cumulative effect of ivermectin, and for a 70% therapeutic coverage, decreasing systematic non-compliance from 5% to 0.1% decreased microfilarial load by ∼69% for annual frequency and by ∼97% for biannual frequency. The corresponding values when no cumulative effect was assumed were ∼23% and ∼53%.
Mathematical models can play a fundamental role in informing control programmes and strategies, but crucially, policy makers must realise that model outputs are highly dependent on implicit and explicit model assumptions [44]. Among the latter and for onchocerciasis in particular, the effects that (yearly or 6-monthly) ivermectin treatments exert on the reproductive biology of O. volvulus represent an area of considerable uncertainty, where further research is urgently needed. Although ivermectin's microfilaricidal effect is well established [17], the embryostatic effect and its repercussions on female worm fertility [18]; whether or not such effects on fertility are irreversible [25], [28]; the rate of resumption of microfilarial production [17]; and possible effects on intranodular sex ratios and insemination rates [45], [46], [47], remain poorly understood. An appropriate and updated incorporation of these effects into models, and an understanding of any enhanced macrofilaricidal activity of ivermectin under increased treatment frequency regimes [45], [47], [48], [49], are essential to reliably inform control policy, and fully assess ivermectin efficacy. Our results illustrate that the question of whether or not the drug effects on microfilarial production are cumulative, is highly influential on the projections of the long-term effect of annual or biannual MDA with ivermectin, particularly in areas with high baseline onchocerciasis endemicity.
The data that informed the model in [25], and presented in [24], comprised longitudinal microfilarial load follow up at various time-points after each of five annual treatment rounds in 74 individuals who received all five annual ivermectin doses from 1987 through to 1991 in an early community trial in the savannah focus of Asubende, Ghana [24]. The focus had been under vector control since 1986 during the OCP, and experienced a 70% reduction in parasite exposure during the trial despite antivectorial measures being interrupted for the first three years of ivermectin treatment. Figure 3 of [25] contrasts two model fits explaining the temporal trends in five annual data points of [24], corresponding to (decreasing) microfilarial counts just before each treatment round. The two hypotheses being tested to explain such trends are a null hypothesis of all—ivermectin-exposed—adult worms regaining their full microfilarial productivity vs. an alternative hypothesis of a 35% reduction in productivity with each treatment round. The authors of [25] concluded that the model assuming the alternative hypothesis provided a better fit to the data. However, given that: a) microfilarial loads were measured per skin snip instead of per mg of skin; b) the weight of a skin snip may range between 0.5 and 3 mg; c) lighter snips more likely yield a false negative result, and d) microfilarial counts originated from snips incubated for only 30 minutes in distilled water [24], [50] (likely to underestimate microfilarial load as microfilaridermia decreases), there is the possibility of considerable measurement error [5]. This is particularly important regarding the last two data points in the dataset (the most influential for discriminating between the two hypotheses), as for the last two years of the community trial in Asubende, the study area was receiving full vector control in addition to ivermectin, making it difficult to disentangle the effects of treatment from those of antivectorial measures. (The authors of [25] indicate, however, that the impact of vector control was taken into account in their model.) By contrast, the study in [27], based on the data presented in [26], which did not detect a cumulative effect of ivermectin on the production of microfilariae by female worms, used longitudinal data from 510 individuals (7 times as many as [24]), who took all five 6-monthly doses of ivermectin from 1998 to 1990 in the absence of vector control in Guatemala, with microfilarial loads measured per mg of skin after 24 h incubation [26].
Since our current model is deterministic, we cannot presently explore the probability of elimination. However, comparison of our projections with those of other models is informative. ONCHOSIM projections indicate that with a coverage of 80%, and an initial intensity of 70 mf/ss (in those aged 20 years and older), a minimum of 25 years of annual ivermectin distribution would be necessary to achieve a 99% probability of elimination [21]. In previous projections with the same model [23], the required duration of ivermectin distribution increases steeply and nonlinearly as heterogeneity in individual variation to vector exposure increases. Our model includes age- and sex-dependent exposure to vector bites [30] but does not consider inter-individual variation. The simulations in [21], [23] assume that ivermectin has a cumulative effect on the production of mf by female worms, and our results suggest that, in the absence of such an effect, ivermectin would have a less pronounced long-term impact. This indicates that if ivermectin does not have a cumulative effect on the fertility of O. volvulus, a longer duration of ivermectin distribution than previously estimated may be required to reach elimination thresholds, especially in areas with a high initial infection intensity and perennial transmission. In some areas of Cameroon that have received 13 years of ivermectin treatment, recent analyses of microfilarial dynamics do not support the operation of a strong cumulative effect of repeated treatments on the microfilarial productivity of female worms [51].
Comparison with provisional thresholds for elimination is also interesting. Operational thresholds based on [19], [21] suggest a microfilarial prevalence <5% in all of the sampled villages, or <1% in 90% of sampled villages. Our results indicate that microfilarial prevalence would remain above 5% after 15 years of annual or biannual treatment if ivermectin does not affect microfilarial production by O. volvulus cumulatively, even with a therapeutic coverage of 80% and only 0.1% of non-compliers (Figure 3B). Our hypothetical baseline infection levels were set at 70% microfilarial prevalence and >40 mf/mg in those aged ≥20 years, and the ABR to 19,000 bites per person per year, with perennial transmission. The baseline prevalence in the Senegalese/Malian foci reporting elimination ranged from mesoendemicity to the lower end of hyperendemicity (20% to >60%), and the CMFL from 10 to 48 mf/ss in 16 (27%) of the villages, with CMFL <10 in the remaining 44 (73%) of the 60 surveyed villages. In addition, transmission in these foci is seasonal as opposed to perennial, enhancing the impact of annual treatment on transmission when ivermectin is distributed just before the start of the rains; microfilarial loads are lowest during the transmission season and there are no blackflies around to ingest mf when these start reappearing in the skin [19]. Also, the difference with a biannual strategy would be less pronounced. These factors may have contributed to the feasibility of elimination in these areas and the reported lack of a significant difference between annual and 6-monthly treatment frequency. Likewise, in the foci located in Kaduna state, Nigeria, the median baseline prevalence was 52%, the median CMFL was 4 mf/ss, and transmission was also seasonal [20]. It should be noted that ONCHOSIM projections are consistent with current observations of elimination [19], [20], [21]. However, as described above, the areas where elimination has currently been achieved had lower baseline endemicity levels, and seasonal vector presence, leading to less transmission during inter-treatment periods. Under these conditions, assumptions of ivermectin effects on adult worms would likely have a lesser effect on models projections.
Our results are compatible with those of other modelling studies [52], which indicate that the higher the transmission intensity, the higher the necessary effectiveness of treatment (a net measure comprising coverage, number of treatment rounds per year and drug efficacy) to reach elimination. However, our study also emphasizes how different modelling assumptions can have profound effects on model outcomes and conclusions (a more extensive summary of the main structural assumptions of different onchocerciasis models is presented in [53]). This further highlights the need, discussed in [44] for helminth modellers to investigate key questions regarding helminth control more collaboratively, exploring the reasons for any disparity between the results of different models using the best available data.
Biannual ivermectin treatment was found to have a large additional benefit in both reducing microfilarial prevalence and intensity compared to annual treatment when no cumulative reduction in parasite fertility was assumed. When such effect was assumed, the model indicated that there would be an initial substantial benefit (as rates of microfilarial production are reduced quickly) of the biannual strategy, but that there would be relatively little difference in microfilarial prevalence at the end of the 15th year compared to annual treatment (Figure 3A). A possible reason for the pronounced difference between the two treatment frequencies, if ivermectin does not decrease worm fertility cumulatively, is that there would be substantially more transmission between annual than between 6-monthly treatments (distributing the drug every 6 months does not allow the adult worms to regain their fertility to a substantial level if there is perennial transmission, but there may be less additional benefit in seasonal transmission scenarios). Understanding ivermectin's effect on the reproduction and survival of adult worms [17], [18], [28], [29], [45], [46], [47], [48], [49] has important policy implications regarding switching to a biannual (or more frequent) treatment strategy in Africa. Three-monthly ivermectin treatments have contributed to acceleration towards local elimination in initially hyperendemic foci in Mexico [54].
Varying therapeutic coverage (for fixed non-compliance) had less effect on the impact achieved with a biannual treatment frequency than it had for annual distribution. This can be explained as the model accounts for the fact that if someone misses a treatment round, there is another chance to get treated during that year, ensuring that at least one annual treatment is received. In annual frequency, a missed treatment would result in a gap of at least two years between treatments, allowing microfilaridermia levels to build-up and contribute to transmission in the between-treatments period. This has implications regarding policy decisions in areas that have been found to have low coverage in the past, and highlights the potential benefit of switching to a biannual treatment strategy. In any case, a higher therapeutic coverage would prevent more disease during the intervention as the intensity of infection would decrease more rapidly. Incidence of blindness [55], and relative risk of excess mortality in sighted individuals [4], [5] depend on microfilarial load. It is also important to bear in mind that our model, at this stage, does not include the possibility of sub-optimal response or resistance to ivermectin or financial costs, in which case, the described benefits of a biannual treatment frequency could be very different.
Assumptions regarding the proportion of systematic non-compliers were found to be just as important as those for overall coverage when projecting the long-term impact of ivermectin distribution. The proportion of systematic non-compliance (for a fixed level of therapeutic coverage) was also found to have a marked influence on the impact achieved by a biannual strategy, particularly when assuming a cumulative effect of ivermectin; the higher the non-compliance rate, the smaller the benefit of biannual treatment. This indicates that the effect of systematic non-compliance may not simply be overcome by increasing treatment frequency and has implications when considering switching to a biannual treatment strategy, as two areas with the same overall coverage but different proportion of systematic non-compliers may lead to very different results regarding the feasibility of elimination [56].
As control programmes move towards elimination goals, the proportion of systematic non-compliers in the population becomes increasingly important. Studies of coverage and compliance for lymphatic filariasis treatment have indicated that, in addition to heterogeneity in transmission and vector density, and missed rounds of MDA, continuing transmission seems to be linked to rates of systematic non-compliance [56]. Therefore, when evaluating the progress of elimination programmes, the proportion of, and factors contributing to, systematic non-compliance should be investigated in addition to those determining overall coverage [36], [57], as an assessment of the latter on its own may mask reasons behind transmission persistence.
Modelling studies should also routinely vary the proportion of systematic non-compliers in addition to levels of treatment coverage as part of their sensitivity analysis to help understand the impact of prolonged treatment in populations. Although there are some data indicating that treatment compliance may depend on host age and sex (Brieger et al. found that older members of the community were more likely to take ivermectin than younger sections of the population, and men were more likely to comply than women in a Cameroon, Nigeria and Uganda multi-centre study [57]), further investigation regarding patterns of systematic non-compliance (i.e. the characteristics of individuals who never take the drug) will be essential to parameterise such modelling studies.
There is substantially more uncertainty surrounding model-derived projections of the long-term impact of, and feasibility of onchocerciasis elimination with ivermectin distribution than previously recognised. This uncertainty arises from an incomplete understanding of the effects of ivermectin on parasite survival, population structure, and reproductive biology, when the drug is administered at the standard dose annually, biannually (or more frequently, e.g. quarterly). Although the results presented in [45], [46], [47], [48], [49] would be invaluable to parameterise mathematical models incorporating such effects, further empirical and theoretical research is needed. Regarding the former, there is a need for well-characterized long-term (individual) longitudinal data (including previous treatment history), to estimate reliably the potential macrofilaricidal effects of ivermectin. However, to avoid the potentially confounding effect of ongoing transmission (which may lead to underestimating macrofilaricidal effects, particularly under annual treatment), studies could be conducted in areas where transmission has been interrupted (in geographical or ecological islands by elimination of the local vector [58], [59]). In areas near to elimination due to ivermectin distribution alone, rates of skin repopulation by mf could be investigated by fitting models to these data under a variety of ivermectin effects assumptions. Regarding the more theoretical aspects, a more adequate formulation of the parasite's mating probability in light of drug effects, decreasing male to female sex ratios [60], and changes in parasite distribution resulting from prolonged treatment [61] would also be important for assessing the feasibility of elimination.
Our results indicate that in areas with high baseline endemicity and perennial transmission, 15 years of annual or biannual treatment with ivermectin may not be sufficient to bring infection levels below potential elimination thresholds. Further incorporation of ivermectin effects into models; comparison of perennial vs. seasonal patterns of transmission; consideration of other O. volvulus–Simulium combinations; calibration of models for a wide range of baseline endemicity levels; assessment of patterns of treatment coverage and compliance; and inclusion of parasite genetic structure regarding sensitivity to ivermectin, will be essential to evaluate uncertainty surrounding model-derived projections. This, together with cost-effectiveness analysis, and development of stochastic frameworks will be crucial for informing control policy regarding annual vs. biannual treatment strategies in Africa, and for exploring the feasibility of elimination in foci with varying degrees of baseline endemicity. Finally, whether prolonged ivermectin treatment has a profound effect on the parasite's reproductive fitness has implications for the risk of ivermectin resistance evolving [35], and the risk of resurgence when treatment ceases. This highlights the importance of post-control surveillance in those foci where treatment is deemed to have been sufficiently successful to be stopped [62], [63], [64].
|
10.1371/journal.pgen.1002283 | Atypical AT Skew in Firmicute Genomes Results from Selection and Not from Mutation | The second parity rule states that, if there is no bias in mutation or selection, then within each strand of DNA complementary bases are present at approximately equal frequencies. In bacteria, however, there is commonly an excess of G (over C) and, to a lesser extent, T (over A) in the replicatory leading strand. The low G+C Firmicutes, such as Staphylococcus aureus, are unusual in displaying an excess of A over T on the leading strand. As mutation has been established as a major force in the generation of such skews across various bacterial taxa, this anomaly has been assumed to reflect unusual mutation biases in Firmicute genomes. Here we show that this is not the case and that mutation bias does not explain the atypical AT skew seen in S. aureus. First, recently arisen intergenic SNPs predict the classical replication-derived equilibrium enrichment of T relative to A, contrary to what is observed. Second, sites predicted to be under weak purifying selection display only weak AT skew. Third, AT skew is primarily associated with largely non-synonymous first and second codon sites and is seen with respect to their sense direction, not which replicating strand they lie on. The atypical AT skew we show to be a consequence of the strong bias for genes to be co-oriented with the replicating fork, coupled with the selective avoidance of both stop codons and costly amino acids, which tend to have T-rich codons. That intergenic sequence has more A than T, while at mutational equilibrium a preponderance of T is expected, points to a possible further unresolved selective source of skew.
| When considering a single strand of DNA, it is not necessarily the case that the frequency of each base should equal its complementary partner, such that A = T and G = C. For the leading strand, it is typically the case that Gs are more common than Cs, and Ts more common than As. This bias is widely thought to arise due to different mutational biases during replication. The Firmicutes exhibit an atypical preference for A over T on the leading strand, and here we show that selection, rather than mutation, can explain this exception. For those bases within coding regions, selection acts to inflate the frequency of A over T in order to avoid stop codons and to use metabolically cheap amino acids. Because genes are not orientated randomly, this manifests as an overall enrichment of A on the leading strand. Furthermore, a direct examination of mutational patterns is inconsistent with the observed enrichment of As. Curiously, our data also point to an unresolved source of selection on synonymous and intergenic sites, which are widely assumed to be neutral.
| Skews in nucleotide usage (compositional asymmetries) are of interest as they provide a window into fundamental processes operating within genomes. Under conditions of equal mutation bias and random gene orientation, the two complementary strands of a bacterial chromosome should be subject to the same sets of substitutions, and hence each should contain approximately equal amounts of a given base and its complement [1]. This condition, where A∼T and C∼G within a given strand, is known as the second parity rule and represents a null expectation of sequence evolution. The division of the replication fork into leading and lagging strands, however, has shaped bacterial sequence evolution contrary to this null, as each strand generally possesses an excess of one nucleotide over its complementary base (called GC and AT skews). Within bacterial genomes, nucleotide skews normally manifest as a richness of G over C and (with a lesser magnitude) T over A on the replicatory leading strand [2]–[5].
These genomic skews indicate some force, be it mutation or selection, is biasing substitutions between the two replicating strands. While it is acknowledged that, in theory, selection for genes to reside in the leading strand coupled with preferences for particular amino acids could result in chromosome-wide skews [2], [5]–[8], such a role for selection in generating large-scale compositional bias remains largely hypothetical and undescribed. Instead mutational biases between the two replicating strands are generally invoked as the cause of nucleotide skew [3], [8], [9]. Mutational differences between transcribed and non-transcribed strands have also been considered [10], [11], and these explanations incorporate a selective element as they require asymmetrically distributed genes between the replicating strands.
It has been argued that strand-specific mutation biases might result from the different amounts of time spent by each strand exposed in the single-stranded state during continuous or discontinuous DNA replication. While cytosine deamination (C→T) in particular was long suspected to play a major role in creating the excess of G and T in the leading strand, it has been shown that similar compositional skews can result from a variety of mutational scenarios [12]. The observation that GC skews tend to be stronger than AT skews also points to contributions from multiple mutation types. As would be expected if they are primarily mutational in origin, detected skews are generally higher in nearly neutral sites such as intergenic regions and fourfold degenerate sites [2], [3], [10].
Staphylococcus aureus is an unusual case in that, like other Firmicutes, it displays an excess of A over T in the leading strand, or positive AT skew given as (A–T)/(A+T) [13]. Why does this AT skew run counter to that observed in most bacteria? One possibility is that unique selective processes might be avoiding T and preferring A in the leading strand. Genes predominately lie in the leading strand in S. aureus, a feature of bacterial chromosomes posited to result from selection to minimize impacts between DNA and RNA polymerases [14] (although the relevance of this mechanism remains unclear). Any pressure to underuse codons rich in T could then result in AT skews simply due to the differential coding content of these two strands. Gene orientation bias is particularly enhanced in low G+C Firmicutes, potentially on account of the replication fork asymmetry induced by the possession of separate α subunits for synthesis of the leading and lagging strands [15], [16]. Alternatively, S. aureus might display a mutational bias which produces AT skew opposite that of most other bacteria, pushing up A over T in the leading strand. Indeed, it was recently suggested that the DNA polymerase-α subunit that replicates the leading strand also determines the direction of AT skew [16]. However, this finding was not repeated in a subsequent study and a direct mutational effect on AT skew resulting from α-subunit possession was called into question [11].
Here we investigate whether mutation or selection best explains the unusual AT skew in S. aureus. Dividing the chromosome into coding and non-coding positions allowed us to assess whether skew is strongest in those sites which should be under weaker purifying selection, such as intergenic and fourfold degenerate sites, or whether skew is most prevalent in non-synonymous sites which are constrained by the need to code for amino acids. Moreover we make use of newly described, high resolution genome-wide SNP data representing a single widespread clone of methicillin-resistant Staphylococcus aureus (MRSA) [17]. As these isolates have diverged from a very recent common ancestor, over a period of 4-5 decades, the data provide an opportunity to infer mutational patterns in S. aureus and contrast AT skews expected under mutational equilibrium to the AT skews observed. Importantly, the false positive rate of SNP calling in these genomes is benchmarked to be less than 1 SNP per genome (Julian Parkhill, personal communication), making these an unprecedentedly high quality resource.
As some of our results focus on coding sites within a single DNA molecule, the published strand, whereas others utilize coding sites in the sense direction on either the leading or lagging strands, we have provided a schematic to illustrate which sites are being considered in different types of analyses (Figure 1).
A plot of AT skew on the published strand in non-overlapping windows confirms that AT skew in S. aureus is unlike that of most bacteria as it is positive in the first half of the published strand (Figure 2), as previously described [13]. Considering only the core (vertically transmitted) without non-core (laterally transferred) regions eliminates irregularities in the AT skew which may arise from the importation of sequences which previously resided on an oppositely-skewed strand (Figure 2). For this reason the core genome only was considered in the rest of our analysis.
Three lines of evidence argue against mutation as the cause of AT skew in S. aureus:
Having established that mutation is not the primary cause of the observed AT strand bias, we sought to determine what selective forces might be responsible. The principle challenge appears to be to explain why AT skew is so profound at first sites in codons, even when compared with codon second sites. As stop codons start with T and cannot feature within the coding sequence, their avoidance provides a potential component of the unusual first site AT skew.
While we are able to make an a priori assumption regarding stop codon usage (since, by definition, they cannot be included within the body of a gene), we have no prior expectation concerning amino acid usage. It was therefore necessary to measure the AT skew resulting from biased gene distribution in S. aureus in the absence of selection on amino acid-encoding codons. To this end we simulated coding sequences preserving the discrepancy in coding content between the two replicating strands of S. aureus. For each simulation, the same number of codons as seen in a given replicatory strand were reconstructed based on the intergenic nucleotide frequencies within that strand, but with the caveat that stop codons were not permitted. Intergenic base frequencies were used to derive codons in order to control for any effects of the baseline nucleotide content as well as any mutational contribution to coding content within the chromosome. AT skew was then calculated in first and second sites for each of the 10,000 randomized coding sequences. These simulations quantified the AT skew expected to result purely from the avoidance of stop codons, indicating that randomized coding sequences display significant AT skew in first positions (Table 3). Thus, we would expect a lack of stop codons to contribute significant AT skew in first positions given such a discrepancy in coding content between the two replicating strands, even with a complete lack of selection on amino acid content. However, the magnitude of the effect owing to stop codon avoidance is unable to explain the full magnitude of the skew that we observe.
To explain the residual AT skew at first sites and all of the AT skew at second sites left unexplained by the avoidance of stop codons in reading frames, we investigated the possibility of further selection within coding sequences to decrease T. The mean codon usage in the randomized coding sequences (where codons are drawn randomly in proportion to the intergenic nucleotide frequencies within the same replicating strand) represents a null expectation of codon usage in the absence of any selection on amino acid content. Comparing this null with the observed codon usage in the TW20 chromosome allows for direct quantification of the over- or under-usage of a given codon (Z, see Methods). A positive Z-score for a given amino acid indicates that amino acid is more commonly used within the S. aureus genome than would be expected according to our null model of codon usage, while a negative Z indicates that amino acid is under-used. Such an approach reveals that T-rich codons are in fact highly under-represented in the S. aureus chromosome, with an enhanced avoidance of T in the gene-rich leading strand (Figure 6). What selective force could account for such a paucity of T in first and, to a lesser extent, second sites? We hypothesize that it might reflect unusual features of the amino acids that start with T.
T-starting amino acids mostly belong to the shikimate pathway and are thought to have been those most recently added to the genetic code [28]. As late-added amino acids tend to be expensive to manufacture [29] and shikimates in particular tend to have complex chemical structures, might it be that T avoidance reflects nothing more than selection against the use of costly amino acids? We find that there is a significant negative correlation between Z and amino acid cost as given in Akashi and Gojobori 2002 [30] (Figure 6), indicating that expensive amino acids are indeed under-used, thus accounting for some of the paucity of T. We have repeated this analysis using alternative cost measures [31] and obtained similar results for six out of eight cost schemas (Figures S2, S3, S4, S5, S6, S7, S8, and Table S2). The two measures (Rglucose and molecular weight) that do not provide significant correlations between Z and amino acid cost are perhaps expected to be less revealing: Rglucose correlates neither with previous cost measures nor with amino acid substitution rate, while molecular weight does not take into account metabolic networks relating to amino acid production [31].
We have shown that a lack of stop codons and avoidance of costly amino acids in asymmetrically distributed open reading frames can in large part account for the positive AT skew in S. aureus. Could similar mechanisms produce the unusual AT skews seen across other Firmicutes? Using phylogenetically independent contrasts (see Methods), we note that, among Firmicutes, an increase in the degree of gene strandedness from one species to another also results in a proportional increase in the extent of positive AT skew (Figure 7). It is therefore likely that the high degree of gene strand bias similarly explains the atypical patterns of AT skew in other Firmicutes.
As a contrast to the Firmicutes, we performed a similar analysis on phylogenies (Tables S3, S4, S5, S6, S7, S8) of the Gram negative Alpha-proteobacteria [32], Delta-proteobacteria [33], Epsilon-proteobacteria [34], Gamma-proteobacteria [35], as well as the Gram positive Actinobacteria [36]. Although the species sampled from these phyla tend to display typical (negative) genomic AT skews, it is possible that the degree of strand bias within these genomes nevertheless modulates the magnitude of these AT skews due to either avoidance of stop codons or selection on amino acid cost. The additional lineages do not however reveal any pattern of regression of ΔATskew on Δgespi, the latter indicating changes in gene strand bias (Figure 7). This is not unexpected since lack of large values of strand bias between terminal node pairs of the non-Firmicute phyla (resulting in lack of large differences in strand bias) means that the points all scatter around 0.
Recently an interest has emerged in whether certain sites in bacterial chromosomes commonly thought to be nearly neutral are in fact under selection as regards their nucleotide content. Two studies [23], [24] both used SNP profiles to estimate the GC content in possibly neutral sites at mutational equilibrium and showed the observed GC:AT bias greatly differs from that expected under the influence of mutation alone, consistent with previous reports of mutational pressure towards AT in E. coli [37]. What these studies were unable to explain, however, was what selective forces might be biasing nucleotide content at third codon and intergenic sites, leaving open the possibility that biased gene conversion and not selection might be acting. Here we also investigate a feature of bacterial chromosomes commonly presumed to be mutational, AT skew, and test whether mutation or selection is responsible for generating the unusual AT skews in S. aureus. Not only do we show that the atypical AT skew pattern in S. aureus is not due to mutational bias, but we are able to delineate to some degree what mode of selection is occurring (at least in terms of coding sequence), and to what end, in order to explain the observed skew pattern.
We find the mutational effect on AT skew in S. aureus (and in another Firmicute, B. anthracis), as derived from intergenic SNPs some distance away from coding sequence, to be inconsistent with, and poorly explanatory of, the observed base composition. Fourfold degenerate sites and intergenic regions display little skew, and intergenic SNP profiles do not support a replication-induced mutational origin of AT skew. In addition to fourfold sites, intra-operonic intergenic regions also display very weak AT skews, and hence any transcriptional effect is likely to be weak. Instead our results support a selectionist basis for compositional bias in S. aureus in which AT skew, the majority of which is observable at first and second positions in the sense direction, results from selection at both the translational level and on gene position. The avoidance of stop codons and codons encoding costly amino acids accounts for a substantial proportion of the skew in first and second codon positions because the majority of genes are on the leading strand. However, we are unable to accurately quantify the contribution of the avoidance of costly amino-acids, because the cost estimates used [31] are only approximate. Nevertheless, we can describe a relationship between the intensity of selection against costly amino acids and the magnitude of skews (Figures S9, S10, and Table S9). Codons encoding more costly amino acids tend to be AT-rich [29] and we observe that, on average, AT-rich genomes encode more costly amino acids (Figure S9A). However, the average cost of amino acids in AT rich genomes while high, is not as high as expected given the AT pressure (Figure S9B). This we interpret as evidence for more efficient selection against costly amino acids in GC-poor strains, which in turn contributes to a higher AT skew (Figure S10).
We further show a phylogenetically controlled positive association between the extent of gene strand bias and positive genomic AT skew across the Firmicutes, indicating that strand bias is likely responsible in part for the atypical AT skews seen across this phylum. Our failure to detect such a relationship in non-Firmicute phyla may in part be due to a lack of genomes in these phyla with very high strand bias, leaving only increases in strand bias of smaller magnitude to investigate and thus much noisier data sets (Figure 7). The pattern of Δgespi versus ΔATskew observed for the Firmicutes is similar to the patterns observed in other phyla when considering the region 0>x<10 (Figure 7), meaning large differences in strand bias between terminal node pairs are required to be able to detect a relationship between the two quantities. This may be why we only see an effect in the Firmicutes, where strand bias is high enough to leave a clear impact upon the magnitudes of genomic AT skews. We conclude that if there is a relationship between gespi and AT skew in non-Firmicutes, our method is not sensitive enough to detect it.
Our results leave several mysteries. First, why do species differ in the degree of strand bias and why is it so high in many Firmicutes? These issues remain enigmatic. A simple model supposes that in fast replicating species the chance of DNA and RNA polymerases colliding must be higher than in slow replicating species. There is, however, no correlation between growth rate and gene strand bias [38]. Rather the higher biases are typically found in chromosomes containing two different (possibly strand-dedicated) DNAP α-subunits at the replication fork which may render them more vulnerable to polymerase collisions [15]. It has also been suggested that strand bias reflects gene essentiality rather than the level of expression [39] although, again, this does not explain the unusually high level of strand bias in Firmicute chromosomes.
Further, while both the observed AT skew in non-coding sites and the pattern of SNPs in intergenic sequence cannot explain the skew seen across the leading strands as a whole, the two approaches are also inconsistent with each other. The relative mutation rates calculated from intergenic SNPs indicate that mutation is acting to bias T over A in intergenic sites, which is the typical direction that AT skew takes in most (e.g. many non-Firmicute) bacteria, suggesting less variable skew-related mutational profiles among bacteria than is commonly assumed. However, intergenic sites on the leading strand have a weak bias in the opposite direction. Such a leading strand bias is consistent with leading strand coding sites also showing slightly higher bias than lagging strand coding sites (Table 1).
What could account for the discrepancy between mutational biases and observed base frequencies at putatively neutral sites? One possibility is that these sites are not yet at mutational equilibrium. This could occur if, for example, there were until recently some unannotated small protein coding genes in the “intergene” spacer. These new pseudogenes would take an appreciable time to reach mutational equilibrium and could well leave a trace of A>T skew if they tended to be on the leading strand. However, in this case it is curious that intergenic skew, intra-operonic skew and skew at four-fold degenerate sites all show a weak A>T bias. An alternative is that the weakly positive intergenic AT skews could reflect ongoing selection. One possibility is that there exists unannotated coding sequence, which, if enriched on the leading strand, would contribute a net A>T skew. Neither missing gene model can explain why skew at four fold degenerate sites is of the same magnitude as in putative intergene spacer. In addition, if mutation alone dictated intergenic AT skews, leading intergenic spacers should skew to roughly -0.4 (Table 2) according to our estimates of mutational equilibria. Given that the average leading AT skew in S. aureus coding sequence is approximately 0.1 across all three codon positions (Table 1), the vast majority of intergenic spacers would need to be unannotated protein-coding sequence in order for missing genes to be able to explain the observed leading intergenic AT skew of 0.0276 (Table 1), a highly untenable scenario. Removal of the few regions with outlier AT skew values does not substantially impact the intergenic AT skew (Figure S11), suggesting that even if we are missing some genes their contribution to skew cannot explain the overall bias. What exactly is generating weakly positive AT skews in leading intergenic regions remains a mystery, but this analysis adds weight to the growing evidence [see e.g. 19,23,24,37,40] that “neutral” sites in bacterial chromosomes may not be quite so neutral after all.
The complete annotated genome of S. aureus subsp. aureus TW20, accession number FN433596 [41] was downloaded from the EMBL Nucleotide Sequence Database (http://www.ebi.ac.uk/embl/). Data analysis was carried out using Tcl and Perl scripts and statistical analyses were done in R 2.9.0 [42]. Core and non-core regions were delineated as in Harris et al. 2010 [17]. Coding sequences labeled as gene remnants or pseudogenes were excluded and both known and putative genes were considered. As we considered intergenic regions to be indicative of mutational pressures, all intergenic regions were subject to a length restriction of 500 bp to decrease the possibility of unannotated genes, and 60 bp were trimmed from each end of all intergenic regions, as these regions display distinct AT skew patterns deviating from that induced by replication alone (see Results and Figure 4). This is most likely explained by ATG initiation context definition (at the 5′ end) and termination sequences (at the 3′ end).
Distinguishing which intergenic regions lie within operons should help reveal the contribution of transcriptionally-induced mutation to AT skew as such regions should be more likely to be transcribed. The operon structure of S. aureus strain MSSA476 [43] was used to deduce operons within strain TW20. Operonic protein-coding and RNA genes in MSSA476 were extracted from NCBI RefSeq NC_002953 (http://www.ncbi.nlm.nih.gov) and were matched via BLAST 2.2.24 against all TW20 protein sequences and gene-encoded RNA sequences respectively. Orthologous operonic genes in TW20 were taken to be those with at least 90 percent identity and an e-value of less than 0.0001. Due to the close evolutionary relationship between the two strains most matches were unambiguous. Matches to pseudogenes and genes in the TW20 non-core chromosome were excluded. TW20 operons were then deduced from these orthologs with the additional caveat that genes within a putative operon be adjacent and transcribed in the same direction. In several cases a gene not contained in an MSSA476 operon was detected inserted into the intra-operonic TW20 intergenic sequence but on the strand opposite to the operonic genes. In these cases the operon was included.
To determine whether the observed AT skew deviates from that expected under mutation alone, we require an estimation of nucleotide content, and the resulting AT skew, at mutational equilibrium. 140 singleton SNPs from core ex-operonic intergenic sites at least 60 bp away from a gene boundary were isolated from what is currently the largest SNP dataset for any bacterial species, that comprising 63 S. aureus ST239 isolates [17]. In order to check for the possibility that some of these SNPs were called erroneously, we went back to the original read data for each of these 140 SNPs individually. These data revealed an average of 20-fold coverage, a maximum of 46-fold coverage, and a minimum (in 4 SNPs) of 12-fold coverage (Table S10). Furthermore, in 126/140 (90%) of cases, the assigned SNP was consistent in all mapped reads. Of the 14 remaining SNPs, a single inconsistent read was noted in 13 cases, and two inconsistent reads noted in one case. Given a sequencing error rate of 0.5% per sequencing reaction, the maximum probability that any SNP has been assigned by error (that is called consistently in at least 12 reads) is of the order of [(0.005×0.3′)∧11)] ≈ 2×10−31. Thus analysis of singleton mutations is an excellent indication of new mutations and does not reflect sequencing errors (see also Results).
These SNPs were used to estimate the mutational profile of S. aureus in ex-operonic intergenic sites. Singletons are SNPs which are seen only once throughout all sequenced isolates. Such SNPs are more likely to represent recent mutational events which selection has not yet had time to act upon, and thus only singletons were considered in order to orientate the direction of changes and minimize the possibility of selection or multiple hits. As all other lineages (aside from that with the singleton) have the same nucleotide at the given location, the assignment of the ancestral state is unambiguous.
As we find the sample size of 13 SNPs falling within intra-operonic intergenic regions too small to calculate the relative mutation matrix for intra-operonic sites, we considered SNPs in ex-operonic intergenic sites only. On both strands singletons were isolated in intergenic sites outside operons and relative rates of mutation were calculated for the leading and lagging strands separately. Nucleotide frequencies at mutational compositional equilibrium were derived from the relative mutation rates by considering that at compositional equilibrium, the loss of any given nucleotide must equal the net gain of that nucleotide at other sites:where f(i) is the frequency of site i and rij is the rate of change from i to j per site i as measured in the extant sequence. The above equilibrium equations were solved simultaneously using Maxima 5.21.1 [44] to yield equilibrium nucleotide frequencies. These equilibrium frequencies were used to calculate the AT skew in ex-operonic intergenic sites expected to result purely from replicational mutation at compositional equilibrium.
Similar mutational equilibrium analyses were performed on polymorphism data from B. anthracis and S. typhi. Intergenic singleton SNPs were extracted from alignments of 18 fully and partially sequenced B. anthracis strains [23] and from the intra-haplotype or haplotype-specific age groups for S. typhi SNP data [45]. For both organisms, only intergenic regions under 500 bp were considered and singletons were only called when sequence data was available for all strains and the SNP at least 60 bp away from a gene. Observed intergenic nucleotide content and AT skew were calculated using NCBI RefSeqs NC_003997 and NC_003198.
To obtain an approximate measure of the robustness the sign of the equilibrium AT skew indicated by the singleton SNP populations, the intergenic (ex-operonic in the case of S. aureus) SNPs were bootstrapped. For each species, the intergenic SNPs used to compute the mutational equilibrium were resampled with replacement 1000 times, and the equilibrium state recalculated as above after each resampling, to yield 95% bootstrap intervals for the equilibrium AT skew estimate.
Selection against stop codons within asymmetrically distributed genes could necessarily impose some amount of AT skew as T might be underrepresented relative to A within first codon sites. We wished to measure the AT skew which results in S. aureus from selection on gene position alone while preventing any selection on amino acid content, which might further increase or decrease the amount of T relative to A within genes in S. aureus, from biasing this measurement. Randomized coding sequences provide a means of estimating the AT skew that would result from the biased gene orientation seen in S. aureus even under a complete lack of selection for amino-acid usage. As both GC content and replication-associated mutational biases can modulate the amino acid content of proteins [46], [47], the baseline nucleotide frequencies of the leading and lagging strands of the TW20 chromosome could favor the presence of certain codons while disfavoring others. A null was devised in which nucleotides were sampled in proportion to their frequency in intergenic regions, which should be neutral or weakly selected, thus controlling for the baseline nucleotide content of the genome as well as any mutational effect on skew. 10,000 protein-coding sequences containing the same number of amino acid-encoding codons as in the leading and lagging strands of the TW20 chromosome were simulated using codons derived from the intergenic nucleotide frequencies in the relevant strand of the TW20 chromosome. Stop and start codons were excluded from randomized sequences. Amino acids with six codons were considered as two separate amino acids—a 4-block and a 2-block—since the frequency of individual nucleotides could differentially influence the usage of these two codon blocks. The resulting AT skew in randomized chromosomes was calculated in first and second sites as (A−T)/(A+T) with respect to the sense direction.
Selective patterns of amino acid usage may, depending on the nucleotide frequencies of the codons involved, also shape AT skew. Determination of whether individual amino acids are over- or under-used in relation to the above null is reflected in the Z score for each amino acid (aa):where the expected usage is the mean usage of that amino acid amongst the 10000 simulated coding regions, the observed usage is that seen in the TW20 chromosome and the standard deviation is that observed through the randomizations. This normalizes for variance seen due to amino acids occupying differing amounts of codon space, controls for the effect that genomic GC content may have on individual codon usage, and allows for comparison of over- or under-usage across different amino acids.
If gene strand bias is responsible for positive AT skews not just within S. aureus but across the Firmicutes, we expect a positive association between gene strandedness and genomic AT skews across the phylum. A simple test for correlation between these two quantities across a wide sampling of Firmicute species might, however, falsely infer a relationship between the two due to over-representation of sequence information in closely related genomes. We therefore investigated the relationship between strand bias and genomic AT skew using phylogenetically independent contrasts. Differences in strand bias and leading genomic AT skew were calculated for phylogenetically independent pairs of terminal node species in a phylogeny of Firmicutes [48] (Table S3) with the expectation that if strand bias does dictate the extent of positive AT skew, an increase in strand bias between species should also result in an increase in AT skew. Gespi values, calculated according to de Carvalho & Ferreira 2007 [49], were used as indicators of the degree of strand bias among these species, with a higher gespi indicating a greater degree of strandedness. As a counterpoint to the Firmicutes, similar analyses were performed on phylogenies of the Gram negative Alpha-proteobacteria [32], Delta-proteobacteria [33], Epsilon-proteobacteria [34], Gamma-proteobacteria [35], and a phylogeny of the Gram positive Actinobacteria [36] (Tables S4, S5, S6, S7, S8).
|
10.1371/journal.pcbi.1002721 | Probability Fluxes and Transition Paths in a Markovian Model Describing Complex Subunit Cooperativity in HCN2 Channels | Hyperpolarization-activated cyclic nucleotide-modulated (HCN) channels are voltage-gated tetrameric cation channels that generate electrical rhythmicity in neurons and cardiomyocytes. Activation can be enhanced by the binding of adenosine-3′,5′-cyclic monophosphate (cAMP) to an intracellular cyclic nucleotide binding domain. Based on previously determined rate constants for a complex Markovian model describing the gating of homotetrameric HCN2 channels, we analyzed probability fluxes within this model, including unidirectional probability fluxes and the probability flux along transition paths. The time-dependent probability fluxes quantify the contributions of all 13 transitions of the model to channel activation. The binding of the first, third and fourth ligand evoked robust channel opening whereas the binding of the second ligand obstructed channel opening similar to the empty channel. Analysis of the net probability fluxes in terms of the transition path theory revealed pronounced hysteresis for channel activation and deactivation. These results provide quantitative insight into the complex interaction of the four structurally equal subunits, leading to non-equality in their function.
| The activation of a receptor protein by a small molecule (ligand) can be quantified by Markovian models. These models, which are widely used in natural sciences, consist of distinct states and transitions between them. In nature receptor proteins are often formed by the assembly of more than one subunit and each subunit can bind a ligand on its own. In such a multimeric receptor protein the translation of the ligand binding into receptor activation is more complex because the subunits interact. This usually limits the application of Markovian models. HCN2 pacemaker channels are tetrameric ion channels that mediate electrical rhythmicity in multiple brain and peripheral neurons and in specialized heart cells. The channels are modulated by cAMP binding to each subunit. We were recently successful to quantify ligand-induced activation for these channels by a complex Markovian model and a full set of rate constants. Herein we applied the transition path theory to further analyze the identified Markovian model, and we quantified time-dependent probability fluxes within the model. Our results provide unprecedented insight into the complex interaction of the four structurally equal subunits of a presumably fourfold symmetric channel that leads to pronounced non-equality of the subunit function.
| In a protein the time scale of conformational changes ranges from picoseconds for the thermal vibration of the atoms to seconds, or even more, for structural arrangements [1]. Since proteins consist typically of hundreds to several thousand amino acids and each amino acid consists of seven or more atoms, the complete space of energetically possible states is huge, and certainly too huge for the experimentalist to study. It turned out, however, that often a set of a great many of these states can be viewed as metastable. This means that a protein typically fluctuates within a set of multiple different structures for a very long time if compared to the picosecond time scale of atom vibrations. Only extremely rarely the thermal energy suffices to leave a metastable set of states to reach another metastable set of states. A striking proof for this concept has been the discovery of the digital behavior of ion channels: Their pore is either closed or open, with dwell times often in the millisecond or even second time range [2]–[4].
Over the past decade there has been considerable progress in studying protein function theoretically by molecular dynamics (MD) simulations [5]. However, the maximum time range of presently possible MD simulations is still too short to become useful for the description of ion channel function. Kinetic analyses of ion channel gating are therefore presently conducted in terms of Markovian models, thereby treating a metastable set of states as one state and specifying transitions between two of these states according to Eyring's transition state theory [6]. Often, however, the accuracy of the experimental data does not suffice to differentiate between plausible models of sufficient complexity, e.g. to describe the action of three or more subunits. As a consequence the models used are often implausibly small or many simplifying assumptions for the equilibrium or rate constants have to be adopted.
Markovian models frequently used to describe the action of ion channels are either of the sequential or the cyclic type. Among the cyclic models the Monod-Wyman-Changeux (MWC) model [7] is of outstanding relevance because of its simple elegance: It assumes that a protein exists in two global conformations, taut and relaxed, and that each binding step systematically shifts the taut-relaxed isomerization and increases the binding affinity. This model, which has been most widely used to describe the sigmoidal oxygen-binding curve to hemoglobin (for review see [8]), has also been applied to describe the gating of ion channels by ligands, as e.g. the nicotinic acetylcholine receptor [9], cyclic nucleotide-gated (CNG) channels [10] or hyperpolarization activated cyclic nucleotide-modulated (HCN) channels [11]. However, as elegant as this model is, it does neither describe all aspects of the oxygen binding to hemoglobin [12], [13] nor is there any evidence that it fully describes the action of any ion channel [14]. For nicotinic acetylcholine and glycine receptors, the analysis of single-channel recordings has allowed the investigators to determine models with substantially more free parameters than typically employed by the MWC model. As a consequence, also the type of cooperativity differed substantially [15]–[18]. However, in all these approaches the ligand binding affinity could only be determined indirectly by the global fit of the channel activity, i.e. by the pore action, but not by direct measurement.
For CNGA2 and HCN2 channels we recently developed a strategy to simultaneously measure ligand binding and activation gating by combining confocal microscopy with patch-clamp fluorometry [19], thereby employing a fluorescently labeled ligand [20], [21]. In a subsequent study on HCN2 channels, pre-activated by a voltage pulse to −130 mV, we combined this approach with the method of concentration jumps. By globally fitting multiple time courses of ligand binding and channel activation, this approach allowed us to determine the equilibrium and rate constants in a Markovian model with 4 binding steps in both the closed and the open channel and 5 closed-open isomerizations (Fig. 1A) [22]. The analysis revealed pronounced cooperativity with respect to the microscopic binding affinity in the surprising sequence ‘positive – negative – positive’ for the binding of the second, third, and fourth ligand, respectively. Moreover, we considered the population of all closed and open states as function of time when jumping the ligand concentration. As a result, the total open probability is dominated by open states with either zero, two or four ligands bound whereas states with one or three ligands bound are only transiently populated (Fig. 1B and C) [22].
Herein we analyze the transition pathways in the C4L-O4L model [23], [24] when disturbing an equilibrium by a sudden change of a parameter. In our case this sudden change is either the application or the removal of a saturating ligand concentration. As a result the probability fluxes as function of time and the net probability fluxes are quantified, thereby identifying relevant and irrelevant transition pathways in channel activation and deactivation.
To specify the probability flux for any transition, we first consider a simple model with the states A and B and the two rate constants, k1 and k−1, specifying the transition rates.(Scheme 1)A and B can also be interpreted as probabilities (with A+B = 1) which change under non-equilibrium conditions with time. At any time, a net probability flux density, f, can be defined by the sum of two opposed unidirectional fluxes. For example, the net probability flux density from A to B is given by fAB = k1A−k−1B. If applying this to the individual transitions in the C4L-O4L model (for rate constants see Fig. 1D, Table S1), the net probability flux density of all transitions can be calculated at each time. In case of the binding reactions the rate constants have to be multiplied by the actual ligand concentration (either 0 or 7.5 µM). The arithmetic signs were chosen such that a probability flux is positive for the binding and opening reactions and, accordingly, negative for the unbinding and closing reactions. Because the probability of the states is time-dependent, the net probability flux density of each transition is also time-dependent.
First, the net probability flux density is considered when stepping from zero to 7.5 µM fcAMP. For the four binding steps in the closed channel, the net probability flux density is given by(1)L is the ligand concentration. As expected, the net probability flux density moves like a wave from C0 to C4 (Fig. 2A, left): fC0C1 ceases after less than 100 ms and fC1C2 after about 500 ms. fC2C3 and fC3C4 are slower and not finished after 1s yet. The net probability flux density from the closed states to the respective open states contributes to the reduction of the net probability flux density between the closed states.
Accordingly, the net probability flux density for the four binding steps of the open channel, that has been pre-activated by voltage, is given by(2)
The respective time courses of fOx−1Ox are basically similar to those of the closed states but larger in amplitude (Fig. 2B, left). For the closed-open isomerizations the net probability flux density is given by(3)
These time courses have a robust amplitude for fC1O1, fC3O3, and fC4O4 and are negligible for fC0O0 and fC2O2 (Fig. 2C, left). The obstruction of the pathway C2→O2 tells that the double liganded channel has a similarly taut structure as the non-liganded channel. The reduced degree of determinateness for fC3O3, and fC4O4 will be considered below.
Second, the net probability flux density of stepping from 7.5 µM fcAMP back to zero is considered. The respective time courses for the unbinding steps in the closed and open channel are given by equation (1) and (2), respectively, by setting L = 0 (Fig. 2A and B, right). The time courses for the open-closed isomerizations are given by equation (3) (Fig. 2C, right). The result is that the net probability flux density is much bigger in the unbinding steps of the open than of the closed channel and that fO3O4 and fO2O3 are much faster than fO1O2 and fO0O1, i.e. the state O2 is a severe obstacle in the net probability flux on the way to lower liganded states. Concerning the open-closed isomerizations, fO4C4 is negligible and fO3C3 is rapidly finished after about 1 s. In contrast, the other net probability fluxes are much slower. However, they all notably contribute to the closing process despite the rate constants for closing, kO2C2, kO1C1, and kO1C1, differ substantially. This shows how important it is to consider the model as a whole, including the rate constants and the time-dependent population of the states.
From the net probability flux densities, fXY, one can easily compute the total net probability fluxes, FXY, as the time integral over the time interval from the concentration jump (t = 0 s) to an end time, tend,(4)tend is either 5 s for the fcAMP pulse or 20 s after removal of fcAMP. Adapted to our C4L-O4L model, FXY indicates how much of the total net probability flux moves along a transition X↔Y. The main results are: (1) The total net probability flux is bigger in the binding steps of the open than the closed channel in both the presence of fcAMP and after its removal (Fig. 3A and B). (2) The total net probability flux between the closed states is larger for the binding than for the unbinding process (Fig. 3A). (3) FC0O0 is negligible in the binding-induced relaxation but significantly present in the unbinding-induced relaxation whereas FC4O4 is negligible in the unbinding-induced relaxation but significantly present in the binding-induced relaxation (Fig. 3C). Together, these results suggest different pathways for activation and deactivation.
To gain a more thorough insight into the transition pathways in our C4L-O4L model, they were analyzed according to the principles of the transition path theory [23], [25]. In this type of analysis one is interested in trajectories between two selected states within the model. This implies that all repeated jumps between two neighbored states in both directions are not counted; i.e. considered are only transitions in one direction. This information is hidden in the total net probability flux for the individual transitions, FXY, which we determined in the previous section. In any model the amount of net flux produced by one set of states must equal the amount of net flux collected by the other set of states. When applying the ligand, the only flux producers are the states C0 and O0. Because the applied ligand concentration was saturating, the main flux collector state is O4 while O3 and O2 also contribute but to an only small extent (Fig. 1C). After removing the ligand, the main flux producer state is O4, and to a small extent O3 and O2. The main flux collector states are C0 and O0. The states C2 and O2 are, to a minor extent, also flux collector states because after 20 s the steady state was not reached (Fig. 1B and C).
In case of a fully reversible Markovian model, which holds for the C4L-O4L model, the total probability flux can easily be decomposed into the pathway net probability fluxes, and no cycles remain because the condition of the detailed balance is fulfilled [22]. The procedure is to choose first a pathway from state X0 to state Xk along the states Xi. Then for this chosen pathway, X0→Xk, the net probability flux Fp,X0Xk is computed according to(5)Fp,X0X is then removed from the flux along all edges of this pathway and the procedure is repeated for the remaining pathway net probability fluxes until the total possible probability flux between these states is zero [25]. It should be noted that the sequence of such a decomposition is not unique because different paths can be chosen. It has become useful to start with the strongest pathway [25].
Let us first consider the pathway net probability flux after applying the ligand at 7.5 µM fcAMP (Fig. 4A). The pathway net probability flux from C0, one of the two flux producers, to O4, the main flux collector, results in five defined pathway fluxes, Fp,, which are, according to equation (5), given by the respective minimum values of FCxOx (x = 0…4). The pathway net probability flux from O0, the other flux producer, to O4 has only one available path because the fluxes of all closed-open transitions are directed to the respective open states. Illustration of the weights of the pathway net probability fluxes in the C4L-O4L model after applying the ligand leads to the result that ligand-induced channel opening proceeds mainly from C3 and C4 (with the uncertainty of the exact attribution; Fig. 3, left), and additionally from C1, but not relevantly from C0 and C2. The same type of analysis was then performed for the pathway net probability flux after removing the ligand from O4, the main flux producer, to C0 and O0, the main flux collectors. Closing of the channels proceeds predominantly from O1 and O0, and to a minor extent from O2 and O3, but not relevantly from O4. Together these results show that there is pronounced hysteresis for ligand-induced activation and deactivation. The pathway flux from O4 to O0, the other flux collector, has again only one available path because the fluxes of all closed-open transitions are directed to the respective closed states.
To demonstrate the function of our model at subsaturating fcAMP concentrations we considered the fluxes to the two main collectors at 0.75 µM fcAMP (O2 and O4; Fig. 5 A,B) and 0.075 µM fcAMP (O1 and O2; Fig. 5 C,D) and the respective reverse fluxes when removing fcAMP. Notably, these selected fluxes are only the dominating fluxes. The results show that at the intermediate concentration of 0.75 µM fcAMP activation proceeds along C1→O1 predominantly to O4 (Fig. 5 A) and to a minor extent to O2 (Fig. 5 B). In addition there is a remarkably big net probability flux in the open channel along O0→O2 and in the reverse direction along O2→O0 which is absent along O0→O4 and O4→O0, respectively. This indicates that O2 is a metastable state. Moreover, this result corresponds to the high energy barrier for the transition O2→O3 [22]. At 0.075 µM fcAMP the predominant activation proceeds along C1→O1 to O2 but not anymore to O4. O1 is only passed in the activation pathway. However, it is of importance as a collector for the probability flux in the open channel (Fig. 5 D).
In our considerations on probability fluxes so far the focus was set on net probability fluxes. If a flux is reversible, as in case of ligand binding and closed-open isomerizations, a net flux consists of two opposed unidirectional fluxes. Generally, very different opposed unidirectional fluxes can cause the same net flux. Large opposed unidirectional fluxes indicate conformational flexibility between two states, whereas small opposed unidirectional fluxes indicate conformational tautness between two states.
For the principal conformational change of the closed-open isomerization we considered the time courses of the unidirectional probability flux densities, fU,CxOx>0 and fU,OxCx<0 (x = 0…4), at the saturating fcAMP concentration of 7.5 µM fcAMP and related them to the net probability flux density, fCxOx (Fig. 6; c.f. Fig. 2C), according to(6)For the time after application of fcAMP, these time courses show that the transitions C0↔O0 and C2↔O2 are functionally irrelevant. In contrast, the net probability flux densities in the transitions C1↔O1 and C3↔O3 are robust and the unidirectional probability flux densities exceed the respective net probability flux density substantially. An extreme surplus of the unidirectional probability flux densities with respect to the net probability flux density is inherent in the transition C4↔O4, suggesting pronounced conformational flexibility when the channel is fully liganded. It is also notable that the empty activated channel (Fig. 6, top left) is much less flexible in this sense compared to the fully liganded channel (Fig. 6, bottom left).
After removal of fcAMP (Fig. 6, right), the fully liganded state is left rapidly (bottom) and the triple liganded state is rapidly passed. For the remaining transitions C2↔O2, C1↔O1, and C0↔O0 both the maximum net and the maximum unidirectional probability flux densities are much smaller than for C4↔O4 and C3↔O3. The finally large total net probability flux in the transitions C2↔O2, C1↔O1, and C0↔O0 (c.f. Fig. 3C, right) is only reached after many seconds. Remarkably, in the transition C2↔O2 the unidirectional probability flux density in the opening direction approximates zero, which is due to the low occupancy of C2, resulting in a close similarity of the unidirectional closing and the net probability flux density. This contrasts to C1↔O1 and C0↔O0 which both show larger unidirectional probability flux densities than the respective net probability flux density. Both the net and the unidirectional flux densities in the transition C0↔O0 reach finally the respective initial values before applying fcAMP (Fig. 6A, top left). The results in Fig. 6 also suggest that the binding of two ligands reduces the conformational flexibility of the closed-open isomerization maximally and, more generally, how different the effects of the four binding steps on the closed-open isomerization are.
We present a detailed analysis of probability fluxes within a Markovian model describing ligand-induced activation of HCN2 channels. The analysis is based on the combined recording of ligand binding and channel activation [22]. The main results are that significant activation of the channel proceeds with one, three, and four ligands bound whereas significant deactivation proceeds relevantly with two, one and zero ligands bound. The consequence of this result is a pronounced hysteresis for channel activation and deactivation. In addition to this we show that the channel is in a flexible conformation with one, three, and four ligands bound whereas it is in a taught conformation with zero and two ligands bound. Notably, we herein considered at each degree of liganding probability fluxes for the whole channel, i.e. we did not use specific stoichiometric factors. This allowed us to avoid any further assumptions concerning equivalence or non-equivalence of the available binding sites.
The high degree of determinateness of our approach was certainly not only caused by analyzing data of ligand binding and gating simultaneously, but also by the specific nature of the effect induced by the concentration jumps: Applying the ligand abruptly transformed the simple C0↔O0 model into the complex C4L-O4L model. Conversely, removing the ligand emptied the C4L-O4L model and led to the initial C0↔O0 model. Only few parameters were not determined. The limited time resolution of our method did not allow us to distinguish some of the rates with three and four ligands bound. It should be emphasized, however, that none of the present conclusions depends on specific assumptions in the fit.
For analyzing the action of proteins great progress has been achieved over the past years by molecular dynamics (MD) simulations and building energy landscapes. In these landscapes often a small set of energy basins can be identified that are separated by energy barriers (for review see [26]). However, present MD simulations are typically limited by the available computer technologies to one microsecond [5] or, very recently, even to a millisecond [27]–[29], thereby already touching the time range of the observables in ion channel gating. In contrast, functional measurements are real and they provide a different kind of information about the action of ion channels compared to MD simulations: The pore opening can be monitored by the ion current, distinct conformational changes by changes in the fluorescence intensity of introduced labels [30]–[32] and, in case of voltage-gated channels, the movement of the gating apparatus by gating currents [33]–[35]. Markovian state models are often used to interpret functional data.
Over the past decade there has been great progress in applying the transition path theory in combination with Markovian state models to learn more about the action of proteins. However, this was done only in theory by MD simulations and in the mentioned short time range [23], [25], [36]. Also graphical visualization of the paths has been performed [37]. So far, the transition path theory has not been applied to the gating of ion channels as performed herein.
Our approach quantifies at which degree of liganding the closed-open isomerization proceeds and demonstrates a complex type of hysteresis in channel gating (Fig. 4). Though this information is, of course, determined by the rate constants, the complexity of the C4L-O4L model precludes an immediate identification. Such a hysteresis might have physiological consequences for the regulation of the channels. For example, molecules regulating the activity of these channels might have different affinities at different conformations. Also, knowledge of the transition pathways might become very helpful for future strategies to develop drugs stimulating or blocking these channels. For example, if one intends to reduce the cAMP effect on the open probability of HCN2 channels in a living cell, it might be a good idea to selectively affect the C1↔O1 transition because it is this transition which mediates the main effect in the lower concentration range.
In addition to the sum of the probability fluxes determined by the transition path analysis, our approach provided us dynamic information: the net and unidirectional probability flux densities. The net probability flux density shows the flux for the different transitions as function of time, i.e. how often per time interval a net transition appears (Fig. 2). There are two important aspects of this information: First, these time courses show the maximum intensities of the transitions. Second, these time courses show how differently rapid these transitions are. This might be also of interest for the development of highly specific drugs modulating the channel activities. Moreover, this knowledge might be of relevance for learning how the subunits interact, how the identified complex cooperativity [22] is generated by subunits that are equal in their amino acid sequence.
The unidirectional probability flux density (Fig. 6) provides another type of information: It specifies all transitions from one state to a neighbor state, including also the repetitive transitions. The unidirectional probability flux density can be much bigger than the net probability flux density. Hence the unidirectional probability flux density provides information how taut or relaxed an HCN2 channel is at a given degree of liganding. The most challenging result is that the tautness is not a monotonous function of the degree of liganding, it is high when none or two ligands are bound and low when one, three and four ligands are bound (Fig. 6, left). In analogy to the T and R conformation in hemoglobin [8], this result suggests that the binding of the first, third and fourth ligand leads to a break of hydrogen bonds whereas the binding of the second ligand promotes the formation of hydrogen bonds.
One might be inclined to relate our results to entropy changes associated with the transition state upon channel activation under the assumption of reducing the complex activation to a single transition and employing Eyring's rate theory [6]. This has been performed for multiple channels, e.g. voltage-gated K+ channels [38]–[40] and voltage-gated Na+ channels [41]. For related CNGA2 channels we reported previously that the entropy of the open channel plus its environment is smaller than that of the closed channel plus its environment [42], similar to the results in the voltage-gated K+ and Na+ channels. For HCN channels respective data are missing.
The fact that the largest unidirectional probability flux density observed herein was observed for the fully liganded channel seems to suggest that the entropy of the channel is higher in the fully liganded state than in the incompletely liganded state. However, this measure of channel flexibility is fundamentally different from a thermodynamic entropy because the environment is not considered. Therefore, the elevated channel flexibility at full liganding indicates a selective property of closed-open isomerization, irrespective of its environment. For further analysis of the gating process in HCN2 channels it would be a good idea to repeat our experiments at different temperatures and study the temperature dependency of the individual rate constants.
We analyzed the gating of homotetrameric HCN2 channels by probability fluxes in a complex Markovian model, the C4L-O4L model. This analysis was not based on simulations but on a fit to functional experimental data. Studying time courses of the net probability flux density, the unidirectional probability flux density, the total net probability flux, and the transition paths within the C4L-O4L model provided us an unusual view on the gating of the channels. Most remarkably, there is considerable ligand-induced channel opening already after the first ligand has bound whereas there is practically no further opening after the second ligand has bound. Moreover, our analysis shows pronounced hysteresis associated with channel opening and closure. Our results should help to better understand the physiological function of HCN channels and, possibly, to develop strategies for a pharmacological modulation of special functional states.
All calculations are based on a kinetic model for the ligand-induced activation of HCN2 channels containing four ligand binding steps in both the closed and open channel and five closed-open isomerizations (C4L-O4L model termed herein; Fig. 1A) [22]. The channels were activated by a voltage pulse from −30 mV to −130 mV and fcAMP was applied only after the voltage-induced activation was maximal. Fixing the voltage to −130 mV provided us the advantage to study the ligand-induced gating independent of any intervening effects of a changed voltage. The rate constants, determined previously by a global fit strategy of multiple time courses of ligand binding and unbinding as well as activation and deactivation gating, are listed together with their s.e.m. in Table S1. The time courses in the present study were computed by using the mean rate constants. The time-dependent occupancies of the states were computed with the Eigenvalue method using Matlab®. Initially, the channel was assumed to be in the equilibrium C0↔O0 because the ligand concentration, L, was zero (Fig. 1). Then L was set to the value of the applied fcAMP concentration (activation) and back to zero (deactivation). For the numerical computation of time integrals, Simpson coefficients were used.
|
10.1371/journal.pcbi.1001088 | Self-Organization of Muscle Cell Structure and Function | The organization of muscle is the product of functional adaptation over several length scales spanning from the sarcomere to the muscle bundle. One possible strategy for solving this multiscale coupling problem is to physically constrain the muscle cells in microenvironments that potentiate the organization of their intracellular space. We hypothesized that boundary conditions in the extracellular space potentiate the organization of cytoskeletal scaffolds for directed sarcomeregenesis. We developed a quantitative model of how the cytoskeleton of neonatal rat ventricular myocytes organizes with respect to geometric cues in the extracellular matrix. Numerical results and in vitro assays to control myocyte shape indicated that distinct cytoskeletal architectures arise from two temporally-ordered, organizational processes: the interaction between actin fibers, premyofibrils and focal adhesions, as well as cooperative alignment and parallel bundling of nascent myofibrils. Our results suggest that a hierarchy of mechanisms regulate the self-organization of the contractile cytoskeleton and that a positive feedback loop is responsible for initiating the break in symmetry, potentiated by extracellular boundary conditions, is required to polarize the contractile cytoskeleton.
| How muscle is organized impacts its function. However, understanding how muscle organizes is challenging, as the process occurs over several length scales. We approach this multiscale coupling problem by constraining the overall shapes of muscle cells to indirectly control the organization of their intracellular space. We hypothesized the cellular boundary conditions direct the organization of cytoskeletal scaffolds. We developed a model of how the cytoskeleton of cardiomyocytes organizes with respect to boundary cues. Our computational and experimental results to control myocyte shape indicated that distinct muscle architectures arise from two main organizational mechanisms: the interaction between actin fibers, premyofibrils and focal adhesions, as well as cooperative alignment and parallel bundling of more mature myofibrils. We show that a hierarchy of processes regulate the self-organization of cardiomyocytes. Our results suggest that a symmetry break, due to the boundary conditions imposed on the cell, is responsible for polarization of the contractile cytoskeletal organization.
| During biological development, evolving forms are marked by distinct functionalities. An interesting example is the organization of myofibrils in striated muscle cells. As the myocyte matures, the myofibrils are rearranged from an irregularly dispersed pattern into tightly organized bundles spanning the length, rather than the width, of the cell [1]. Although assembly of the myofibril from its molecular constituents has been extensively investigated [2], [3], [4], how myofibrils build this specialized architecture and its functional consequences remains unanswered. This is important because changes in muscle structure accompany not only morphogenesis, but also pathogenesis [5], [6].
Myofibrils mature in a force-dependent manner [7], [8], [9], suggesting that the contractility of a cell may play an important role in polarizing the myofibrillar network. This has been shown in nonmuscle cells where the cytoskeletal architecture within a geometrically-defined microcompartment becomes polarized with increasing tractional forces [10], [11]. Thus, we hypothesized that geometric cues in the extracellular matrix (ECM) can organize the intracellular architecture and potentiate directed myofibrillogenesis. Because of the difficulty in identifying de novo sarcomeres in primary harvest muscle cells in culture, one strategy for studying myofibrillogenesis is to coax the disassembly and reassembly of myofibrils by forcing myocytes to assume shapes that are not commonly observed in vivo using engineered substrates in vitro [10], [11]. To guide these experiments, we developed a computational model of myofibrillar patterning to show the sensitivity of the intracellular architecture to the extracellular space. With these tools, we sought to understand the critical events in the global assembly and organization of the contractile apparatus in cardiac myocytes. By comparing experimental results with our computational model, we were able to elucidate the role of maturing myofibrils, their parallel coupling, and their functional attachment to the focal adhesion assembly and how these processes are guided spatially by the boundary conditions imposed on the cell. After determining the roles of these parameters in myofibrillogenesis, we then expanded our model to test the functional implications of these architectures. We developed a novel method for micropatterning on soft substrates and were able to engineer myocyte shape on substrates that would allow us to measure the contractility of these artificial shapes and compare them with the model results. Together, these results suggest that the self-assembly and -organization of the contractile apparatus is facilitated by a symmetry-breaking event that is potentiated by either a geometric cue in the extracellular space or a random event in the intracellular space.
Our theoretical approach focuses on the interaction between the myofibril and the ECM, as well as adjacent myofibrils (Fig. 1). Inherent to our model are two key assumptions: 1) the force that the myofibrillar bundle exerts on the substrate is fiber length-dependent [12] and 2) adjacent myofibrils affect each other to facilitate lateral coupling, which is akin to them exerting torque on each other. We have modeled only the maturation of cytoskeletal structural elements responsible for contraction and integrin binding to the ECM. We define these components using coarse-grained variables that are experimentally observable. This eliminates the computational complexity required to model detailed molecular interactions and the effect of different protein isoforms. The nomenclature for the immature and mature versions of the myofibril vary with different qualitative models (reviewed by Sanger and colleagues [2]). Here we refer to the immature state as the premyofibril, and the quasi-mature state as the nascent myofibril [1], [13].
Our mathematical approach differs from others [14], [15] in that we incorporate focal adhesion (FA) kinetics, mutual alignment of adjacent contractile fibers, and the dependence of contractile forces on fiber length [16]. The variables used in our approach are: (1) the density of bound and unbound integrin, and , respectively; with the bound integrins connected to premyofibrils and nascent myofibrils labeled as and , respectively; (2) the net force exerted on the bound integrin, ; (3) the local density, , orientation, , and the orientational order parameter, , of the premyofibril network and the nascent myofibril network; and (4) the resultant 2D stress field exerted by the cell on the substrate, T. Previously, we reported [17] that when cardiac myocytes are constrained on 2D islands, their vertical dimension, orthogonal to the plane of the culture surface, is uncontrolled. In that study, we reported that myofibrils are predominantly located under the nucleus, in a plane parallel to the culture surface. However, as that study also showed, several layers of myofibrils may be present, and the nucleus and microtubule organizing center may represent an obstacle to a symmetrical array of myofibrils in the thicker regions of the cell. Our model and analysis is restricted to the 2D intracellular plane closest to the culture surface. Instead of solving the steady state for all of the variables, we numerically simulated their spatiotemporal profiles. This allows us to trace the effect of local symmetry-breaking events such as the mutual alignment of fibers on myofibrillar patterning, which cannot be easily predicted by conventional steady-state analysis. The local symmetry-breaking event may result from a static cue or a transient perturbation.
In our simulation, we began with randomly distributed densities of the unbound integrin, unless fitting parameters, in which case we examined several sets of initial conditions. The unbound integrin can initially become bound through a random process, with the rate proportional to its local concentration. The fraction of bound integrins connected to the fibrils is modeled as an adsorption process, and is calculated using the Langmuir isotherm. The force exerted between FAs is assumed to be proportional to the product of fiber connections at each site [16]. The net force at a local FA is computed by integrating the tension contributed by all connected contractile elements (Fig. 1A). The net force governs the growth rate of local FAs, which in turn modulates the premyofibril network [18], [19]. The assembly of FAs and the bundling of its associated fibers is coupled by a positive feedback loop via forces exerted on the FA [16], [18]. As a consequence of the positive feedback, when the net force on a FA is not zero, both the FA and its associated fibers are structurally reinforced (Fig. 1B–D) [20]. If the net force is zero, the bound integrins will disassemble at each time step and disassociate the attached fibers (Fig. 1E–G) [18], [21]. As time lapses, the premyofibrils are converted to the nascent myofibrils. The local orientation of the nascent myofibril is primarily determined by the antecedent premyofibril network, but also can be modulated by adjacent myofibrils due to their lateral coupling [1], [22]. In some cell shapes, polarization of the myofibrillar array can only be achieved by the lateral alignment of adjacent myofibrils, which occurs at a much slower time scale than that of fiber assembly [1], [22]. The effect of the lateral coupling is modeled as a biasing potential field that distributes the free integrins, such that the nascent myofibrils are moved towards each other through the course of normal integrin recycling. To visualize the amount of parallel, or lateral, coupling of the fibers, we define a variable, ψ, which varies from zero for no local coupling, to unity for the maximal local coupling. The model's calculations are ordered as depicted in Fig. 1H.
To fit the parameters of the computational model, we chose an uncommon cell shape, a stair-shaped myocyte, that we could model computationally in silico and repeatably in vitro with cell engineering techniques (Fig. S1). The parameters were fit on a variety of initial conditions (Fig. S2) such that the steady state results were the same for each. In Fig. 2A we show the temporal results for an initial condition with a random distribution of free integrins. Initially, there are no fibers in the cell, as no integrins are bound (Fig. 2A ). The geometrical symmetry of the stair-shape cell potentiates the initial appearance of fibers predominantly along the diagonal. As the fibers form, the fiber density is mostly uniform throughout the cell, as evident from the line segment thickness (Fig. 2A ). When the nascent myofibrils form and begin to laterally couple, they are distributed diffusely within the cell (Fig. 2A ). As time progresses, the positive feedback increases, i.e. greater number of fibers produces a greater force which drives the clustering of bound integrins and fibers. As a result, the myofibrils achieve a distribution very similar to the steady state (Fig. 2A ). For the rest of the simulation the nascent myofibrils mutually align and exhibit greater degrees of parallel coupling (Fig. 2A ). Myocytes were cultured on stair-step shaped islands for three days and then stained against actin filaments (Fig. 2B). At equilibrium, most nascent myofibrils are coupled and aligned with the major diagonal, as shown experimentally in Fig. 2B and in simulation (Fig. 2A ). The parallel coupling of the nascent myofibrils emerges later in the simulation, as suggested by previous reports [1], [21], [22], [23]. In summary, the simulated dynamics visualized for nascent myofibril bundling and realignment show that well-aligned myofibrils first occurred in the center of the cell, followed the longest diagonal, and recruited additional adjacent fibers to form a bundled, parallel arrangement.
To test our hypothesis, we examined the sensitivity of myocytes and our model to various cellular boundary conditions. We reasoned that when myocytes are constrained by a heterogeneous boundary curvature, triangles (Fig. 3A) and squares (Fig. 3F), the distinct geometrical cues at the cell boundaries would potentiate unique cytoskeletal architectures, but when cells are constrained by a homogeneous boundary curvature, a circle (Fig. 3K), there is no external cue to break the symmetry of the isotropic network. Thus, we examined two cases of the cell with heterogeneous curvature at the periphery: the square shaped cell, where the longest axes are on the diagonal, and the equilateral triangle shaped cell, where the long axes are along the cell periphery. We also tested cells with homogeneous boundary curvature: the circular shaped cell, in which no major axis is defined. To ensure that the observations resulted from geometric considerations alone, we used the same parameter values from the previous simulations.
Fluorescent staining of actin filaments in myocytes cultured on square and triangular ECM islands for 72 hrs revealed that polymerized actin fibers were densely arranged along the longest axes (Fig. 3). The fibers are regularly punctuated along their length, indicating the presence of sarcomeres (Fig. 3B, G). At steady state, modeled triangular and square cells displayed the same cytoskeletal arrangement as the in vitro results, with enhanced parallel bundling occurring along the longest axis of these cells (Fig. 3C, H). Fluorescent staining of vinculin revealed elongated FAs in the corners of the square and triangular cells that were oriented in parallel with their attached myofibrils (Fig. 3D, I). Numerical results revealed the same accumulation pattern of FAs, as indicated by the density of bound integrin located in the corners (Fig. 3E, J). The dynamics of the simulation results are depicted in Fig. 3C, E, H, and J and Video S1, S2, S3, and S4. As previously observed in the simulation shown in Fig. 2, the predominant orientation of the premyofibrils occurs quickly and the parallel bundling increased with time to further stabilize the myofibrillar architecture with respect to the geometric cues in the ECM. These data suggest that FAs localize and mature at the corners because the premyofibrils that align along the longest axes of the cell are the strongest by virtue of their greater propensity for parallel bundling and binding myosin motors [24], [25].
In contrast, myocytes cultured on circular ECM islands (Fig. 3K) for the same period of time have random myofibrillar architectures (Fig. 3L) [26], which is recapitulated in the model (Fig. 3M). Without an external cue to break the geometric symmetry, computer simulations suggest that myofibrillar polarity will emerge after a longer period of time, (almost five times as long as other shapes). Transient multi-pole patterns develop within cellular microcompartments (Video S5) and at equilibrium there is local bundling and nascent myofibril formation, but no overall cell organization (Fig. 3M, Video S5). In vitro, vinculin stains irregularly around the myocyte perimeter (Fig. 3N). In silico, after a similarly prolonged simulation, FAs appear as opposing bands along the cell periphery (Fig. 3O, Video S6). It is important to note that this patterning is due to a random, intercellular, symmetry-breaking event and that while the model will always converge, circular cells both in silico and in vitro, after 2–3 days in culture, often display irrepeatable cytoskeletal structures. Together, the simulation and experimental results summarized in Fig. 3 suggest that the orientation of the premyofibrillar network is regulated by ECM cues. These cues promote stabilization of the network and FAs, facilitating parallel bundling of the nascent myofibrils. Furthermore, our model predicted that the polarized myofibrillar network has a preference to align along the longest axis of cells.
Proper functioning of myocytes requires the correct myofibrillar configuration for coordinated contraction [5]. To correlate myofibrillar structure with contractile function, we investigated the spatial patterning of sarcomeric proteins and conducted traction force microscopy on the cultured myocytes. Fluorescent micrographs of myocytes immunostained against sarcomeric α-actinin revealed distinct myofibrillar patterning on ECM islands of heterogeneous boundary curvature (Fig. 4A, F). The sarcomeric Z-lines register in the internal angles of the corners of both the square and triangle and are perpendicular to the orientation of the actin fibers. To measure myocyte contractile stresses, we engineered ECM islands on soft substrates. When freshly harvested myocytes are cultured on these substrates, they remodel to assume the shape of the island in the same manner as they do on rigid substrates (Fig. 4B, G). Unlike myocytes cultured on the rigid substrates, myocytes on soft substrates do not contract isometrically and can be observed to shorten as in traditional assays of single myocyte contractility (Fig. 4C, H, Video S7 and S8). To visualize substrate deformation due to myocyte contraction, fluorescent beads were embedded in the substrate and bead movement was detected using high speed fluorescence microscopy. The nominal stress field exerted on the substrate due to systolic contraction, with the resting myocyte position defined as the reference state, was calculated from substrate deformation with the known substrate mechanical properties and assuming that the substrate is linearly elastic. In the videos (Videos S9 and S10), the substrate displacement vectors, as depicted by the white arrows, are directed inward during systole, indicating that the substrate is pulled towards the center of the myocyte by the shortening FA-anchored myofibrils. During diastole, they reversed direction as the elastic recoil of the myocyte pushed the substrate back to the rest position. The myocytes generate a unique contractile footprint that mimics the position of the FAs depicted in Fig. 3, with the highest systolic stresses exerted on the substrate at the corners of the myocyte (Fig. 4D, I). Note that even though the model does not differentiate between systolic and diastolic stresses, the experimental substrate stress field pattern matches the simulated results (Fig. 4E, J).
In myocytes of homogeneous boundary curvature, the myofibrillar patterns are not reproducible. However, structural coordination of the myofibrils on a preferential axis was observed, as evidenced by the well-demarcated Z-lines that continuously traversed the 1 to 7 o'clock axis in the circular myocyte shown in Fig. 4K. Similarly, the circular shaped myocytes cultured on soft substrates appear to shorten concentrically during contraction (Fig. 4L, M, Video S11), where a principal axis of shortening is apparent at peak systole but does not occur with the same spatial regularity of the square and triangular cells (Fig. 4N, Video S12), consistent with previous findings with nonmuscle cells [10]. Our model predicted a similar contractile signature (Fig. 4O), with the peak stresses coincident with the location of the widest FA bands observed in Fig. 3O. Thus, these data suggest that muscle cells depend on extracellular spatial cues to efficiently and functionally organize the myofibrils and contracion.
We hypothized that a hierarchy of mechanisms may be responsible for myofibrillar organization. We reasoned that our model would allow us to determine which of the two model features, the fiber length-force dependence and parallel coupling of fibers, was dominant in organizing the myofibrillar architecture. We also reasoned that the nature of the cell boundaries may determine the sensitivity of the cell to these two mechanisms. To test this hypothesis, we ran simulations where these two features were either on, or turned off, within the cell. In the staire-shaped cell, we ran simulations where: 1) there is no mutual alignment mechanism but fiber contractility is fiber length-dependent (L = ON, τ = OFF, refer to Eq. (1) & (4)); 2) the fiber contractility is not myofibril length-dependent but there is mutual alignment of fibers (L = OFF, τ = ON); and 3) there is neither fiber force-length dependence nor any mutual alignment of fibers (L = OFF, τ = OFF). In simulations where the nascent myofibrils have fiber force-length dependence, fibers will predominantly organize along the major diagonal (Fig. 5A) as shown experimentally (Fig. 2B), however, when there is no fiber force-length dependence, fiber bundles follow both the long and short diagonals (Fig. 5B).
We compared the mean degree of parallel coupling as a function of time for all conditions (Fig. 5C). This analysis reveals that the force-fiber length dependence is an essential contributor to the emergence of an organized equilibrium in the myofibrillar network. In these simulations, the absence of the force-length dependence potentiated a less organized nascent myofibril network, whereas mutual alignment of nascent myofibrils enhanced parallel coupling. Eliminating the mutual alignment alone (grey dot-dashed line), produces a minor effect in the stair cell as shown in the inset of Fig. 5C, however, previous reports suggest that the effect of mutual fiber alignment is seen at longer time scales [1], [21], [22], [23].
We asked how mutual fiber alignment would effect myofibrillar organization in the circular cell, whose homogeneous boundary curvature requires an internal, random symmetry break to achieve equilibrium. By eliminating the ability of fibers to cooperatively align in circular cells (grey-empty circle line Fig. 5C), we show that the increase in parallel fiber coupling is solely depended on the ability of the nascent myofibrils to mutually align. The importance of mutual alignment is illustrated by contrasting the steady state fiber organization in circle cell with mutual alignment (Fig. 5D) and no mutual alignment (Fig. 5E). In the case of no mutual fiber alignment the fibers in the circular cell remain randomly organized, which is contradicted by experimental results (Fig. 3L and Fig. 4N). In summary, our data suggests that the fiber length-force dependence is necessary to reproduce myofibrillogenesis in all cell shapes, while the importance of mutual fiber alignment effect increases in cells with homogenous boundary conditions.
Muscle morphogenesis is a hierarchal, self-organizing process spanning from nanometer scale conformational changes in proteins to bundled fibers sometimes a meter in length. We reasoned that boundary constraints are a physical signal that is conserved over all of these length scales and spatially organizes this broad range of coupled structures. Based on previous experimental evidence [17], [26], [27], [28], we hypothesized that geometric cues in the extracellular space help organize the assembly of the contractile apparatus in the cytoplasm and developed computational and experimental models to recapitulate these events.
We report that distinct cytoskeletal architectures arise from two temporally-ordered, organizational processes: the cooperative interaction between premyofibrils and focal adhesions, as well as the mutual alignment and parallel bundling of nascent myofibrils. Our model assumes that the assembly of FAs and the parallel bundling of actin based fibers is coupled by a positive feedback loop and that the growing force on the FA potentiates its structural reinforcement, as suggested by previous experimental work [7], [8], [9]. By modeling the amount of bound and unbound integrin and by marking the maturation of the premyofibril to a nascent myofibril simply by increased contractility, we are able to predict the organization of the contractile apparatus in cardiac myocytes cultured on engineered substrates in a computationally efficient manner. To achieve this efficiency, we ignore the details of the molecular constituents of the assembly of myofibrils [2], [3], [4]. However, we were able to account for all the dominant factors in a course grained manner as indicated by the match between all our models and experiments. By experimenting with our assumptions in silico and comparing them to data from in vitro experiments, our results suggest that the force that the myofibrillar bundle exerts on the substrate is fiber length-dependant [10], [11], [12] and that the adjacent myofibrils exert “torque” on one another to facilitate coupling [24], are necessary to describe how these myocytes build and organize their internal cytoskeleton relative to extracellular cues. Our computationally efficient model recapitulates the elegant protein choreography of the sarcomere assembly, where an ensemble of proteins assembles repetitively along the length of the actin fiber template.
Several models of cell cytoskeleton assembly and mechanics have been reported and it is worthwhile to compare and contrast the efforts [14], [15], [29]. Our model is similar to the model by Novak and colleagues [16] in that we have used reaction kinetics to simulate the dynamic self-assembly and – organization of the cytoskeleton. These approaches differ from that of Deshpande, et al [14], [15], [16] who report a solid mechanics model and Pazek and colleagues [14], [15], [29] who use a mechanochemical model. All four of these models simulate the bound and free states of integrins in some form and also model the increasing stabilization, or maturation, of focal adhesions with increases in exerted force. The Despande and Pazek models offer detailed mechanical analysis of the cell-substrate interface, whereas our model, like the Novak model, does not. While the Pazek, et al., model does not recapitulate stress fibers, our model, like the Novak and Deshpande models, does. Our model accounts for the specialized case of the maturing striated muscle cell by mimicking the transition of a premyofibril to the nascent myofibril, modeled by an increased ability to generate tension. The Hammer and Novak models omit the fiber length-force assumption that is critical to our model's ability to recapitulate our experimental data. Similarly, the Desphande and Novak models explicitly do not account for mutual alignment of fibers, whereas ours does. Our model, like the Desphande et al. and Pazek models, calculates the load exerted on the substrate by the contracting cell, where the Desphande and Pazek models offer detailed descriptions of the solid mechanics at this interface. Both our model and that by Novak et al., are similar to larger scale models of myofibril adaptation in the left ventricle [30], in the assumption that there is a network of fibers where all integrins are connected to all other integrins. Each model, including the one reported herein, varies in approach and further work is required to test all of these models against experimental data as we have attempted.
We were able to reproduce the results shown by Novak et al., [16], who predicted that with no fiber tension-length dependence, and homogeneous boundary conditions the FAs would aggregate to the perimeter. However, as our in vitro work shows even with a homogeneous boundary condition, i.e. the circular cell, there occurs a symmetry break, therefore it is necessary to introduce fiber tension-length dependance and mutual alignment of fibers for in silico experiments. We can also utilize the model to explore the effect of cell boundary curvature, cell aspect ratios and combinations of multiple cells on the myofibril distribution, as well as the relative importance of mutual fiber alignment in three dimensions. Additionally, it will be possible to integrate our model with adhesion dynamics models using the same methods as Paszek et al., to explore integrin clustering with contractile cells on substrates with different material properties [29]. This combination of a mechanical model with our myofibrillogenesis model could also allow for simulations of the rearrangement of the extracellular matrix by contractile cells.
In summary, our study suggests that hierarchal organization of muscle requires localized cues that guide myofibrillogenesis. Specifically, a local symmetry break is required to potentiate the assembly and organization of FA and actin complexes that are the template for myofibrillar organization. Such cytoskeletal symmetry-breaking has also been widely observed in other important biological behaviors such as cellular migration [11], cellular division [31], and formation of tissue sheets [32]. The symmetry-breaking can arise from a static, external cue, such as a geometric feature in the boundary conditions imposed on the cell, or from a dynamic internal cue, such as a local overlapping of long fibers. The multiple time scales of these interacting events suggest a hierarchy of post-translational, self-organizational processes that are required for coupling cellular form and function.
All experiments were conducted in accordance with the guidelines of the Institutional Animal Care and Use Committee of Harvard University. Trypsinized ventricular tissue isolated from 2-day old neonatal Sprague Dawley rats (Charles River Laboratories, Wilmington, MA) was serially dissociated into single cells by treating the ventricular tissue 4 times with a 0.1% solution of collagenase type II (Worthington Biochemical, Lakewood, NJ) for 2 minutes at 37°C. The myocyte fraction was purified and pre-plating the cells twice for 45 minutes each time. Purified myocytes were plated onto micropatterned substrates prepared as described below at a density of 100,000 cells per coverslip and kept in culture at 37°C with a 5% CO2 atmosphere. The culture medium was M199 (Invitrogen, Carlsbad, CA) base supplemented with 10% heat-inactivated Fetal Bovine Serum, 10 mM HEPES, 20 mM glucose, 2mM L-glutamine, 1.5µM vitamin B-12, and 50 U/ml penicillin. The medium was changed 24 hours after plating to remove unattached and dead cells and every 48 hours afterwards. After 72 hours in culture, most cardiac myocytes beat spontaneously and were used either for immunostaining or traction force measurements.
Micropatterned substrates containing square, triangular, or circular adhesive islands were prepared for immunostaining and traction force microscopy, as follows. For immunostaining, the substrates were micropatterned using a microcontact printing procedure similar to that described by Tan et al. [35]. Micropatterned substrates for traction force experiments were created by adapting the published techniques [10], [36]. Briefly, a thin layer of 10% by weight poly-N-iso- propylacrylamide (PIPAAM) prepared in 1-butanol was spin coated on a silicon wafer (Fig. S1a). A 50∶75 µm layer of photoresist (SU-8, MichroChem Corp, Newton, MA) was spin-coated on top of the PIPAAM (Fig. S1b), UV light treated through a photolithographic mask (Fig. S1c), and developed to obtain a complementary master that contained holes with the same size and shape as the desired adhesive islands (Fig. S1d). The master was immersed in ice water to dissolve the PIPAAM and the photoresist membrane was released from the wafer (Fig. S1e). Polyacrylamide gels (0.1% bis and 5% acrylamide; 90 µm thick) containing 1∶500 volume of carboxylate-modified fluorescence latex beads (0.2 µm Fluospheres, Molecular Probes, Eugene, OR) were fabricated on 25 mm coverslips. The Young's modulus of the gel was estimated to be ∼3 KPa using atomic force microscopy as described previously [37]. The photoresist membrane was placed on the surface of the gel and 1 mM sulfo-SANPAH (sulfosuccinimidyl- 6-4-azido-2-nitrophenylamino-hexanoate; Pierce, Rockford, IL) in 50 mM HEPES was added through the holes in the photoresist membrane. The whole system was then placed under vacuum for 3 minutes to ensure that the sulfo-SANPAH reached the gel surface. The surface of the gel that contacted with the sulfo-SANPAH was photoactivated by UV light exposure (Fig. S1f). After excess sulfo-SANPAH was removed, fibronectin (FN) 100 µg/mL was added to the membrane and the gel was placed under vacuum for another 3 minutes to remove bubbles from the holes (Fig. S1g). The FN was allowed to react with the photoactivated gel for at least 4 hours at 37°C to create FN-coated adhesive islands. Excess FN was washed away with PBS. After removal of the photoresist membrane, the gel was immediately used for cell plating (Fig. S1h).
In vitro studies show that the maturation of sarcomere can be determined by measuring the distance between two adjacent α-actinin rich spots that are supposed to be the precursors of the Z-band [13]. A 2–2.5 µm spacing between the sarcomeric α-actinin rich spots indicates a matured sarcomere [2]. We used a fast Fourier transform (FFT) to calculate the spacing of sarcomeric -actinin rich spots in Fig. 4A, F, K. An intensity profile of the sarcomeric α-actinin stains was chosen along myofibrils spanning the long axis of the cells. The profile was then detrended, weighted with a Hamming window and transformed into the spatial frequency domain by FFT. The spatial frequency at peak power of the first-order harmonic in the spatial frequency domain was identified and converted into the spatial domain to yield the sarcomere length. The results reveal that the sarcomere lengths are 2.4±0.1 µm, 2.2±0.1 µm, and 2.4±0.2 µm for the cell in Fig. 4A, F, K, respectively, indicating that they are mature sarcomeres.
Coverslips containing the beating myocytes were removed from the incubator, mounted onto a custom-made microscope stage containing a bath chamber, and continuously perfused with 37°C normal Tyrode's solution (1.192 g of HEPES, 0.901 g of glucose, 0.265 g of CaCl2, 0.203 g of MgCl2, 0.403 g of KCl, 7.889 g of NaCl and 0.040 g of NaH2PO4 per liter of deionized water, reagents from Sigma, St. Louis, MO). Fluorescence images of gels containing fluorescent beads immediately beneath the contracting myocytes were taken at 28.1 Hz. The duration of image acquisition was long enough to include at least two complete cycles of contraction-relaxation of individual myocytes. Consecutive images were paired and the prior image was used as a reference to measure the change of the position of the fluorescence beads using the algorithm described previously [38]. This yielded the discretized displacement field between two consecutive frames. The calculated displacements were summed up for a whole systolic cycle to determine the overall 2D displacement field. The systolic traction field was calculated from the displacement field by adapting the algorithm previously developed [39], [40]. This algorithm solved the inverse of the Boussinesq solution from the displacement field on the surface of an elastic halfspace to obtain the traction field when the mechanical properties of the gel are known. The Poisson ratio of the gel was assumed to be close to 0.5 [10]. The interior of the cell was subdivided into 4×4 µm2 squares to approximate the discretized localization of contractile forces. The ability of a particular solved traction field to explain the observed displacements was estimated with statistics. In addition to a zero-order Tikhonov regularization, a constraint that the forces should not become exceedingly large was used to minimize and stabilize the solution [40]. The L-curve criterion, as previously described [40], was used to determine the optimal balance between the data agreement and the regularization.
Cardiac myocytes stained for actin (Alexa 488 Phalloidin, Molecular Probes), vinculin (clone hVIN-1, Sigma), and sarcomeric α-actinin (clone EA-53, Sigma) were fixed in 4% PFA with 0.01% Triton X-100 in PBS buffer at 37°C for 15 minutes and equilibrated to room temperature during incubation. Secondary staining was performed using tetramethylrhodamine- conjugated goat anti-mouse IgG (Alexa Fluor 594, Molecular Probes), and nuclei were visualized by staining with 4′,6′-diamidino-2- phenylindole hydrochloride (DAPI, Molecular Probes). All fluorescence and traction force microscopy was conducted with a Leica DMI 6000B microscope, using a 63× plan-apochromat objective. For traction force experiments, images were collected with a Cascade 512b enhanced CCD camera, while immunofluorescence images were collected with a CoolSnap HQ CCD camera (both from Roper Scientific, Tucson, AZ) controlled by IPLab Spectrum (BD Biosciences/Scanalytics, Rockville, MD).
|
10.1371/journal.pgen.1002188 | Ongoing Phenotypic and Genomic Changes in Experimental Coevolution of RNA Bacteriophage Qβ and Escherichia coli | According to the Red Queen hypothesis or arms race dynamics, coevolution drives continuous adaptation and counter-adaptation. Experimental models under simplified environments consisting of bacteria and bacteriophages have been used to analyze the ongoing process of coevolution, but the analysis of both parasites and their hosts in ongoing adaptation and counter-adaptation remained to be performed at the levels of population dynamics and molecular evolution to understand how the phenotypes and genotypes of coevolving parasite–host pairs change through the arms race. Copropagation experiments with Escherichia coli and the lytic RNA bacteriophage Qβ in a spatially unstructured environment revealed coexistence for 54 days (equivalent to 163–165 replication generations of Qβ) and fitness analysis indicated that they were in an arms race. E. coli first adapted by developing partial resistance to infection and later increasing specific growth rate. The phage counter-adapted by improving release efficiency with a change in host specificity and decrease in virulence. Whole-genome analysis indicated that the phage accumulated 7.5 mutations, mainly in the A2 gene, 3.4-fold faster than in Qβ propagated alone. E. coli showed fixation of two mutations (in traQ and csdA) faster than in sole E. coli experimental evolution. These observations suggest that the virus and its host can coexist in an evolutionary arms race, despite a difference in genome mutability (i.e., mutations per genome per replication) of approximately one to three orders of magnitude.
| To examine the ongoing changes driven by host–parasite interactions, we have constructed a coevolution model consisting of Escherichia coli and the lytic RNA bacteriophage Qβ (Qβ) in a spatially unstructured environment. In coevolution through 54 daily copropagations of the parasite and its host, E. coli first evolved partial resistance to infection and later accelerated its specific growth rate, while the phage counter-adapted by improving release efficiency with a change in host specificity and a decrease in virulence. Whole-genome analysis of E. coli and Qβ revealed accelerated molecular evolution in comparison with Qβ propagation in this study and E. coli sole passage reported previously. The results of the present study indicated that, despite the large difference in mutability of their genomes (approximately one to three orders of magnitude difference), a host with larger genome size (4.6 Mbp) and a lower spontaneous mutation rate (5.4×10−10 per bp per replication) and a parasite with a smaller genome size (4,217 bases) and a higher mutation rate (1.5×10−3 to 1.5×10−5 per base per replication) were capable of changing their phenotypes to coexist in an arms race.
| Host–parasite coevolution has been a topic of intense research interest in various fields from basic science of molecular evolution to agricultural and medical applications [1]–[5]. According to the Red Queen hypothesis or arms race dynamics, coevolution leads to complex but continuous change, adaptation, and counter-adaptation of the phenotypes of interacting organisms [2], [6], [7]. Futuyma and Slatkin suggested that investigation of coevolution could raise and help provide answers to many questions regarding the history of evolution, e.g., whether parasites tend toward specialization or toward benign or even mutualistic relationships with their hosts [8].
There have been many previous observational and theoretical studies on natural host–parasite dynamics. With regard to the relationships between bacteria and phages, Rodríguez-Valera et al. proposed the constant-diversity dynamics model in which the diversity of prokaryotic populations is maintained by phage predation [9]. Moreover, an observational study supported the model by analyzing the dynamics of bacteria and phages in four aquatic environments using a metagenomics method and showed that microbial strains and viral genotypes changed rapidly [10]. In addition, experimental models in simplified environments have been employed to analyze the ongoing process of coevolution. Various pairwise combinations of bacteria and phages and one with Caenorhabditis elegans and bacteria have been subjected to long-term laboratory cultivation [11]–[15]. These studies indicated that coevolution proceeded on a laboratory time scale [11]–[14], accelerated molecular evolution of parasites [16], [17], and broadened the host range of parasites [14]. However, the changes in genetic information and phenotype of parasites and their hosts through coevolution remain to be elucidated, and the changes in host specificity and virulence of the parasites through the arms race have not been determined in sufficient detail because ongoing adaptation and counter-adaptation in simplified experimental model systems have not been analyzed at the levels of population dynamics and molecular evolution.
To examine the ongoing changes driven by host–parasite interactions, we have constructed a coevolution model consisting of Escherichia coli and the lytic RNA bacteriophage Qβ (Qβ) in a spatially unstructured environment. Qβ is a simple RNA bacteriophage that infects and lyses E. coli cells, taking about 1 h for its burst without escaping into a lysogenic state. It has a single-stranded RNA genome of 4,217 bases encoding four genes for A2, A1 (read-through), coat protein, and RNA replicase β subunit [18]. Due to a high misinsertion rate and lack of a proofreading mechanism, ribovirus RNA replicase (including that of Qβ) has a high mutation rate [18]–[22], which allows us to monitor the evolutionary changes on a laboratory time scale.
Here, we report that in coevolution through 54 daily copropagations of the parasite and its host, E. coli first evolved partial resistance to infection and later showed acceleration of its specific growth rate, while the phage counter-adapted by improving release efficiency with a change in host specificity and a decrease in virulence. Fitness analysis indicated that these phenotypic changes occurred within an arms race, i.e., accompanied with a monotonic fitness increase of either the parasite or its host. Whole-genome analysis indicated that the phage accumulated 7.5 mutations mainly in the A2 gene 3.4-fold faster than in Qβ propagation evolution where the phage was transferred daily to freshly prepared E. coli cultures, while E. coli showed fixation of two mutations (in traQ and csdA) faster than in sole E. coli experimental evolution. The results indicated ongoing adaptation and counter-adaptation through a host–parasite arms race.
Evolution experiments were carried out with copropagation of E. coli and Qβ and with propagation of Qβ only (Figure 1A). In the copropagation experiment, the ancestral E. coli strain HL2 (Anc(C)) and Qβ derived from cloned Qβ cDNA [23] (Anc(P)) were mixed, cultivated, and diluted so that the next daily culture was initiated at approximately 1×107 E. coli cells/ml. We calculated the replication generations of Qβ genome as the cumulative generations of each passage, (Nfinal/Ninitial) = 2g, where Nfinal and Ninitial represent final and initial free phage density of each passage in plaque forming units (PFU/ml), respectively, and g represents replication generation. We also calculated E. coli cell generations as the cumulative generations of each passage, (Nf/Ni) = 2n, where Nf and Ni represent the final and initial colony forming units (CFU/ml) of each passage, respectively, and n represents cell generation. In the very early phase of the copropagation experiment, the cell generation was underestimated due to cell lysis by infection. The copropagation experimental population was divided into two on day 18, equivalent to 59 replication generations and 62 cell generations. Culture was continued to a total of 54 days (lines 1 and 2), equivalent to 163 replication generations and 163 cell generations for line 1, and 165 replication generations and 164 cell generations for line 2 (Figure 1B). Two Qβ propagation experiments, lines 3 and 4, were conducted in parallel for 18 days, equivalent to 169 and 168 replication generations where the phage population was separated daily by centrifugation from the host and transferred into fresh logarithmic cultures of the host Anc(C) (Figure 1B).
The population dynamics of the copropagation experiment demonstrated the coexistence of E. coli and Qβ (Figure 2), although Qβ is lytic and has no lysogenic state. The daily Qβ population density fluctuated over the course of the copropagation experiment, while the E. coli population density was stable probably due to the constant initial density of the host at each daily coculture. The degree of phage amplification in the copropagation experiment (2–20-fold per single coculture) was substantially lower than that in the Qβ propagation experiment (approximately 1,000-fold), even though the initial multiplicity of infection (MOI) in each passage was approximately 0.5 (approximately 107 phages/ml over 2×107 E. coli cells/ml) for the Qβ propagation experiment and was not higher than that for the copropagation experiment (Figure 2B and 2C). These observations suggested that the biotic environment for phage amplification, i.e., the cellular state of the host E. coli, changed during the copropagation experiment.
Cross-cocultures were conducted to determine the changes in fitness of E. coli and Qβ in the copropagation experiment. Four hosts (Anc(C), M54(C), M163(C), and M165_2(C)), and the four corresponding phages (Anc(P), M54(P), M163(P), and M165_2(P)) at the 1st, 54th, 163rd, and 165th replication generations in the copropagation experiments of lines 1 and 2 were cocultured to measure fitness in each pairwise combination. Here, the fitness of E. coli is defined as the ratio of the initial to the stationary optical density at 600 nm (OD600), while the fitness of the phage is the ratio of the initial to the stationary free phage density (PFU/ml) (Table 1, Table 2 and Figure S1).
The host E. coli evolved partial resistance along with its increase in fitness (Table 1). The evolved hosts M54(C), M163(C), and M165_2(C) showed phage amplification ratios two to three orders of magnitude lower than Anc(C), regardless of whether the ancestral or evolved phage was used (host: Anc(C), M54(C), M163(C), and M165_2(C), parasite: Anc(P), M54(P), M163(P), and M165_2(P): one-way ANOVA F3,24 = 213, P<0.01; post hoc Tukey–Kramer test, P<0.01; Table 2). The resistance was only partial, allowing phage amplification of only approximately one order of magnitude. On the other hand, the host E. coli infected with Anc(P) gradually showed an increase in amplification ratio along with host evolution (one-way ANOVA F2,3 = 469, P<0.01; post hoc Tukey test detected significant differences between all combinations: Anc(C) vs. M54(C) and Anc(C) vs. M163(C), P<0.01; M54(C) vs. M163(C), P<0.05; Table 1). The growth curve of M163(C) inoculated with the phages became similar to that of the uninfected host (Figure S1A, third from left), suggesting that the host population evolved, increasing its fitness, to become almost oblivious to the phages.
Despite the development of partial resistance by the host, the phage also increased its fitness through changes in host specificity (Table 2). The evolved phage M54(P) and M163(P) showed higher fitness on the evolved host M54(C) with partial resistance than Anc(P) on the same host (one-way ANOVA F2,3 = 19.7, P<0.05; post hoc Tukey test, P<0.05; Table 2). In line 1 and line 2, the most evolved Qβ, M163(P) or M165_2(P) showed the highest fitness on corresponding E. coli, M163(C) or M165_2(C), respectively. Briefly, there was a significant difference in fitness among the host–parasite combinations (host: M163(C) or M165_2(C), parasite: Anc(P), M54(P), M163(P), or M165_2(P), one-way ANOVA F2,3 = 60.8, P<0.01; post hoc Tukey test, P<0.05 for M163(C); F2,3 = 37, P<0.01; post hoc Tukey test, P<0.05 for M165_2(C); Table 2). The phage evolved through natural selection to show greater amplification on the corresponding host, although the amplification ratio itself decreased from approximately 104 to 101.
The phage, while responding adaptively to the evolutionary changes of its host, showed a decrease in amplification ratio on the ancestral host strain, leading to a decrease in virulence. The amplification ratios of the phage on the host Anc(C) gradually decreased over the course of the copropagation experiment (host: Anc(C), parasite: Anc(P), M54(P), M163(P), and M165_2(P), one-way ANOVA F3,4 = 17.4, P<0.01; post hoc Tukey test, P<0.05; Table 2). A decline in phage amplification was also observed as a reduction in plaque size (Figure 3). Consequently, the evolved phage showed less cell killing effect against the ancestral strain Anc(C), resulting in better growth of the ancestral bacterial strain (host: Anc(C), parasite: Anc(P), M54(P), M163(P), and M165_2(P), one-way ANOVA, F3,4 = 340, P<0.01; post hoc Tukey test, P<0.01; Table 1 and Figure S1A, left), i.e., the phage showed a decrease in virulence. In addition, the phage evolved in the Qβ propagation experiment (S94_3(P)) showed the greatest amplification ratio (host: Anc(C), parasite: Anc(P), M54(P), M163(P), M165_2(P), and S94_3(P), one-way ANOVA F4,5 = 37.8, P<0.01; post hoc Tukey test, P<0.05; Table 2 and Figure S1B, left) and similar virulence against Anc(C) with Anc(P) (host: Anc(C), parasite: Anc(P) and S94_3(P), Welch's t test, t = 12.7, P = 0.45; Table 1 and Figure S1A, left). These results suggest that the decrease in virulence was not due to simple degeneration through the long-term passage experiment, but was probably at the expense of increasing the fitness of the phage in the arms race with its host.
To examine how E. coli and Qβ improved their fitness during coevolution, free phages, infected E. coli cells, and total E. coli cells from the copropagation of line 1 were monitored hourly by determining the numbers of PFUs in the supernatant and pellet after centrifugation and CFU, respectively (see Materials and Methods).
E. coli was found to first evolve partial resistance to Qβ, which was followed by a later increase in the specific growth rate. After 3 hours of incubation with Anc(P), almost all of the ancestral host Anc(C) cells were infected, while infection ratios of the evolved hosts M54(C) and M163(C) were only 0.03% and 0.08%, respectively (Figure 4A, 4B, and 4D, left). The observed partial resistance was likely due to a very low adsorption rate of E. coli cells to the phage (Figure 5). As most of the evolved host cells remained uninfected, they were able to proliferate, while the ancestral cells could not. In addition, the uninfected cells of the most evolved hosts, M163(C) and M165_2(C) for lines 1 and 2, respectively, showed higher specific growth rates than those of M54(C) (see legend of Figure 6 for specific growth rates, ANCOVA, F2,24 = 18.0, P<0.001; post hoc Tukey test, P<0.001). Although the OD600/CFU seemed to have changed over the copropagation experiments (Figure 6), the same conclusion was obtained using specific growth rates based on CFU values (data not shown). Briefly, M54(C) eliminated Anc(C) from the population by developing partial resistance to phage infection, and M163(C) and M165_2(C) finally took over the population due to acceleration of specific growth rate.
The phages evolved to show increased release efficiency, i.e., the number of phages released from a single infected cell per unit time. As the phage evolved, the speed of free phage amplification for either M54(C) or M163(C) increased (the amplification rates of free phage density of Anc(P) and M54(P) on M54(C) were 0.11 h−1 (r2 = 0.43) and 0.31 h−1 (r2 = 0.93), respectively, two-tailed t test t = 3.24, P<0.01; Figure 4B and 4C, right, and those of Anc(P), M54(P), and M163(P) on M163(C) were 0.25 h−1 (r2 = 0.79), 0.24 h−1 (r2 = 0.89), and 1.38 h−1 (r2 = 0.93), respectively, ANCOVA; F2,18 = 35.3 P<0.01; post hoc Tukey test, P<0.01; Figure 4D, 4E, and 4F, right), while the infection efficiency, i.e., the rate of increase in infected cells, did not change significantly for the same hosts (the rates of increase in infected cells of M54(C) infected with Anc(P) or M54(P) were 1.21 h−1 (r2 = 0.89) and 1.26 h−1 (r2 = 0.98), respectively, two-tailed t test, t = 0.39, P = 0.70; Figure 4B and 4C, left, and those of M163(C) infected with Anc(P), M54(P), or M163(P) were 2.40 h−1 (r2 = 0.98), 2.80 h−1 (r2 = 0.92), and 2.72 h−1 (r2 = 0.99), respectively, ANCOVA, F2,12 = 0.95, P = 0.41; Figure 4D, 4E, and 4F, left). The acceleration of free phage amplification rate could be attributed to either an increase in burst frequency per unit time or burst size. There was no significant difference in burst size between M163(P) and Anc(P) on M163(C) determined by the method of analysis of burst sizes in single cell [24] (data not shown). Therefore, the phage seems to have evolved to burst more frequently from infected cells per unit time. It is noteworthy that the most evolved phage, M163(P), inoculated onto M163(C) showed a marked increase in number of free phage at 3 h, probably leading to further infection of surrounding uninfected hosts and an increase in proportion of infected cells beyond the inoculated free phage concentration (4.8×105 PFU/ml) (Figure 4F). Other phages stopped increasing the number of infected cells at around the inoculated free phage concentration (Figure 4D, 4E, and 4G).
We performed whole genome sequence analyzes of all of the Qβ populations indicated in Figure 1B to determine how molecular evolution of the phage proceeded in response to the adaptation of the hosts. First, mutations were shown to be accumulated in a biased manner in the A2 gene, which encodes a multifunctional protein related to infection and cell lysis (Figure 1B and Table 3). The A2 gene, accounting for 30% of the whole genome, accumulated 65.5% of all mutations, and this bias was shown to be statistically significant (P<0.05, two-tailed binomial test). A similar substantial accumulation of mutations in genes related to host infection was demonstrated previously in an evolution experiment using the DNA bacteriophage Φ2 [16]. The mutation fixation rate in phage was higher in the copropagation experiment (1.0×10−5±6.0×10−7 per base per generation) than that in the Qβ propagation experiment (3.2×10−6±5.1×10−7 per base per generation) (two-tailed Welch's t test t = 4.3, P<0.01), suggesting that the phage showed accelerated molecular evolution through coevolution with its host (Figure 7).
Whole-genome analysis of E. coli revealed the process of molecular evolution in the host cells. We analyzed the whole genome sequence of M163(C) using an Illumina Genome Analyzer IIx (GAIIx; Illumina, San Diego, CA) and confirmed the mutations in M163(C) together with Anc(C) and M54(C) by the dideoxynucleotide chain termination sequencing method [25]. A single mutation in traQ (S21P) encoded on the F plasmid was detected in M54(C) and an additional mutation was detected in csdA (D340N) in M163(C) (Table S1, Table S2). As discussed below, the protein products from these genes may contribute to resistance to phage infection and the increase in fitness of E. coli.
In the copropagation experiment, E. coli adopted a simple strategy with only two mutations, while the phage accumulated more mutations within its small genome as counter-adaptation against the evolutionary changes in the host. The host first developed resistance to phage infection via a non-synonymous mutation in traQ. This gene encodes TraQ, the conjugal transfer pilin, which is a component of the F pilus, and is a chaperone for inserting propilin into the inner membrane. Propilin was reported to be unstable in traQ− cells [26], and amino acid 21 of TraQ where the mutation was detected in this study interacts with propilin [27]. F pilus assembly from membrane F-pilin requires many Tra proteins [28]. As no mutations were detected in other Tra protein genes in the copropagation experiment, the mutation on TraQ may result in a decrease in the amount of inserted propilin, leading to the partial resistance observed in this study. E. coli then showed further mutation in csdA, which encodes CsdA, an enzyme related to Fe/S biogenesis and a new sulfur transfer pathway that is related to the fitness of these cells, especially in stationary phase [29]. Therefore, this mutation could be beneficial as the host was passaged daily at the stationary phase in the evolutionary experiment.
On the other hand, the phage evolved to increase release efficiency by accumulating mutations mostly in the gene encoding the A2 protein. A2 is a multifunctional protein with roles in host cell lysis, adsorption to the F pilus of E. coli, RNA binding during capsid assembly, protection of the 3′ terminus, penetration into the cytoplasm of the host, and blockage of cell wall biosynthesis by inhibiting the catalytic step from UDP-GlcNAc to UDP-GlcNAc-EP catalyzed by MurA [18], [30]–[32]. Due to the cell lytic activity of A2, it is unsurprising that these mutations might have resulted in an increased burst frequency and release efficiency. In fact, the experimental lag period between infection and detection of the increase in free phage became shorter by approximately 1 hour in the cross-culture experiment (e.g., 3 h for Anc(P) and 2 h for M163(P) on M163(C), Figure 4D and 4F, right, respectively). It should be noted that the uninfected M163(C) reached the stationary phase, which was not susceptible to phage infection, approximately 1 hour earlier than the other hosts (see legend of Figure 6 for the time to reach the stationary phase, one-way ANOVA, F2,3 = 693.8, P<0.001; post hoc Tukey test, P<0.001). Thus, it is possible that the increased burst frequency of M163(P) for the earlier phage release evolved as a counter-adaptation on the evolved host M163(C) due to the shorter period available for infection. Previous experiments using DNA bacteriophages indicated that shorter latent periods were favored in the presence of a high density of highly susceptible host cells [33], [34].
It is of interest that Qβ evolved to show reduced virulence toward the ancestral host. Many studies have indicated that phages with low or moderate virulence were favored in vertical transmission or in structured environments [35]–[37], while Qβ has no lysogenic state and evolved reduced virulence in this experiment. The decrease in virulence observed in this study may have been a side effect of the increase in burst frequency. If fact, the evolved phage M163(P) with increased burst frequency on the evolved host M163(C) showed lower virulence and lower fitness on Anc(C) than Anc(P) on Anc(C) (Figure 4 and Figure S1 left), suggesting that Qβ may have co-evolved to increase the burst frequency in reducing some benefits that can be gained if the host reverts to the Anc(C)-like phenotype.
The single non-synonymous mutation at position 221 of the A2 gene found in M54(P) seems to have resulted in reduced virulence and a change in host specificity. As the non-synonymous mutation was only observed in the copropagation experiment and two others were also observed in Qβ propagation and the deposited sequence (NCBI accession no. AY099114), the mutation at 221 and/or the combinations with the mutation and two other mutations may have resulted in the decrease in virulence and the change in host specificity observed in M54(P).
In coevolution between Qβ and its host, E. coli, the phage showed accelerated molecular evolution (Figure 7). In the Qβ propagation experiment, the molecular evolution of the phage proceeded but seemed to slow down after the 94th generation. On the other hand, the phage coevolving with E. coli retained a 3.4-fold faster molecular evolution rate throughout the copropagation experiment. The higher evolution rate may be attributable to the changes occurring in the host E. coli. If the host had stopped evolving, e.g., at the 54th generation, the M163(P) or M165_2(P) phage would not have been fixed into the population as it had fitness similar to or less than that of M54(P), leading to deceleration of evolutionary rate. It should be noted that neutral mutations cannot be fixed in the copropagation experiment because 163 replication generations is too short for them to become fixed. The fixation of neutral mutations is known to require generations approximately as long as the effective population size (Ne) [38]. The effective population size in the copropagation experiment was roughly estimated as the bottleneck size of the population (approximately 103 phages) assuming that 1% of the minimum initial 106 phages infect and burst to release approximately 107 phages. Thus, even synonymous mutations observed here were positively selected [38], consistent with the influence of the RNA secondary structure on Qβ genome replication reported previously [39]–[42]. Some synonymous mutations may have physiological impacts on phage growth because of genomic secondary structure; it has been reported that some synonymous mutations or mutations in intergenic regions show lethal effects in Qβ [42].
It is noteworthy that the fixation rate of the E. coli genome in the copropagation regime (2.6×10−9 per bp per generation) calculated as 2 mutations in 4.73 Mbp per 163 generations was one order of magnitude higher than that under conditions of E. coli sole passage, maintaining log phase at 37°C (1.7×10−10 per bp per generation) [43] or 20,000 generations (1.6×10−10 per bp per generation) (Poisson distribution, P<0.01) [44]. In summary, these observations indicated that molecular evolution rates of both the parasite and its host were accelerated through adaptation and counter-adaptation.
Based on the observed fitness changes in the host E. coli and in the Qβ phage, we propose a plausible coevolution path to depict the arms race between Qβ and the host E. coli. As the order of phage fitness on the ancestral E. coli Anc(C) was Anc(P)>M54(P)>M163(P), the population in the coculture seemed to first take a route not in the direction of phage evolution (upward) but in the direction of host evolution (right), increasing host fitness by increasing its resistance to Qβ (Figure 8). The arrows in Figure 8 reflect the experimentally determined finesses changes (Table 1 and Table 2). Arriving around the pair position of M54(C) vs. Anc(P), the population could take either the upward or rightward direction, but happened to take the direction of phage evolution due to the occasional appearance of a single non-synonymous mutation at position 221 in the phage genome that was detected only under copropagation conditions (Table 3). The population of M54(P) and M54(C) could not fix a phage mutant like M163(P) with the same fitness as M54(P), but fixed the E. coli mutant M163(C) with fitness higher than that of M54(C). Due to the host change from M54(C) to M163(C) accompanied with an additional single non-synonymous mutation in csdA, the phage mutant M163(P) was fitter than M54(P) and was therefore fixed in the final population. Taken together, these findings indicated that the evolutionary path seemed to be an arms race involving adaptation of E. coli and counter-adaptation of the phage.
We showed that parasites, such as RNA viruses, and hosts, such as E. coli, have the potential to coexist even in an arms race. When a parasite encounters its host, the host may become extinct through the evolution of high parasite virulence, or the parasite may become extinct through the evolution of host resistance. However, both may also change their phenotypes by genomic mutation in a synchronized manner and thus coexist. The results of the present study indicated that a host with a larger genome size (4.6 Mbp) with a low spontaneous mutation rate (5.4×10−10 per bp per replication) [45] and a parasite with a smaller genome size (4,217 bases) and a higher spontaneous mutation rate (1.5×10−3–10−5 per base per replication) [18]–[22], despite the large difference in mutability of their genomes (approximately one to three orders of magnitude difference), were capable of changing their phenotypes to coexist in an arms race. Further studies linking the phenotype mutability and genome complexity will help to elucidate the dynamic host–parasite relationship.
The E. coli HL2 strain was used as the coculture host strain and A/λ [46] was used as an indicator strain for the titer assay. The E. coli HL2 strain was constructed by conjugation with DH1ΔleuB::(gfpuv5-Kmr) [43] and HB2151 [47]. We mixed log-phase DH1ΔleuB::(gfpuv5-Kmr) and HB2151 for 2.5 hours and screened for kanamycin-resistant clones on LB agar medium supplemented with 25 µg/ml kanamycin. F′ retention of HL2 was checked by PCR with the primers TraU_f (5′-ATGAAGCGAAGGCTGTGGCT-3′) and TraU_r (5′-GCAGCTTGAACGCCATGCGT-3′) and the ability of HL2 to amplify Qβ was confirmed. Before the evolution experiments, HL2 was grown in mM63gl (62 mM K2HPO4, 39 mM KH2PO4, 15 mM ammonium sulfate, 1.8 µM FeSO4·7H2O, 15 µM thiamine hydrochloride, 2.5 mM MgSO4·7H2O, 0.04% glucose, and 1 mM l-Leu) for several passages until the specific growth rate had become stable, and the strain with stable growth rate was used as the ancestor strain (Anc(C)). The OD600 of stationary-phase Anc(C) cultured in mM63gl medium was approximately 0.4 (approximately 3×108 CFU/ml) because of glucose limitation. Qβ was kindly provided by Dr. Koji Tsukada (Osaka University, Japan), which was generated from Qβ genomic cDNA [23]. Qβ particles were diluted with LB medium and plaque assay was performed according to the standard method [48]. Polypropylene centrifuge tubes (15 ml, No. 430791; Corning Incorporated, Corning, NY) treated with 0.1% BSA for at least 15 minutes to prevent attachment of phages to the tube walls were used for all the experiments as culture tubes.
Copropagation experiment: 4.8×107 cells Anc(C) and 5.1×107 PFU Anc(P) were mixed and copropagation was started in a culture volume of 3 ml at 37°C with shaking at 160 rpm. Mixed cultures were divided into 2 lines on the 18th day, equivalent to 59 replication generations, and propagated independently for a further 36 days (line 1 and line 2 in Figure 1B). Serial transfer was conducted by daily transfer of the cultures with cells and phages. The portion of cultures calculated based on the final OD600 were transferred into fresh medium with dilution to an initial OD600 of 0.05. Daily culture samples were divided into thirds: one for preparing −80°C frozen stocks with 15% glycerol, one for CFU determination by dilution and spreading on low divalent cation mM63gl agar medium with 0.2 mM MgSO4·7H2O, and the other for PFU analysis using the supernatant after centrifugation. Qβ propagation experiment: Two lines (line 3 and line 4 in Figure 1B) were independently propagated from Anc(P) for 18 days, equivalent to 168–169 replication generations, at 37°C with shaking at 160 rpm. Serial passages consisted of infection of a host culture, followed by about 6 h of phage growth, and extraction of the phage from the culture. Each serial passage was performed as follows: uninfected Anc(C) cultures were grown at 37°C overnight and transferred into new medium with dilution to OD600 of 0.03. When OD600 became 0.06–0.07 (approximately 1×107 CFU/ml) after 2–2.5 h, cells were infected with phage to approximately 1.0–2.0×107 PFU/ml from the previous passage. The cultures were grown for about 6 h. E. coli cells were removed by centrifugation, and the supernatant was subjected to filtration with 0.2 µm syringe filters (Minisart RC15 filters; Sartorius Stedim Biotech, Goettingen, Germany), and phage solution was stored at 4°C for infection on the next serial passage. The replication generation number of the phage population (n) was calculated as n = ln2 (Nf/Ni), where Ni and Nf are the phage density (PFU/ml) at the initial and final time points of each passage, respectively. The initial value (Ni) was calculated by dividing the Nf of the previous passage by the dilution rate.
The evolved E. coli populations (M54(C), M163M(C), and M165_2(C)) and Qβ phage populations (M54(P), 163(P), and M165_2(P)) were purified from mixed cultures to analyze the phage genome sequence and to determine their fitness. To purify the evolved E. coli population, cultures stocked at −80°C including evolved E. coli and phage were streaked on mM63gl agar medium and then passaged several times in low divalent cation medium, 0.2 mM MgSO4·7H2O mM63gl, to prevent further phage adsorption to E. coli. We checked the purity of evolved E. coli by confirming that no plaques were observed in the passaged and chloroform-treated cultures. To purify the evolved phage population, cultures stocked at −80°C including evolved E. coli and phage were cultured in mM63gl at 37°C with shaking at 160 rpm for 1 day and filtrated with 0.2 µm syringe filters (Minisart RC15 filters; Sartorius Stedim Biotech). These particles were used for RNA genome sequencing analysis as described below. For analysis of phage fitness, these filtrated particles and Anc(P) were dialyzed to remove carry-over glycerol from the −80°C stock using Microcon centrifugal filter devices with 10,000 nominal molecular weight limit membranes (Millipore, Billerica, MA). We confirmed that the dialysis step did not affect plaque forming ability.
The RNA genomes of the Qβ population noted in Figure 1B, i.e., 8 kinds of genome derived from approximately 108 PFU particles of the Anc(P), M54(P), M109(P), M163(P), M165_2(P), S94_3(P), S169_3(P), and S168_4(P) phage populations, were extracted using a QIAamp Viral RNA mini kit (Qiagen, Hilden, Germany) according to the manufacturer's instructions. To analyze the full-length RNA genome sequence, samples were prepared as follows. Poly(A) was added at the 3′ end using poly(A) polymerase (Applied Biosystems/Ambion, CA). cDNA was synthesized using the primer Qt (5′-CCAGTGAGCAGAGTGACGAGGACTCGAGCTCAAGCTTTTTTTTTTTTTTTTT-3′) with SuperScript™ III Reverse Transcriptase (Invitrogen, Carlsbad, CA), and then RNA was degraded with RNaseH. The first-strand cDNA was purified and poly(A) was added at the 3′ termini of the cDNA with terminal deoxynucleotidyl transferase (Roche Diagnostics, Basel, Switzerland) and dATP. The cDNA with poly(T) at the 5′ terminus and poly(A) at the 3′ terminus was purified. To obtain the 5′ end of the Qβ phage genome sequence, second-strand DNA was prepared using Qt primer and PfuUltra II Fusion HS DNA Polymerase (Stratagene, La Jolla, CA). PCR was performed separately with high fidelity DNA polymerases for the whole Qβ genome divided into 5 regions as shown in Table S3. In total, PCR products were obtained from cDNA template derived from approximately 106 PFU particles. The templates, primers, and polymerase used for PCR and the primers used for sequencing are listed in Table S3. Sequencing was performed by the dideoxynucleotide chain termination method [25] on both strands, but we conducted direct sequencing in one direction using 2 sets of primers only for the 5′ end of the genome (Table S3). When a double peak appeared in the sequencing chart, positions where the height of the smaller peak was over half that of the larger peak were defined as polymorphic sites.
Genomic DNA was extracted from over 109 cells of Anc(C), M54(C), and M163(C) using a DNeasy Blood & Tissue Kit (Qiagen) according to the manufacturer's instructions. The genomic DNA of M163(C) was sequenced with an Illumina GAIIx (Illumina) using 51-bp of single-read format by Hokkaido System Science Co., Ltd. (Sapporo, Hokkaido, Japan). The GAIIx produced 18,718,549 reads and 954,646 kb. All reads were aligned to the reference sequence of E. coli DH1 genome sequence (GenBank accession number, CP001637.1; genome size, 4,630,707 bp) and F plasmid sequence (GenBank accession number, NC_002483.1; size, 99,159 bp) using MAQ [49] guaranteed to find alignments with up to 2 mismatches in the first 24 bp of the reads and 5 mismatches in 51 bp. Mean depth was 196 dividing 927,788 kb mapped bases by 4,730 kb of total reference sequences. After mapping and consensus base calling, SNPs were filtered with the same threshold values as reported previously [49]. The alignment view was also confirmed with Mapview [50], and SNPs were also scored with the following parameters: Phred quality score ≥20, variant frequency ≥0.40, coverage sum ≥5. In addition, Tablet [51] also has the ability to align short reads to the reference sequence, thus allowing us to score sites with a deletion, insertion, and/or regions with large insertions or deletions. SNPs, insertions, and deletions were scored for the F plasmid and for the M163(C) genome, and they were confirmed by sequencing using the dideoxynucleotide chain termination method for the genomes derived from over 107 cells of Anc(C) and M163(C) using the primers listed in Table S1. Two different positions observed in Anc(C) and M163(C) were sequenced for M54(C) by the dideoxynucleotide chain termination method.
The cross-coculture experiment, time course analysis after infection, and adsorption rate constant analysis were conducted according to the methods described in Text S1.
Fitness and the time to reach the stationary phase were compared by one-way ANOVA with the post hoc Tukey test [52]. The virulence of Anc(P) and S94_3(P) to Anc(C) was compared with Welch's t test. The differences in growth rates calculated from semi-logarithmic plots of E. coli or phage densities were tested by two-tailed t test and ANCOVA with the post hoc Tukey test [53].
|
10.1371/journal.pgen.0030056 | Lifespan Regulation by Evolutionarily Conserved Genes Essential for Viability | Evolutionarily conserved mechanisms that control aging are predicted to have prereproductive functions in order to be subject to natural selection. Genes that are essential for growth and development are highly conserved in evolution, but their role in longevity has not previously been assessed. We screened 2,700 genes essential for Caenorhabditis elegans development and identified 64 genes that extend lifespan when inactivated postdevelopmentally. These candidate lifespan regulators are highly conserved from yeast to humans. Classification of the candidate lifespan regulators into functional groups identified the expected insulin and metabolic pathways but also revealed enrichment for translation, RNA, and chromatin factors. Many of these essential gene inactivations extend lifespan as much as the strongest known regulators of aging. Early gene inactivations of these essential genes caused growth arrest at larval stages, and some of these arrested animals live much longer than wild-type adults. daf-16 is required for the enhanced survival of arrested larvae, suggesting that the increased longevity is a physiological response to the essential gene inactivation. These results suggest that insulin-signaling pathways play a role in regulation of aging at any stage in life.
| The lifespan of an animal is determined by both environmental and genetic factors, and many of the mechanisms identified to increase lifespan are evolutionarily conserved across organisms. Previous longevity screens in C. elegans have identified over 100 genes, but ∼2,700 essential for normal development were excluded from analysis. Paradoxically, these essential genes are five times more likely to be highly conserved in phylogeny than genes with no obvious developmental phenotypes. We screened these 2,700 essential genes for increased adult lifespan by initiating the gene knockdown once the animal had reached adulthood, thus bypassing earlier developmental roles. We identified 64 genes that can extend lifespan when inactivated postdevelopmentally. More than 90% of the genes we identified are conserved from yeast to humans. Many of the newly identified longevity genes extend lifespan as robustly as the most well-characterized longevity mutants. It is possible that the homologues of these genes may also regulate lifespan in other organisms as well. Genetic analysis places some of these genes in known pathways regulated by insulin-like signaling, although many of these gene inactivations function independently of this mechanism of lifespan extension. Surprisingly, a subset of these gene inactivations that induce potent developmental arrest also facilitate enhanced survival in the arrested state, suggesting that aging at any stage may be subject to regulatory control.
| The lifespan of an organism is regulated by both genetic and environmental influences in many species [1]. Recent work has identified specific components from a variety of cellular processes that regulate lifespan. In C. elegans loss-of-function mutations in the insulin/insulin-like growth factor-1/daf-2 signaling pathway can more than double the lifespan of an animal [2–6]. The regulation of lifespan by DAF-2 occurs during adulthood [7]. The insulin-signaling pathway negatively regulates the forkhead (FOXO) transcription factor DAF-16, which ultimately functions to both positively and negatively regulate transcription of metabolic, chaperone, cellular defense, and other genes [8–11]. The regulation of lifespan through an insulin-like signaling cascade is an evolutionarily conserved mechanism and has been demonstrated in flies and mice [12–15]. The regulation of DAF-16 activity is also modulated by the JNK signaling pathway, the SIR-2.1 deacetylase, and HSF-1, LIN-14, and SMK-1 in the nucleus [16–20].
In many organisms the rate of aging is tied to reproduction. In C. elegans germline proliferation produces a DAF-16 and KRI-1 mediated signal that negatively regulates lifespan while the somatic gonad promotes lifespan extension [21–23].
Caloric restriction (CR) also extends lifespan across species including yeast, worms, flies, and mice [24–29]. The sir-2.1 and let-363 genes in C. elegans regulate lifespan via CR [30,31]. Unlike insulin signaling, in flies, imposing CR at anytime can increase lifespan [32]. Finally, perturbations in mitochondrial function have been shown to increase lifespan [33–36]. Diminished function of the mitochondrial electron transport system can further extend mutations in the insulin-like signaling pathway, however mutations in the ubiquinone biosynthesis gene clk-1 do not further increase the longevity phenotype from CR. The mechanism of longevity induced by defective mitochondria is thought to occur during development, as previous attempts to use RNA interference (RNAi) to inhibit mitochondrial function in adults has not been shown to increase lifespan [34].
Recent genome-wide RNAi screens for increased longevity have identified ∼100 potential regulators of lifespan in C. elegans from diverse cellular pathways, many of which are evolutionarily conserved [37,38], but genes essential for viability are underrepresented in genome-wide RNAi screens for postdevelopmental phenotypes such as aging. These RNAi screens for adult longevity preclude the identification of gene inactivations that cause lethality, larval arrest, sterility, and/or other developmental pleiotropies (i.e., essential genes), unless these genes are inactivated after their required developmental roles.
Here we report 64 genes that, when inactivated postdevelopmentally by RNAi, increase adult lifespan. We also report the enhanced survival phenotype of the animals arrested during development by these essential gene inactivations.
To identify essential genes that function in adulthood to regulate lifespan, we selected 2,700 RNAi clones that, if fed from the previous generation or from the L1 larval stage, cause arrest at embryonic or larval stages, and screened them for increased lifespan after initiating RNAi at the L4 larval/young adult stage (Figure 1). We found three observations that validate this approach: First, C. elegans lifespan can be extended when fed dsRNA targeting the insulin receptor/daf-2 at any developmental larval stage through adulthood [7]; second, conditional daf-2 alleles cause dauer arrest if the gene is inactivated at the L1 stage but increased longevity if inactivated in adults [39]; and third, null alleles of daf-2 are lethal [40] demonstrating the necessity of uncoupling the developmental and aging phenotypes of essential genes as well as the utility of producing non-null phenotypes by RNAi.
We performed the RNAi screen utilizing the eri-1(mg366) [41] strain to improve RNAi in all cells types including neurons, which are normally refractory to RNA interference. In fact, many of the longevity genes we identify are expressed in the nervous system, which has been implicated in insulin regulation of longevity (Figure S1; Table S1) [42].
Gene inactivations that caused increases in mean lifespan of at least 10% were scored as positive in our screen. Endocrine signaling, stress adaptation, metabolism, and reproduction are potent and evolutionarily conserved regulators of aging [43]. Our screen identified genes from these canonical longevity-promoting pathways in addition to pathways not previously implicated in aging (Figures S2 and S3; Table 1).
Essential genes are more conserved in phylogeny than genes with no obvious developmental phenotype; more than 90% of the genes we identify are conserved from yeast to humans. The theory of antagonistic pleiotropy suggests that any genes that function in postreproductive longevity control should be under natural selection at prereproductive stages. Our data support this notion: our yield of major lifespan regulators is 64 gene inactivations out of 2,700 tested (∼2.4%), a 4-fold increased yield compared to the previous 89 gene inactivations out of 16,000 screened (∼0.6%), and a higher proportion of the gene inactivations cause large increases in longevity (percent increase compared to vector control denoted in parentheses for each RNAi clone), although the use of an enhanced RNAi strain may also contribute to the increased sensitivity.
One established method of validating genes that emerge from RNAi screens is to test gene knockouts for the same phenotype. For these essential genes, it may prove possible to study the longevity of the arrested animals. However, to study adult longevity conditional alleles will be required. Such alleles tend to emerge from detailed genetic analysis rather than genomic knockout projects and are not currently available.
The aging research community has characterized several central mechanisms that mediate lifespan regulation including insulin/insulin-like growth factor signaling, CR, and mitochondrial function. To classify the pathways represented by these new genes, we performed secondary assays: DAF-16 localization, sod-3 expression, arrested larval survival, suppression of polyglutamine aggregation, and aberrant fat metabolism and clustered the genes by the phenotypes observed (Figure 2; Figure S4; Table 2; Tables S1 and S2).
In this study, we have screened a large portion of the genome previously underrepresented in genome-wide based RNAi screens for aging phenotypes and identified 64 genes that normally function to shorten lifespan. Characterization of these gene knockdowns by epistasis with the insulin-signaling pathway and use of biomarkers of aging places them into distinct classes. Because these molecules are predicted to function in complex cellular pathways (Figure S5), future work will focus to dissect the mechanisms employed by these essential processes to regulate lifespan.
Strains were maintained and cultured using standard techniques [65]. We used the following C. elegans strains and mutant alleles: wild-type N2 Bristol, eri-1(mg366)IV, daf-16(mgDf47)I;eri-1(mg366)IV; zIs356, (daf-16p::daf-16::gfp; rol-6[su1006]); muIs84, (sod-3p::gfp); huIs33, (sod-3p::gfp; rol-6[su1006]); and rmIs133: (unc-54p::Q40yfp).
Eggs were isolated from gravid eri-1(mg366) worms and synchronized by hatching overnight in the absence of food. The synchronized L1 larvae were then placed on OP50-containing agar plates and allowed to develop to L4-stage larvae at 20 °C. The L4-stage larvae were washed thoroughly, cleaned by sucrose flotation, and placed on 12-well plates with Escherichia coli expressing double-stranded RNA (dsRNA) (described below). We carried out a large-scale RNAi screen using the enhanced RNAi strain eri-1(mg366) (Figure S1). Briefly, each RNAi colony was grown overnight in LB with 50 μg/ml ampicillin and then seeded onto 12-well RNAi agar plates containing 5 mM isopropylthiogalactoside (IPTG). The RNAi bacteria were induced overnight at room temperature for dsRNA expression. We then added ∼30 synchronized L4-stage animals to each well, allowed worms to develop to adults, and then added 5-fluorodexoyuridine (FUdR) solution to a final concentration of 0.1 mg/ml. Worms were kept at 20 °C, and their lifespan was monitored. Worms feeding on bacteria carrying the empty vector or targeting eri-1 were used as negative controls. At least 96 wells of the empty vector control were included. At the time when all of the control worms were dead, each well containing the different RNAi bacteria was scored for live worms. RNAi wells in which live worms were observed were scored as positives. RNAi clones that were scored as positive in the first and second passes of screening (Figure S1) were retested in duplicate during a third pass using a conventional longitudinal RNAi lifespan assays (see below).
RNAi bacteria were prepared as described above. daf-2, eri-1, and daf-16 RNAi clones were included as positive and negative controls in blind fashion in the conventional lifespan assays (daf-2 RNAi construct kindly provided by M. Vidal, Harvard Medical School, Boston, Massachusetts, United States). Synchronous L4/young adult stage eri-1(mg366) animals treated with FUdR were prepared as above and placed on 6-well plates containing the positive RNAi clones and controls in duplicate. Approximately 40–60 adult animals were scored in each well (two wells for each RNAi clone, in two biological replicates). The animals were kept at 20 °C and scored every two days by gentle prodding with a platinum wire to test for viability. To ensure the continued efficacy of RNAi knockdown, animals were fed freshly induced RNAi bacteria every five to seven days. Lifespan is defined as the first day of adulthood (adult lifespan = 0) to when they were scored as dead. Worms that died of protruding/bursting vulva, bagging, or crawling off the agar were censored from the analysis. For epistasis analysis, RNAi lifespan assays were performed as described above, except that daf-16(mgDf47);eri-1(mg366) worms were used.
Statistical analyses were performed using SPSS software (http://www.spss.com). The survival experience of each RNAi-treated population is compared with that of the population treated with control RNAi using the log rank test. A p-value <0.05 was considered as significantly different from control.
Synchronized L1-stage eri-1(mg366) or eri-1(mg366);daf-16(mgDf47) animals were placed on RNAi clones that induce larval arrest phenotypes at 20 °C. To ensure that any animals bypassing lethality did not reproduce the plates were moved to 25 °C to exploit the temperature-sensitive sterility associated with eri-1(mg366). The wells were qualitatively monitored every two days for arrested larvae survival.
Synchronized L1-stage animals or freshly egg-prepped embryos carrying the integrated transgenes sod-3p::gfp, daf-16p::daf-16-gfp, and myo-3p::Q40yfp were each placed onto RNAi bacteria as described above. The fluorescence intensity of each population was monitored two and three days following RNAi treatment.
Nile Red experiments were performed as previously described [66] except that L4-stage eri-1(mg366) animals were fed RNAi clones induced on plates containing 5 mM IPTG.
Promoter elements were amplified by PCR from wild-type genomic DNA and fused to RFP [67]. The transcriptional reporters were then injected into the gonads of wild-type adult hermaphrodites. Transgenic animals harboring the extrachromosomal arrays were then imaged.
The primer pairs used to amplify promoter elements are as follows:
c56g2.1: F: 5′-CAGACAGGTGAAGCTGAGCGTGGC-3′, R: 5′-AAACGCAGAAAACGTCGGTGACGGAATG-3′;
egl-45: F: 5′-CCAGCCAGGAAAAATCGATTATATTAAG, R: 5′-AGTTGTGCTCGGATTACCGCTGAATTG-3′;
unc-62: F: 5′-CCCTGAAATTGTTGCGAAAGTTTCTG, R: 5′-GTTCCTGCAAGAGAGAAATATTAAATTTTG-3′;
htp-3: F: 5′-CTCCCGAAGATTCCGCATTTGCTC, R: 5′-TTTGACACTTAAAATATTTTAAAACATTTTTTTTAA-3′;
f26a3.4: F: 5′-GACCGGAACAGGTGGGCAATGTCGAC, R: 5′-TTTGTAGTGTATCTGTAATCATATTAAATTTGATTC-3′;
inf-1: F: 5′-GTATGTGTTTATGGTGTGTGCACAAG, R: 5′-GACAGGTGGGTTGAAAAGTTAAAAATTAAC-3′;
f08b4.1: F: 5′-CTAACCGATTCCTCAAGCCACGTGGG, R: 5′-TTTTGATTATTGATATTTCATTCGAATTTGCCAG-3′;
y54e10br.4: F: 5′-CTCTCCCGATTCCGCCATAATGCCCG, R: 5′-TCTCTGAAATATCGAAAAGAAATGAGATAATTG-3′;
sem-5: F: 5′-GGGTTAGAGCACTCTTAATGAGTCATG, R: 5′-CGTCTCGCTACCTGAAATATACTCTT-3′;
zk686.2: F: 5′-GGTTCCGGAGATAACCAAGCAGTATTGG, R: 5′-TATCTGGAGAAATTAAAATATGAACCAAAAAATGCG-3′. |
10.1371/journal.pntd.0004555 | Trypanosoma cruzi Needs a Signal Provided by Reactive Oxygen Species to Infect Macrophages | During Trypanosoma cruzi infection, macrophages produce reactive oxygen species (ROS) in a process called respiratory burst. Several works have aimed to elucidate the role of ROS during T. cruzi infection and the results obtained are sometimes contradictory. T. cruzi has a highly efficiently regulated antioxidant machinery to deal with the oxidative burst, but the parasite macromolecules, particularly DNA, may still suffer oxidative damage. Guanine (G) is the most vulnerable base and its oxidation results in formation of 8-oxoG, a cellular marker of oxidative stress.
In order to investigate the contribution of ROS in T. cruzi survival and infection, we utilized mice deficient in the gp91phox (Phox KO) subunit of NADPH oxidase and parasites that overexpress the enzyme EcMutT (from Escherichia coli) or TcMTH (from T. cruzi), which is responsible for removing 8-oxo-dGTP from the nucleotide pool. The modified parasites presented enhanced replication inside murine inflammatory macrophages from C57BL/6 WT mice when compared with control parasites. Interestingly, when Phox KO macrophages were infected with these parasites, we observed a decreased number of all parasites when compared with macrophages from C57BL/6 WT. Scavengers for ROS also decreased parasite growth in WT macrophages. In addition, treatment of macrophages or parasites with hydrogen peroxide increased parasite replication in Phox KO mice and in vivo.
Our results indicate a paradoxical role for ROS since modified parasites multiply better inside macrophages, but proliferation is significantly reduced when ROS is removed from the host cell. Our findings suggest that ROS can work like a signaling molecule, contributing to T. cruzi growth inside the cells.
| The parasite Trypanosoma cruzi is the causative agent of Chagas’ disease, which affects 10 million people, mainly in Latin American. Macrophages are one of the first cellular actors facing the invasion of pathogens and during T. cruzi infection, produce reactive oxygen species (ROS). To deal with oxidative stress, T. cruzi has an antioxidant machinery and, to repair DNA damage triggered by ROS, this parasite possesses enzymes of the oxidized guanine DNA repair system. The understanding of the role of ROS in the infection by T. cruzi can provide us with good insights on T. cruzi biology and virulence. While some studies suggest that ROS is related to parasite control, others have demonstrated that ROS is important for proliferation of this parasite. To investigate the contribution of ROS in T. cruzi infection, we utilized mice deficient in the production of ROS (Phox KO) and parasites that overexpress the enzymes related to DNA repair. Our results show that ROS is not only important for the battle against pathogens, but suggest that ROS can also work as a signal that contributes to the growth of this parasite.
| Macrophages are one of the first lines of defense against intracellular pathogens [1]. During Trypanosoma cruzi infection, these cells are activated to produce ROS, a process called respiratory burst [2–4]. The detection of infectious agents leads to activation of the membrane bound NADPH oxidase, a multi-subunit complex that utilizes NAD(P)H as an electron donor to reduce oxygen (O2) to superoxide (O2●−) within the phagosome. The anionic nature of O2●− restricts its diffusion through membranes, confining its actions to the site of formation. Superoxide radicals can spontaneously or enzymatically dismutate into hydrogen peroxide (H2O2), an oxidant with higher diffusional capacity. Metal transition ions in the presence of H2O2 can generate hydroxyl radical (●OH), an oxidant that, owing to its high reactivity, encloses poor selectivity against cellular targets and may not be highly toxic [5,6]. Alternatively, O2●− may react with iNOS-derived nitric oxide (●NO) in a diffusion-controlled reaction, to produce peroxynitrite (ONOO−), a cytotoxic effector molecule against T. cruzi [4,7,8].
T. cruzi has a highly efficiently regulated antioxidant machinery to deal with the oxidative burst and adapts to the conditions imposed by their digenetic life cycle [9,10]. There are different pathways to detoxify hydroperoxides, within different substrate specificities and in different compartments such as mitochondria, glycosome, endoplasmic reticulum and cytosol [11]. In this intricate network, reducing equivalents from NADPH, produced by the pentose phosphate pathway, are delivered to a variety of detoxification enzymes. This reducing equivalents are delivered from trypanothione (T(SH)2) to tryparedoxin (TXN) and glutathione (GSH), which transfers them to the several peroxidases. T(SH)2 is maintained in its reduced state by the NADPH-dependent trypanothine reductase (TcTR) [12]. Several peroxidases have been characterized: two cysteine-dependent glutathione peroxidases, one ascorbate-dependent hemoperoxidase (TcAPX) and two tryparedoxin peroxidases [13–17]. The tryparedoxin peroxidases differ in their subcellular location: a cytosolic and a mitochondrial form (TcCPX and TcMPX, respectively) and catalyze the reduction of H2O2, small-chain organic hydroperoxides and ONOO− [11,18]. In the endoplasmic reticulum, TcAPX and glutathione-dependent peroxidase II (GPX-II) metabolize H2O2 and lipid hydroperoxides respectively [16,19]. There is growing evidence that this antioxidant network may play an important role in parasite virulence and success of infection [4,11,20–22].
Despite this efficient antioxidant system, the parasite macromolecules, particularly DNA, may still suffer oxidative damage that may be deleterious if not repaired. Due to its low redox potential, guanine (G) is the most vulnerable base. The oxidation of guanine results in formation of 8-oxo-7,8-dihydroguanine (8-oxoG), a cellular marker of oxidative stress [23]. When 8-oxoG assumes syn configuration, it is particularly mutagenic because it functionally mimics thymine. When 8-oxoG is inserted during DNA replication, it can generate double-strand breaks, making this lesion deleterious [24,25]. To repair lesions caused by 8-oxoG, most organisms possess the oxidized guanine (GO) DNA repair system (GO system). This repair pathway is composed by the enzymes MutT, MutY and MutM in bacteria [26] and by corresponding enzymes MTH1, MUTYH and OGG1 in humans [27]. Studies on T. cruzi genome have demonstrated that this parasite has homologs to the enzymes OGG1, MUTYH [28] and MutT [29]. MutM excises the oxidized base from 8-oxoG:C base pairs and the MutY excises adenine where it has been erroneously incorporated opposite to unrepaired 8-oxoG during replication [27]. The MutT enzyme catalyzes the hydrolysis of 8-oxo-dGTP in the nucleotide pool, by substitution at the rarely attacked beta-P, to yield monophosphate nucleotide and pyrophosphate. This prevents errors in DNA replication, since the monophosphate form cannot be incorporated into nascent DNA [30,31].
Although ROS is clearly involved in control of several infections, increasing evidences point to a role of ROS as promoters of infection [32]. Several works have aimed at elucidating the role of reactive oxygen species during T. cruzi infection and the results obtained are sometimes contradictory. Although some studies have suggested that ROS produced during the respiratory burst have an important role in T. cruzi control [2,4,10,33], other authors have demonstrated that ROS is important to cellular signaling and proliferation of this parasite [11,34–36]. Studies performed with bacteria [37,38], fungus [39], viruses [40–42] and Leishmania [43] also associated ROS and parasite proliferation. In order to investigate the contribution of ROS in T. cruzi infection, we utilized mice deficient in the gp91phox (Phox KO) subunit of NADPH oxidase [44] and parasites that overexpress the enzyme EcMutT or TcMTH and are more resistant to DNA damage caused by ROS (parasites with the E. coli mutT gene and parasites that overexpresses TcMTH gene, a T. cruzi MutT homolog) [29]. We found that modified parasites multiply better inside macrophages than wild type, but their proliferation is significantly reduced when ROS production is inhibited in host cell. Our results suggest that low concentration of ROS may work like a signaling molecule, contributing to growth of T. cruzi inside the cells in vitro and increasing the levels of parasitemia in vivo.
This study was conducted in strict accordance with the recommendations in Guide for the Care and Use of Laboratory Animals of the Brazilian National Council of Animal Experimentation (http://www.cobea.org.br/) and the Federal Law 11.794 (October 8, 2008). All animals were handled in strict accordance with good animal practice as defined by the Internal Ethics Committee in Animal Experimentation (CETEA) of the Universidade Federal de Minas Gerais (UFMG), Belo Horizonte, Minas Gerais, Brazil. The protocol number 214/11 was approved by CETEA.
T. cruzi epimastigotes (CLBrener, wild-type) were cultured at 28°C in BHI (brain heart infusion) medium. Parasites overexpressing EcMutT, TcMTH and parasites transfected with the empty vector pROCK (TcROCK) were obtained as described previously [29,45]. Transformed cells were cultured in BHI medium containing 250 μg·ml−1 of hygromycin (Sigma Aldrich, St. Louis, MO, USA). T. cruzi trypomastigotes were obtained from the supernatant of infected mono layers of LLC-MK2 cell cultures (grown in 2% FBS, 1% penicillin-streptomycin and 2 mM glutamine supplemented DMEM (Dulbecco´s Modified Eagle´s Medium, Sigma Aldrich) and purified by incubation of the pellet for 2 hours at 37°C, followed by collection of motile infective trypomastigotes in the supernatant. This project was approved by National Technical Biosafety Commission (CTNBio) under the process number: 01200.003883/97-02.
Four- to 8-week-old male and female C57BL/6 mice were obtained from CEBIO (Instituto de Ciências Biológicas, Universidade Federal de Minas Gerais.Belo Horizonte, MG, Brazil). Phox KO [44] and IFN-γ KO [46] mice were purchased from The Jackson Laboratory (Glenville, NJ, USA). Mice were kept in conventional conditions with barriers, controlled light cycle and controlled temperature. Animals were fed a commercial diet for rodents (Labina, Purina, SP, Brazil) ad libitum.
The macrophages used in this study were isolated from the peritoneal cavity of mice 4 days after injection of 2 mL of 3% thioglycollate medium (BD, Le Pont de Claix, France) into the peritoneal cavity. After this time, mice were euthanized and the peritoneum cells were harvested by repeated cycles of aspiration and re-injection with 10 ml of cold PBS in 10ml syringe with a 24G needle. More than 80% of the cells harvested were macrophages. The cells were centrifuged at 4°C, 1,500 g for 10 minutes and re-suspended in DMEM supplemented with 10% fetal bovine serum (FBS) (Cultilab, Campinas, SP, Brazil), 1% penicillin-streptomycin and 2mM glutamine. Macrophages were counted in a hemocytometer prior to seeding 5x105 or 1x106 cells into each well of a 24-well or 72-well plate respectively and incubated at 37°C, 5% CO2 for 2 hours. The parasites were purified, counted and diluted in DMEM medium, and infection was performed for 2 hours, at a five-parasite-to-one-macrophage ratio. Immediately after macrophage infection, the cells were washed four times with phosphate-buffer saline (PBS, pH 7.3) to remove extracellular parasites. The cells were fixed or reincubated with medium for 48 and 72 hours before fixation with methanol. Coverslips with attached macrophages were stained with Panótico (Laborclin, Pinhais, PR, Brazil) and a minimum of 300 macrophages per coverslip were counted. The results were expressed as an infection index ([percentage of infected macrophages x number of amastigotes]/total number of macrophage). Cells from 96 well plates were used to count released parasites in the supernatant (3–7 days after infection). The following drugs were used in these assays: apocynin (APO) (300μM; Sigma-Aldrich); N-acetyl-cysteine (NAC) (1mM; Sigma-Aldrich); H2O2 (100 μM); superoxide dismutase–polyethylene glycol (SOD) (25 U/well, Sigma-Aldrich) and catalase–polyethylene glycol (CAT) (40 U/well, Sigma-Aldrich). Drugs were added to the cells 30 minutes (H2O2) or 2 hours (apocynin, catalase, NAC, SOD-PEG) before and immediately after infection. Parasites were treated with 100μM H2O2 for 30 minutes before the infection.
Luminometry assays were performed to evaluate the production of ROS by macrophages. The cells, obtained as described before, were centrifuged at 4°C, 1,500 g for 10 minutes, and resuspended in complete RPMI without phenol red. Macrophages (1 x 106 cells/well) were plated in 96 well opaque plates (NUNC, Rochester, NY, USA) and pre-incubated with 300μM of APO, 1mM of NAC, 25u of SOD or 40u of CAT for 2 hours. After this time, 0.05 mM luminol (5-Amino-2,3-dihydro-1,4-phthalazinedione; Sigma-Aldrich) and T. cruzi trypomastigotes (Y strain) in the proportion of 10 parasites to 1 macrophage were added in each well. Measurements were taken for 120 minutes with two-minute interval between measurements. Production of ROS was assayed by the light intensity generated by the reaction between ROS and luminol and expressed as relative light units.
Macrophage iNOS was induced by pre-incubating the cells with 100 units/mL of IFN-γ (BD, San Diego, CA, USA) and 10μg/mL of LPS (Invivogen, San Diego, CA, USA) for 2 hours. Then, control and IFN-γ/LPS-activated macrophages were infected with T. cruzi trypomastigotes (5 parasites: 1 host cell) for 2 hours and washed with phosphate-buffer saline (PBS, pH 7.3) to remove extracellular parasites. Following incubation (for 48 h), supernatants were collected and the concentration of nitrite was determined spectrophotometrically (Microplate Spectrophotometer System, model SPECTRAmax 340, Molecular Devices, Sunnyvale, CA, USA) at 540nm using the Griess method with NaNO2 as the standard [47].
T. cruzi epimastigotes (TcWT, EcMutT and TcMTH) were treated with 200μM of H2O2 for 30 minutes. After incubation, cells were centrifuged at 800 g for 10 min at 25°C and washed twice in DPBS (Dulbecco’s PBS, pH 7.3; Sigma-Aldrich). Parasites (1 × 109 cells/mL) were incubated for 30 min at 28°C in DPBS containing 50 μM DHR (Molecular Probes, Life Technologies, Eugene, OR, USA). After incubation, cells were centrifuged at 800 g for 10 min at 25°C and washed twice in DPBS in order to eliminate non-incorporated DHR. Detection of intracellular Rhodamine 123 (RH 123), the oxidation product of DHR, was performed after exposure to the 0.1mM peroxynitrite donor 3-morpholinosydnonimine hydrochloride (SIN-1, Sigma-Aldrich). The detection of intracellular RH 123 was performed using a FACS-Calibur flow cytometer (Becton-Dickinson, Rutherford, NJ, USA).
Epimastigotes (3x108 cells) were replicated 3 days consecutively to maintain parasites in the logarithmic phase of growth and after this process were incubated with 50μM H2O2 for 30 minutes, washed twice and prepared for determination of antioxidant enzyme contents. Parasites (3x108 cells) were centrifuged at 800 g for 10 min at 25°C, washed three times in DPBS pH 7.3, re-suspended in 250 μL lysis buffer (10 mM Tris–HCl, 1 mM EDTA and 0.5% (v/v) Triton X-100) and incubated on ice for 15 min. Cell extracts were clarified (13,000 g for 30 min at 4°C) and supernatants supplemented with loading buffer (30 mM Tris–HCl, pH 6.6, 1% (w/v) SDS and 5% (v/v) glycerol) were stored at -80°C until use. Protein extracts (50 μg), were resolved by 15% SDS–PAGE and then blotted into nitrocellulose membranes (Hybond-C extra, GE Healthcare Life Sciences, USA). After transfer, proteins were stained with Ponceau-S solution (Aplichem, Daermstadt, Germany) and blocked using 3% dry milk in PBS for 1 h at 25°C. Membranes were then probed with anti-TcCPX (1:2000) diluted in PBS 0.1% (v/v) Tween 20 for 1 h at 25°C following 1 h incubation with anti-rabbit IRE-800 (LI-COR, Lincoln, NE, USA) diluted 1:10.000 in PBS 0.1% (v/v) Tween 20. Membranes were imaged with the LI-COR Odyssey Infrared Imaging System. Protein content relative to total protein loaded (Ponceau-S staining [48]) in the different extracts analyzed was determined by densitometric techniques using ImageJ (National Institute of Health, Bethesda, MD, USA). Results are expressed as relative enzyme content respect to total protein content [20]. To determinate the quantity of TcMPX and TcSODB, the parasites were fixed in paraformaldehyde (4% v/v in PBS) and incubated with anti-TcMPX, anti-SODB, and anti-cruzipain (1:2000) for 1 hour at 37°C. The parasites were washed and incubated for 1 h with Alexa Fluor 488 goat anti-rabbit IgG (Life Technologies, Eugene, OR, USA) diluted 1:10,000, washed again and analyzed by flow cytometer (FACS-Calibur).
Streptavidin has previously been shown to bind with high specificity to 8-oxoG [49,50] and was therefore used for the 8-oxoG measurements. Epimastigotes and amastigotes parasites were treated with H2O2 for 30 minutes, fixed in paraformaldehyde (2% v/v in PBS) at 25°C for 15 min and thereafter incubated for 15 min in PBS with 0.1% Triton X-100 v/v. Cells were then incubated with Alexa488-conjugated streptavidin (Invitrogen) (1:100) in PBS for 1 h at 37°C and evaluated by flow cytometer (FACS-Calibur).
T. cruzi trypomastigotes were maintained by blood passage in IFN-γ KO (TcWT and TcMTH strain) or Swiss (Y strain) mice every 7 or 9 days respectively. Trypomastigotes were obtained from heparinized blood, counted and used to infection. In some experiments, blood parasites (Y strain) were treated with 100μM H2O2 for 30 minutes before the infection. Experimental infection was performed in C57BL/6 WT and Phox KO mice by intraperitoneal injection of 106 TcWT, TcMTH or Y strain blood trypomastigotes. Parasitemia was assessed by counting trypomastigotes in 5 μL of tail vein blood, every day from the 3rd day post-infection until the time at which the parasites became undetectable. The number of parasites per mL was calculated as previously described [51]. Mortality of infected mice was monitored daily.
Statistical analysis in this work was performed using the GraphPad Prism 5.0 program (GraphPad Software Inc., CA, USA). Data are presented as the mean ± standard deviation (SD), and all experiments were repeated at least three times. Data were analyzed for significant differences using ANOVA, and differences between groups were assessed with Bonferroni post-test. The level of significance was set at p < 0.05.
Internalization of T. cruzi trypomastigotes by macrophages triggers the assembly of the NADPH oxidase complex to yield O2●− [4]. To establish that infection promotes respiratory burst in our conditions, we performed chemiluminescence experiments using luminol which can serve as a probe for O2●− and ONOO− [4]. Infection-increased chemiluminescence triggered by parasites was almost twice that observed with non-infected cells (Fig 1). Luminol chemiluminescence increase was not detected when we performed these experiments in Phox KO macrophages, due to the lack of phagocyte NADPH oxidase (phox) activation and thus O2●− production (Fig 1A and 1C). As expected, pretreatment of macrophages with the phox inhibitor apocynin, a compound that prevents p47phox subunit translocation and therefore assembly of the enzyme complex, prevented the increase in ROS induced by infection, and brought chemiluminescence values down, similarly to Phox KO macrophages (Fig 1A and 1B). Addition of the antioxidants NAC, SOD and CAT also reduced chemiluminescence intensity (Fig 1B and 1D). Once determined in our experimental setup that T. cruzi could stimulate ROS production by macrophages, we proceeded to investigate if oxidative stress would affect the course of infection.
To investigate the importance of oxidative stress on the success of infection of macrophages with T. cruzi, we used parasites that over-express EcMutT enzyme and are more resistant to DNA damage by the oxidation of guanine [29]. To investigate the influence of over-expression of MutT/MTH on T. cruzi invasion process in host cells, macrophages were exposed to parasites for two hours, washed to eliminate extracellular parasites, and fixed. EcMutT heterologous expression does not affect the invasion process (Fig 2A, 2B and 2C). The number of internalized trypomastigotes (Fig 2A) and the number of infected macrophages (Fig 2B) was similar between the two populations of parasites. However, after 48 hours, the number of infected macrophages (Fig 2B) and the number of amastigotes per macrophage (Fig 2A) was elevated for EcMutT parasites in comparison with TcWT parasites. To better express the obtained data the infection index was determined, considering simultaneously the number of infected macrophages and the number of amastigotes in relation to total macrophages. The infection index shows that EcMutT presented increased replication inside murine inflammatory macrophages when compared with wild-type parasites (Fig 2C). This enhanced replication of EcMutT parasites inside macrophages was corroborated by counting the number of trypomastigotes released at the supernatant of infected cells (Fig 2D). Hence, removal of 8-oxo-dGTP from the nucleotide pool increased the success of T. cruzi inside murine macrophages. A replicate of this experiment is presented in S1 Fig.
EcMutT parasites express E. coli MutT enzyme, whereas TcMTH parasites overexpress a T. cruzi MutT homolog [29]. Both parasites multiply better in macrophages than wild-type (TcWT) and the wild-type parasite transfected with the empty vector pROCK (TcROCK) (Fig 2 and [29]). To determine if EcMutT and TcMTH are resistant to oxidative stress and what would be the mechanism for this resistance, we evaluated DHR oxidation by flow cytometry in parasites exposed to peroxynitrite donor, SIN-1. Oxidation of the DHR loaded into epimastigotes into fluorescent rhodamine 123 indicates that SIN-1 reaches the parasite cytosol (Fig 3A). The highest intracellular DHR oxidation yield was obtained when TcWT epimastigotes were pre-incubated with H2O2. DHR oxidation was not increased in TcMTH and EcMutT pre-incubated with hydrogen peroxide and further challenged with peroxynitrite (Fig 3A). We observed that pre-conditioning with H2O2 did not promote increase in expression of cytosolic tryparedoxin peroxidase (Fig 3B and S2 Fig gel), but increases expression of mitochondrial tryparedoxin peroxidase (Fig 3C and [29]) in both TcWT and TcMTH. Interestingly, EcMutT over-expressed cytosolic tryparedoxin peroxidase. Superoxide dismutase B (Fig 3D) was similar among parasites and after pre-treatment with H2O2. Cruzipain expression was used as control (Fig 3E). Treatment with H2O2 did not alter parasite viability (S3 Fig).
We had already demonstrated that MutT/MTH-expressing cells contained fewer nuclear DNA lesions [29]. We now evaluated the accumulation of 8-oxoG in DNA after parasite exposure to H2O2. The modified base was detected using streptavidin conjugated with Alexa-488 by flow cytometer. We observed an increase of 8-oxoG in DNA after H2O2 treatment in TcWT epimastigotes. On the other hand, TcMTH epimastigotes did not show increased 8-oxoG in DNA after H2O2 treatment, demonstrating the functional activity of the MTH enzyme in the T. cruzi over-expressers (Fig 4).
Our results so far indicate that the over-expression of genes related to repair of oxidative damage favors the growth of parasites inside macrophages. To understand better the role of ROS in T. cruzi infection, we infected macrophages from Phox KO mice. Macrophages from these mice produced less ROS than cells from C57BL/6 WT mice upon infection (Fig 1A and 1C). Phox KO macrophages showed reduced parasitism, when infected with Y (Fig 5A and 5B) and CL Brenner (Fig 5C and 5D) strains of T. cruzi, as compared to C57BL/6 WT macrophages. After 48 hours of infection with Y strain, the number of parasites was increased in C57BL/6 WT macrophages, but was reduced in Phox KO macrophages (Fig 5A). In addition, the number of trypomastigotes released in the supernatant of C57BL/6 WT infected macrophages was greater than in Phox KO macrophages after infection with the Y strain (Fig 5B). The same result was obtained with the CL Brenner strain of T. cruzi (Fig 5C and 5D). We did not observe significant differences in ●NO production between Phox KO and C57BL/6 WT macrophages. In both cells, infection with T. cruzi did not induce the production of ●NO, which was similar to basal levels obtained by non-infected cells. The treatment with IFN-γ/LPS induced significant amounts of ●NO both in Phox KO and C57BL/6 WT macrophages. ●NO production was not different between Phox KO and C57BL/6 WT submitted to the same treatment. Additionally, ●NO production by previously stimulated cells was increased by T. cruzi infection (Fig 6).
Phox KO macrophages showed reduced parasitism when compared with C57BL/6 WT macrophages. Our next step was to investigate if ROS were responsible for the lack of growth in Phox KO macrophages. Thus, we inhibited ROS production by C57BL/6 WT macrophages using different antioxidants. Pre-treatment with anti-oxidants was performed in order to ensure the status of the macrophage at the time of infection. The antioxidants SOD-PEG, CAT-PEG, NAC, and apocynin all reduced parasitism in C57BL/6 WT macrophages. This inhibition was more striking when we infected macrophages with TcMTH, since the levels of infection obtained with this parasite were greater and easily visualized after 48 hours (Fig 7A and 7B). When we observe the number of trypomastigotes released in the supernatant, the differences in infection are still more striking, for both TcWT and TcMTH. The number of trypomastigotes released in the supernatant of infected macrophages is reduced after treatment with apocynin (Fig 7E). The treatment of macrophages with NAC and apocynin also reduced the parasitism of the cells after infection with Y strain (Fig 7C and 7D).
Our results suggest that exposure to ROS promotes parasite replication. Some works have demonstrated that ROS could act as signal molecules to cells [11,36]. We propose that T. cruzi needs a signal provided by ROS produced by macrophages to thrive in this host cell. To test this hypothesis, we treated Phox KO macrophages with H2O2 before and after infection, and evaluated the infection index. We also treated NAC-treated C57BL/6 WT macrophages with H2O2. In both cases, the infection index was increased after treatment with H2O2 (Fig 8A). To clarify further this issue, we treated T. cruzi with H2O2 30 minutes before infection. Our results show that parasites treated with H2O2 can infect Phox KO macrophages similarly to C57BL/6 WT macrophages (Fig 8B). In addition, treatment of parasites with H2O2 did not affect infection of C57BL/6 WT macrophages (Fig 8B). The results displayed in Fig 8A differ slightly from the ones presented in Fig 7, in that TcMTH parasites grew better in Phox KO macrophages than TcWT. This result was not repetitive, that is, in some experiments we observed this difference and in others we did not. The reason for the discrepancy is not clear to us at this point, but may be due to differences among parasite cultures obtained in different days. However, consistently parasites grew better in WT macrophages than in Phox KO macrophages, and TcMTH grew better than TcWT in WT machrophages.
So far, our results seem paradoxical: oxidative-stress-resistant parasites multiplied better inside macrophages, but cells deficient in ROS production did not sustained T. cruzi infection. There is a broad response to oxidants in cells: after exposure to a relatively high concentration of ROS, oxidative stress damage generally occurs, while lower concentrations can exert important physiological roles in cellular signaling and proliferation [11]. Thus, we treated parasites with different concentrations of H2O2 and used these parasites to infect macrophages and mice. We found no differences in infection index of C57BL/6 macrophages using up to 200μM H2O2 (Fig 9A). However, in a higher concentration (300 μM) H2O2 was toxic to parasites. Lower concentrations of H2O2 (50μM and 100μM) promoted replication of parasites in Phox KO macrophages, while 200μM H2O2 brought the parasitism back down, and 300 μM H2O2 was also toxic to parasites in Phox KO macrophages. Our next step was to evaluate if the treatment of the parasite with H2O2 could affect T. cruzi capacity to infect mice. We treated blood trypomastigotes of the Y strain of T. cruzi with 100μM of H2O2 for 30 minutes, infected C57BL/6 WT mice by intraperitoneal injection of 103 blood trypomastigotes and followed the course of infection. Our results indicate that C57BL/6 WT mice infected with treated parasites presented significantly higher parasitemia eight days after infection, compared with animals infected with control non treated parasites (Fig 9B). Our results suggest that the lower concentrations of ROS used contribute to growth of the parasite inside the cells, working like signaling molecules to the parasite.
To investigate whether the increase in the intracellular growth rate observed for the EcMutT parasites would also affect the course of infection in vivo, C57BL/6 WT and Phox KO mice were infected with one million TcWT or TcMTH, and parasitemia was evaluated from day 3 post-infection. The data obtained revealed that TcMTH-infected mice presented significantly higher parasitemia compared with animals infected with TcWT parasites and this difference was more prominent at 5 days post-infection (Fig 10A). This result corroborates the result obtained when the infection was performed in Swiss mice [29]. In addition, no difference in parasitemia was found between mouse strains. C57BL/6 WT and Phox KO mice displayed similar parasitemia, which peaked around 5 days post-infection (Fig 10A) and was subsequently controlled. However, mice deficient in functional NADPH oxidase do not survive infection (Fig 10B). While C57BL/6 WT mice presented 100% of survival after day 40 of infection, Phox KO animals exhibited high mortality when compared to C57BL/6 WT, starting at day 9 and reaching 100% mortality by 12 days of infection (Fig 10B and as previously published for the Y strain, [52]).
Oxidative stress, resultant from a deregulated ROS production, has been involved in pathogenesis of several diseases [53,53–55]. On the other hand, in higher eukaryotic cells, reactive oxygen species (ROS) recently emerged as important players in cellular signaling involved in cell growth and differentiation [56]. The regulated increase in free radicals in a temporary imbalance represents the physiological basis for redox regulation [57] and, in this case, ROS can act as secondary messengers in the intracellular signal transduction pathways [56,58,59]. In this paper, we show dual role for reactive oxygen species during infection with T. cruzi.
In agreement with earlier observations in higher eukaryotes [60,61] some authors have pointed evidences for a role of ROS in growth and signaling events of pathogens [11,38,40–42]. In Leishmania, iron uptake controls H2O2 generation, which can act as a signaling molecule, initiating differentiation of promastigotes into infective amastigotes [43]. Sub-lethal doses of the superoxide-generating drug menadione and H2O2 results in increased resistance to H2O2 toxicity and increased virulence of L. chagasi promastigotes [62]. Corroborating these findings, the inhibition of ROS production by treatment with NAC reduced parasite burden in BALB/c mice infected with Leishmania amazonensis [63]. The exposure of T. cruzi to sub-lethal doses of H2O2 caused an increase in the level of antioxidant enzymes, and confers resistance to this oxidant [11,29]. Moreover, some authors have demonstrated different ROS-independent mechanisms used by cells to kill T. cruzi, contesting the necessity of these molecules in killing of parasites [64,65].
However, other studies relate ROS with the killing of parasites. Macrophages treated with phorbol myristate acetate (PMA), which triggers respiratory burst, are incapable of releasing ROS upon subsequent re-stimulation. In these cells, PMA pre-treatment contributes to growth of T. cruzi, pointing to the importance of a respiratory burst mechanism in killing of intracellular parasites [66]. In another work, the ability to release H2O2 and the ability to kill trypanosomes were correlated in macrophages [67] and strong evidence for peroxynitrite as a mediator of T. cruzi killing was also found [4]. Using a variant clone derived from the cloned macrophage cell line J774, which lacked the capacity of producing ROS, Tanaka et al. demonstrated that T. cruzi grew better in this variant cell line [33] and that H2O2 is associated with the killing of parasites [68].
Hence, from the exposed above, ROS may be friend to the parasite or foe.
T. cruzi is exposed to oxidative stress conditions in its life cycle [2,10,69] and this may generate oxidized nucleotides, causing DNA damage. We had already demonstrated that parasites with enhanced 8-oxo-dGTPase activity multiply better than wild type parasites in Swiss mice [29]. In the present work, we show that, although both wild-type and recombinant parasites had the same capacity of invading macrophages, modified parasites presented improved growth in macrophage cultures and confirm that these parasites multiply better in vivo, using C57BL6/WT mice, a different animal model. Hence, protection of DNA against oxidative stress is beneficial to the parasite performance both in vivo and in vitro. The reason for this better performance may be as simple as the quicker replication when there is less necessity of DNA repair, or a more complex mechanism involving increased expression of protective enzymes, as discussed below, and which exact cause is currently unknown.
The hydrolysis of 8-oxo-dGTP could prevent DNA lesions and this could explain the greater replicative capacity of parasites with enhanced 8-oxo-dGTPase activity. Indeed, we demonstrated that modified parasites prevent 8-oxo-dGTP incorporation into DNA when exposed to H2O2. Furthermore, TcMTH and EcMutT parasites expressed more TcCPx and TcMPx after exposure to H2O2 than WT parasites [29] and, importantly, incorporate less peroxynitrite (Fig 3). The enzymes TcCPX and TcMPX have the capacity to detoxify ONOO−, H2O2 and small-chain organic hydroperoxidases [18,21], which would explain the smaller concentrations of peroxynitrite inside TcMTH parasites compared to TcWT. Thus, these latter set of data speak for a more complex reason for higher proliferation of TcMTH parasites, which could involve increased expression of antioxidant enzymes.
T. cruzi contains four iron superoxide dismutases (FeSODs) that eliminate superoxide radicals by dismutation into H2O2 and molecular oxygen [70]. The levels of cytosolic SODB did not increase after oxidative treatment with H2O2, possibly because this enzyme is not involved with H2O2 detoxification. Furthermore, there are no differences in the levels of SODB between modified and wild type parasites. T. cruzi cytosolic FeSODB is particularly resistant to peroxinitrite inactivation, suggesting it participates mainly as an antioxidant defense enzyme, while mitochondrial FeSODA may act as an oxidative stress sensor participating in O2●−-mediated redox process of cell signaling [21,71]. Some works show that levels of TcCPX, TcMPX and mitochondrial SODA are up-regulated in infective forms of the parasite [20,72,73]. The relationship of these enzymes with the infective capacity of the parasite further helps to explain why modified parasites replicate more successfully in vitro and in vivo.
Thus, the results obtained here reinforce the idea that ROS are deleterious to T. cruzi, since modified parasites grow better, probably because they are better able to deal with oxidative stress conditions.
Phox KO macrophages infected with CL Brenner strain of T. cruzi had decreased parasitism compared to C57BL/6 WT macrophages. This difference is not related to the uptake of parasites, since our results demonstrate that both macrophages presented the same parasite uptake. Similarly, knockout mice to p47 subunit of NADPH oxidase (p47Phox KO) were not compromised in parasite uptake capacity [74]. We also show that the treatment with antioxidants reduce parasite replication. These results indicate that the parasite needs a signal provided by macrophages to replicate efficiently within these cells. These data are in agreement with a work published by Paiva et al [35] using Y strain of T. cruzi, but are in contrast with work published by Dhiman and Garg [74] using Sylvio X10/4 strain and p47Phox KO mice. This latter work shows no differences in the number of trypomastigotes released in the supernatants of infected macrophages from p47Phox KO and WT mice [74]. This difference could be related to the type of strain used or to levels of ROS detected in macrophages after parasite infection. Our results show that T. cruzi infection did not induce alterations in ROS levels detected in Phox KO macrophages, which are similar to levels observed in resting cells. The same result was obtained when we used zymosan as a stimulus (S4 Fig). Dhiman and Garg, however, showed that ROS levels detected in p47Phox KO cells are reduced when compared to levels produced by WT cells, but are significantly higher when compared to basal levels of production in non-infected cells [74].
Our results suggest that the parasite needs a signal given by ROS in order to grow inside the cells. It is unlikely that O2●− is the reactive oxygen species responsible for signaling in T. cruzi, because of the anionic nature and restricted capacity of this molecule in to cross membranes. So, the exposure of parasites in the cytosol to this radical would be unlikely. On the other hand, H2O2 is an oxidant with higher diffusional capacity. Although H2O2 is known for its cytotoxic effects, recently it has emerged as an important regulator of signal transduction in eukaryotic cells. This positive (signaling) or negative (damage) effect is dependent of the level of H2O2 and of the cell type under investigation [75]. We show here that parasite growth in Phox KO macrophages could be triggered if we treated Phox KO macrophages or antioxidant treated-WT macrophages with H2O2 before infection. Treatment of parasites with H2O2 before infection also induced the recovery of replicative capacity in Phox KO macrophages. In addition, treatment of blood-derived parasites with H2O2 used for in vivo infection increased parasitemia levels in C57BL/6 WT mice. Although the mechanism by which ROS promotes parasite proliferation remains to be elucidated, our data suggest that this signal is given to the parasite instead the macrophage, since pre-treatment of parasites is sufficient to promote growth in Phox KO macrophages and in vivo.
Further evidence that H2O2 is the signal for parasite replication is that treatment of macrophages with catalase, an enzyme that promotes H2O2 detoxification, also reduced parasitism. This signal could be provide by a direct or an indirect effect of H2O2. Removal of H2O2 by antioxidant enzymes prevents H2O2 signaling, but recent studies have identified several peroxide-signaling mechanisms in which antioxidant enzymes act as H2O2 sensors. The high affinity of some peroxidases for H2O2 makes them suited for hydrogen peroxide-sensing [75]. The peroxidase class of H2O2-scavenging enzymes has conserved cysteine residues in their catalytic sites, which are targets for oxidation by H2O2 [76]. The initial H2O2 sensing event would be the oxidation of an antioxidant enzyme, which then leads to changes in the activity of associated components of the signaling pathway [75]. A recent work shows that H2O2 signaling could be sensed by cysteine-containing proteins, such as thiol peroxidase peroxiredoxin-2 (PRX-2), which would become oxidized and would transmit oxidative equivalents to the redox-regulated transcription factor STAT3. Prx2 catalyzes the formation of disulfide-linked STAT3 oligomers, which compromises its capacity to promote transcription [77]. This could influence the cellular response, activating or inhibiting different cell pathways related with pathogen invasion.
Some studies have suggested that oxidative stress is important for T. cruzi proliferation [35,36]. One evidence for the role of ROS in signaling events is that ROS or heme-induced ROS activate a CaM Kinase II-like pathway muttriggering the proliferation of the epimastigote forms of T. cruzi [36]. In addition, the oxidative stress generated in response to Y strain of T. cruzi contributes to the maintenance of high parasite burdens in macrophages [35]. The treatment with antioxidants inhibited epimastigote proliferation in vitro [36] and reduced T. cruzi parasitemia [78].
Once in the vertebrate host, trypomastigotes invade cells at the inoculation site (e.g., fibroblasts, macrophages, and epithelial cells) [79,80]. T. cruzi multiplies inside resident macrophages and disrupts these cells, which release infective trypomastigote forms that reach blood circulation and disseminates to other cells, like myocardium and autonomic nervous system ganglion cells that innervate esophagus and intestine walls. We found no differences in parasitemia in Phox KO and C57BL/6 WT mice. This could be because ROS are important to signaling events in macrophages, but not in other host cells, like for example fibroblasts [35]. These data are in contrast with the reported increased parasite burden in apocynin-treated mice and in p47Phox KO mice infected with strain Sylvio X10/4 [74,81] and also in contrast with the reported reduced parasitemia in Phox KO mice infected with Y strain [35]. Data with p47Phox KO mice show that these mice succumbed to infection with SylvioX10/4 strain probably because of a compromised CD8+T cell response, leading to increased parasite burden and pathogenesis [74]. Our results on in vivo infection with CL Brenner strain are in agreement with the previously published work by our group with Y strain [52]. In that paper, we showed that infected Phox KO mice succumb to infection probably due to low blood pressure caused by excess ●NO, which was not quenched by superoxide. Apocynin treatment increased parasitemia in C3H/HeN mice infected with the SylvioX10/4 strain. The reason for the discrepancy between data obtained in Phox KO mice in our hands and apocynin-treated mice described earlier [74,81] may be several, including mouse strain and parasite strain. The SylvioX10/4 strain grows more slowly in mice [74,81], and it is not clear if parasitemia in apocynin-treated mice was determined in the acute or in the chronic phase of infection. In addition, it is possible that apocynin inhibits other oxidative mechanisms independent of NOX2. In vitro, apocynin has been shown to have an oxidative effect [82]. In our hands, apocynin effects were consistent with the inhibition of NOX2 in macrophages.
The exact mechanism by which low oxidant production enhances T. cruzi infection remains to be elucidated. One possibility is that ROS could generate the oxidation of 8-oxoG, resulting in the monophosphate form, 8-oxodGMP, which could be acting as a second messenger to the cell, indicating the presence of oxidative stress and preparing the parasite to be more resistant. This is currently under investigation.
In the present study we attempted to clarify the importance of ROS in T. cruzi infections. We found that modified parasites, more resistant to DNA damage by ROS, multiply better inside macrophages, but their proliferation, as well as the proliferation of wild-type parasites, is significantly reduced when ROS production is inhibited in the host cell. A possible explanation is that parasites need minimal levels of ROS, which would work as a signal for replication. However, high levels of ROS are deleterious to the parasite, inducing, for example, DNA damage. In this way, parasites over-expressing 8-oxo-GTPase could be more fit (since less DNA repair is necessary, for instance) and escape from the negative effects induced by ROS, by decreasing double-strand breaks and thus lethal lesions, increasing their replicative capacity.
|
10.1371/journal.ppat.1002376 | EBV Tegument Protein BNRF1 Disrupts DAXX-ATRX to Activate Viral Early Gene Transcription | Productive infection by herpesviruses involve the disabling of host-cell intrinsic defenses by viral encoded tegument proteins. Epstein-Barr Virus (EBV) typically establishes a non-productive, latent infection and it remains unclear how it confronts the host-cell intrinsic defenses that restrict viral gene expression. Here, we show that the EBV major tegument protein BNRF1 targets host-cell intrinsic defense proteins and promotes viral early gene activation. Specifically, we demonstrate that BNRF1 interacts with the host nuclear protein Daxx at PML nuclear bodies (PML-NBs) and disrupts the formation of the Daxx-ATRX chromatin remodeling complex. We mapped the Daxx interaction domain on BNRF1, and show that this domain is important for supporting EBV primary infection. Through reverse transcription PCR and infection assays, we show that BNRF1 supports viral gene expression upon early infection, and that this function is dependent on the Daxx-interaction domain. Lastly, we show that knockdown of Daxx and ATRX induces reactivation of EBV from latently infected lymphoblastoid cell lines (LCLs), suggesting that Daxx and ATRX play a role in the regulation of viral chromatin. Taken together, our data demonstrate an important role of BNRF1 in supporting EBV early infection by interacting with Daxx and ATRX; and suggest that tegument disruption of PML-NB-associated antiviral resistances is a universal requirement for herpesvirus infection in the nucleus.
| Persistent infection by Epstein-Barr virus (EBV) is associated with a variety of diseases, including lymphoid and epithelial tumors. Despite a wealth of information on the mechanism of viral persistence, relatively little is known about the early steps of EBV infection and viral gene activation. Host cells actively mount resistances against viral infection, which viruses need to overcome to invade the cell. We have found that among the proteins packaged in the EBV viral particle, BNRF1 plays an important role of counteracting cellular defenses. We show that EBV protein BNRF1 binds to the cellular protein Daxx and disassembles the Daxx-ATRX complex, where both Daxx and ATRX are cellular proteins known to inhibit viral gene expression. We also confirm that BNRF1 can promote expression of early viral genes, and that Daxx-binding by BNRF1 is required for this function. Finally, we demonstrate that Daxx and ATRX repress viral gene expression during latency. We conclude that BNRF1 disassembles cellular antiviral defense machinery to promote expression of viral genes in the host cell.
| Epstein-Barr virus (EBV) is a member of the human gammaherpesvirus subfamily that infects over 90% of the global adult population [1], [2]. EBV preferentially establishes latent infection in B-lymphocytes but can also infect epithelial cells [3], [4]. EBV primary infection is one of the main causes of infectious mononucleosis (IM); while EBV latent infection is associated with multiple malignancies such as nasopharyngeal carcinoma, Burkitt's lymphoma, and Hodgkin's lymphoma [3], [4]. Furthermore, EBV is responsible for the majority of lymphoproliferative diseases associated with AIDS and immunosuppression following organ transplant [5]. Like all herpesviruses, EBV exists in a dynamic balance between productive and latent infection. The factors that regulate the fate decisions for lytic reactivation from latency have been investigated in some detail, but relatively little is known about the fate regulation during the earliest stages of primary infection.
Upon entry into the nuclear compartment, herpesvirus DNA genomes must confront several intrinsic anti-viral resistances that restrict viral gene expression and replication. One prominent nuclear structure involved in antiviral resistances is the PML nuclear body (PML-NB), also referred to as nuclear domain 10 (ND10). PML-NBs are nucleoplasmic protein aggregates mainly consisting of (but not limited to) the components PML, Sp100, Daxx, and ATRX [6], [7]. The size and abundance of PML-NB is interferon inducible [8], [9], [10], and over-expression of the PML protein represses viral infection [11]. PML-NB is the nuclear localization site of many DNA viruses, including Herpes Simplex virus (HSV-1), Human Cytomegalovirus (HCMV) and Adenovirus (Ad5) [12], [13]. These viruses then modify the morphology and/or protein composition of PML-NBs shortly after infection [12], [14]. The mechanism of PML-NB-mediated antiviral repression is not clearly determined. PML, Sp100, and Daxx are all associated with transcription repression, and this function may act on viral genomes [15]. Daxx can act as a transcription co-repressor of many cellular transcription factors [16], [17], [18], [19], and forms repressive transcription complexes with histone deacetylases (HDACs) [20], [21] and DNA methyltransferase I (DNMT I) [22], [23]. Daxx has been shown to induce heterochromatin markers on the HCMV genome and repress viral gene expression in a HDAC dependent manor [24], [25]. Daxx also forms a chromatin-remodeling complex with ATRX [26] and both can form a repression complex at heterochromatin [27]. Furthermore, RNA interference (RNAi) studies have shown that knockdown of Daxx or ATRX can result in a higher infection level of HCMV [28], [29], [30] and also relieve the infection defect of mutant HSV deficient in disrupting PML-NB [31].
Herpesviruses confront intrinsic anti-viral resistances immediately upon entering the host cell nucleus, and therefore must counteract these resistances at the earliest possible time points to initiate viral gene expression. Herpesvirus tegument proteins, which are pre-packaged and delivered with the infectious virion, are strategically positioned to counteract the intrinsic anti-viral defenses and support the early steps of infection [32]. Both alpha- and beta- herpesviruses encode tegument proteins that regulate early events during lytic replication, including the disruption of the PML-NBs. HSV-1 immediate early gene ICP0, disrupts PML-NB structure by degrading the core component PML [33], [34], [35] and eliminating SUMO-modified Sp100 [36]; while HCMV tegument protein pp71 displaces ATRX and subsequently degrades Daxx [24], [30]. Both ICP0-deficient HSV-1 and pp71-deficient HCMV mutants are deficient in infection, where viral gene expression is shutdown, resulting in a dormant viral genome [29], [37], [38]. Interestingly, it has been reported that disruption of PML-NB by ICP0 is mediated by de novo synthesized ICP0, instead of tegument delivered ICP0 protein, suggesting that this event is coordinated with early viral gene activation or, perhaps, reactivation from latent infection [35]. We have previously shown that EBV genomes localize to and then disrupt PML-NB during lytic replication; while latent EBV episomes are segregated away from PML-NBs during latency [39]. EBV regulatory proteins, including the lytic cycle immediate early gene Zta (also referred to as BZLF1, ZEBRA, and Z), and latency associated EBNA1 and EBNA-LP, have been implicated in PML-NB interactions [40], [41], [42]. However, it remains unclear if PML-NBs regulate early events associated with viral gene expression upon EBV nuclear entry, and if an EBV tegument protein modulates this intrinsic defense.
The EBV major tegument protein BNRF1 is one of the most abundant tegument proteins in the virion [43] and is essential for the establishment of viral latent infection [44], yet its function is largely unknown. BNRF1 homologues are present in all gammaherpesviruses but absent in the alpha- and beta- herpesvirus subfamilies. All BNRF1 orthologues share regions homologous to the cellular enzymes Phosphoribosylformylglycineamide Amidotransferase (FGARAT) and Aminoimidazole ribonucleotide (AIR) synthetase, ATP-dependent enzymes in the 4th and 5th steps of the purine de novo biosynthesis pathway. However, no enzymatic activity has been found in any BNRF1 orthologues. In a knockout study, transfected BNRF1-deficient EBV genomes can reactivate from latency, produce morphologically normal virions, and the progeny can enter cells with little observed defects [44]. Yet, upon infection of B cells the mutant virus showed a 20-fold lower expression of a viral latency associated gene EBNA2 and failed to induce B cell transformation [44]. This suggests an important role of BNRF1 in supporting early infection. Furthermore, the BNRF1 orthologue encoded by murine herpes virus 68 (MHV68), tegument protein ORF75c, induces PML degradation and is essential for initiation of viral gene expression [45], [46].
Here, we demonstrate that EBV BNRF1 is a novel PML-NB-interacting viral protein, and that this interaction is important for supporting EBV primary infection. We first show that Daxx is a primary cellular interaction partner of BNRF1. BNRF1 co-localizes with Daxx at PML-NB foci while disrupting the Daxx-ATRX complex. Furthermore, we identify a novel Daxx interaction domain on BNRF1. This domain is essential for BNRF1 to interact with Daxx, localize to PML-NB, and displace ATRX from Daxx. We then show that BNRF1 supports EBV primary infection and promotes the expression of viral genes soon after viral genomes enter the cell, and that the Daxx interaction domain contributes to these functions. Lastly, we show that knockdown of either Daxx or ATRX results in disruption of viral latency, suggesting that Daxx and ATRX play a role in the restriction of viral gene expression. Our study suggests that EBV tegument protein BNRF1 disassemble the Daxx-ATRX antiviral resistance complex to enable viral gene expression after cell invasion, and likely regulate the chromatin organization for the establishment of latent infection.
To characterize the biological properties of the EBV major tegument protein BNRF1, we took a proteomic approach to screen for potential cellular interaction partners. BNRF1 was cloned into a 3x FLAG tag expression vector under the control of a CMV promoter. 293T cells were then stably transfected with either FLAG-vector or FLAG-tagged BNRF1. Nuclear extracts from stable cell lines were subject to immunopurification (IP) with a FLAG antibody, and then analyzed by SDS-PAGE (Fig. 1A). Bands unique to the BNRF1 lane (B) were cut out and analyzed by liquid chromatography-tandem mass spectrometry (LC/MS/MS). The major identified species was BNRF1, but substoichiometric proteins enriched in the BNRF1 IP were also identified, including Daxx, nucleophosmin (NPM1), and PARP1 (Fig. 1C). We subsequently confirmed in BNRF1 transiently transfected 293T cells that Daxx co-precipitates with BNRF1 (Fig. 1B) by both FLAG pull-down and the Daxx reverse pull-down, indicating a stable in vivo interaction between BNRF1 and Daxx. Neither PARP1 nor NPM1 interaction with BNRF1 could be validated by subsequent co-IPs (data not shown), we therefore focused our efforts on characterizing the interaction with Daxx.
To further characterize the interaction between BNRF1 and Daxx, we introduced serial deletions on the FLAG-BNRF1 expression plasmid. We first made five deletion constructs of BNRF1, sequentially deleting regions coding for 300 amino acids (Fig. 2A, constructs d1 through d5). We then performed IPs with either control IgG, αFLAG, or αDaxx on lysates of cells transfected with the BNRF1 deletion constructs. Daxx co-precipitated in the FLAG IP for all of the BNRF1 mutants with the exception of the BNRF1 300–600 aa deletion mutant (d2) (Fig. 2B, middle panels). Similarly, all of the FLAG-BNRF1 mutants, with the exception of d2, co-precipitated with Daxx IP (Fig. 2B, right panels). Since d2 was expressed and recovered by FLAG IP to similar levels as other BNRF1 mutants capable of interacting with Daxx, we conclude that a putative Daxx-interaction domain is located in the region between 300-600aa of BNRF1.
We then further made six serial deletions of 60 amino acids in the 300–600 aa region (Fig. 3A, constructs d21 through d26) to narrow down the suspected Daxx-interaction domain to a smaller region. After a subsequent round of IP pull-downs, we found that all BNRF1 deletions, with the exception of d21, were defective in binding Daxx (Fig. 3B), suggesting that the 360–600 aa region of BNRF1 is responsible for interaction with Daxx. To determine if this region was sufficient for interaction with Daxx, we expressed only the 300–600 aa region in the FLAG-expression vector (Fig. 3A, construct DID) and performed IP pull-downs. We found that this region bound Daxx as efficiently as WT-BNRF1, in both FLAG IP and in the reverse IP with anti-Daxx antibody (Fig. 3C). Notably, we failed to find sequence homology of this Daxx interaction domain with any known protein motif, and this domain is also distinct from the FGARAT and AIR synthetase homology regions. These findings suggests that BNRF1 utilizes a previously unknown motif to bind Daxx, and that the Daxx interaction domain (300–600 aa) may contain a complex protein fold sensitive to smaller truncation deletions.
Daxx forms a chromatin remodeling complex with ATRX [26] and ATRX has been implicated in the transcriptional repression of both HSV-1 and HCMV during the early steps of infection [30]. Moreover, both HSV-1 and HCMV utilize viral encoded proteins that disrupt the interaction between Daxx and ATRX [30], [31]. To determine if BNRF1 also disrupted the interaction between Daxx and ATRX, we assayed the effect of WT and mutant BNRF1 proteins on the co-IP of Daxx with ATRX. We observed that WT BNRF1 disrupted the interaction between Daxx and ATRX (Fig. 3C, 2nd panel from top, right). However, deletion mutants d22 and d26, which fail to interact with Daxx, did not disrupt ATRX binding in Daxx IP assays (Fig. 3C, 2nd panel from top, right). Interestingly, the Daxx interaction domain by itself (DID), which binds Daxx efficiently, could only partially disrupt ATRX binding. This suggests that Daxx binding by BNRF1 is necessary, but not sufficient for the disruption of ATRX with Daxx. We also found no evidence that BNRF1 co-IPs with PML (Fig. 3C, 3rd panel from top).
To determine whether any other domains of BNRF1 contribute to the disruption of ATRX from Daxx, we assayed FLAG-BNRF1 IPs for ATRX binding using the set of larger BNRF1 deletions examined in Figure 2 (Fig. 3D). We found that WT BNRF1 did not co-IP with ATRX, although it efficiently pulled down Daxx. The BNRF1 d2 mutant failed to pull down Daxx or ATRX, as expected. In contrast, the BNRF1 d3 and d4 mutants, which disrupts most of the FGARAT and AIR synthetase homology regions, efficiently pulled down both ATRX and Daxx. The d1 and d5 truncations, which lie outside of the FGARAT and AIR synthase homology regions, pulled down only Daxx but not ATRX, suggesting it efficiently disrupted the ATRX-Daxx interaction similar to WT. These data suggest that the FGARAT and AIR synthetase homology regions of BNRF1 may contribute to the disruption of ATRX-Daxx complex.
Daxx is a prominent component of PML nuclear bodies [47], and Daxx localization at these nuclear bodies are disrupted by viral proteins of both HSV-1 and HCMV [24], [34]. Thus, it is important to investigate the sub-cellular location of BNRF1-Daxx interaction, and check if BNRF1 disrupts Daxx localization to the nuclear bodies. For immunofluorescence (IF) microscopy studies, we selected Hep2 carcinoma cell lines because of their larger size and prominent PML nuclear bodies, and their common use in many previous studies with herpesvirus protein interactions with PML-NBs. Hep2 cells were transiently transfected with empty FLAG vector (V) or BNRF1 constructs WT, d26, or DID. Cells were then fixed two days post transfection and subject to IF staining. We found that WT BNRF1 partially co-localized with nuclear foci containing Daxx (Fig. 4A, S1A and Table S1), PML (Fig. 4B, S1B and Table S1), and Sp100 (Fig. S1D), suggesting that BNRF1 interacts with Daxx at the PML nuclear bodies. We also noticed that DID itself is sufficient for localizing to PML nuclear bodies, while the d26 deletion mutant showed a weak dispersed pattern in the cell (Fig. 4A and B, S1 and Table S1). The co-localizations were also confirmed by line scan analysis, where the BNRF1 WT and DID intensity peaks overlap with Daxx and PML peaks (Fig. S2). To ensure that the diffuse pattern of the d26 mutant is not due to deficient protein expression, the same set of transfected cells as used for IF were also assayed by Western blot for total expression levels of BNRF1 proteins (Fig. 4D). We found that d26 protein was expressed at levels similar to that of WT, despite its diffuse staining in IF studies, confirming its protein expression in the cells used in our microscopy study. These findings indicate that the interaction with Daxx is necessary and sufficient for BNRF1 to localize to the nuclear bodies.
To understand the BNRF1 disruption of Daxx-ATRX complex in a sub-cellular spatial context, we also examined ATRX by IF in BNRF1-transfected Hep2 cells (Fig. 4C and S1C). Again, we found ATRX foci co-localizing with WT BNRF1 and DID but not d26, which is also confirmed by line scan analysis (Fig. S2C). However, we also found a substantial reduction in ATRX foci intensity when cells were transfected with WT BNRF1, but no apparent reduction when transfected with d26 or DID-mutants (Fig. 4C). The failure of DID to disperse ATRX is consistent with its only partial disruption of ATRX from Daxx IP (Fig. 3C). Quantification of Daxx (Fig. 4E), PML (Fig. 4F), and ATRX (Fig. 4G) nuclear foci in BNRF1-expressing cells compared to non-expressing cells revealed that BNRF1-expressing cells contain a significantly lower (p<0.0001) average number of ATRX nuclear foci than non-expressing cells (Fig. 4G). In contrast, we found no significant difference in the number of Daxx (Fig. 4E) and PML nuclear foci (Fig. 4F) in BNRF1 transfected cells. Taken together, these results suggest that BNRF1 not only disrupts the Daxx-ATRX complex, but also actively disperses ATRX away from nuclear bodies.
HSV-1 ICP0 and HCMV pp71 each induce the degradation of PML and Daxx proteins respectively, yet we did not observe any evidence of this with BNRF1 in our microscopy studies. To investigate the potential degradation of PML, Daxx and ATRX proteins by BNRF1, we examined the stability of these proteins in BNRF1 stably transfected cells (Fig. 5A). 293T cells stably transfected with control vector (clone C) or WT BNRF1 (stable transfection clones 3 and 9) were lysed and subject to Western blot analysis. We found no evidence of degradation or gross post-translational modification of PML, Daxx, nor ATRX in BNRF1-expressing cell lines. We also analyzed the protein stability of PML and Daxx in Hep2 cells transiently transfected with control vector, WT BNRF1 or d26, and again found no evidence of BNRF1-induced protein degradation (Fig. S3). This suggests that BNRF1 does not mimic the protein degradation function of HCMV pp71 or HSV-1 ICP0, but rather, disrupts Daxx-ATRX interactions through alternative mechanisms.
The diffuse distribution of the BNRF1-d26 mutant raised the question of whether the Daxx interaction domain of BNRF1 correlated with nuclear localization. To test this, we utilized biochemical fractionation methods to isolate nuclear and cytoplasmic proteins from 293T cells (Fig. 5B). We found that WT BNRF1 localized to both cytoplasmic (∼60%) and nuclear (40%) fractions. The d26 mutant, which is deficient in both Daxx interaction and nuclear bodies localization, was expressed at lower amounts yet showed a cytoplasmic to nuclear distribution similar to WT (Fig. 5B). This is consistent with d26 having a weak diffuse nuclear and cytoplasmic staining in IF (Fig. 4). Meanwhile, the DID mutant, which binds Daxx and co-localizes with PML nuclear bodies in the nucleus, was isolated at low, yet detectable levels in the nucleus; although substantially more was recovered in the cytoplasmic fraction. The efficiency of the fractionation was confirmed by the presence of PARP1 exclusively in the nuclear fractions, and α-tubulin exclusively in the cytoplasm. These findings suggest that BNRF1 can localize to both cytoplasmic and nuclear compartments, and that the Daxx interaction domain might contribute partially to the nuclear entry or stability of BNRF1.
A previous study using an EBV bacmid with a BNRF1-knockout demonstrated that BNRF1-mutant virions can be generated from producer 293 cells and can enter the cytosol of infected B-cells; yet mutant virus failed to express one of the first expressed latent genes, EBNA2, upon primary infection of B cells, and were incapable of inducing B-cell proliferation [44]. To understand the role of BNRF1-Daxx interaction in primary infection, we took a complementation rescue approach with the BNRF1-mutant virus. 293 cells stably transfected with either wild type or BNRF1-knockout EBV bacmids were used for virus production (Fig. 6). As the EBV bacmids also encode GFP, cells infected with this bacmid-derived virus could be visualized by the presence of green fluorescence. To induce viral production, bacmid containing cells were co-transfected with the EBV transactivator Zta and BALF4. To complement for BNRF1 deletion, production cells were also transfected with either control vector, WT BNRF1 (WT), or the BNRF1 deletion mutant (d26) which fails to interact with Daxx. Three days after transfection, the media was collected and used to infect primary B cells isolated from human peripheral blood mononuclear cells (PBMCs). We detected high-levels of GFP positive proliferating B-cell clusters when infected with virus generated from wild type bacmid (Fig. 6Ai), but no GFP positive or clumped cells were detected when infected with no virus (Fig. 6Aii) or virus from un-complemented ΔBNRF1 bacmids (Fig. 6Aiii). However, when ΔBNRF1 virus was complimented with WT BNRF1 we were able to detect GFP positive cells and proliferating B-cell clusters (Fig. 6Aiv). Notably, ΔBNRF1 virus complimented with the d26 mutant BNRF1 failed to express GFP or induce B cell proliferation (Fig. 6Av), showing a similar defect as ΔBNRF1 virus with no complementation. Quantification of at least three independent infections confirmed that GFP positive and proliferating B-cells were detectable only when ΔBNRF1 bacmid virus was complemented with WT, but not with d26 mutant BNRF1 (Fig. 6B). To ensure the infections between each complemented virus were comparable, virus titer was quantified by real time PCR for virion DNA. The viral titers of either empty vector or WT BNRF1 complemented virus was found to be similar, while some reduction in virus titer was observed with d26 virus (Fig. 6C). We also tested by Western blotting for incorporation of FLAG-BNRF1 proteins in virions, and found that WT and d26 mutant BNRF1 proteins were both packaged into virions to similar per particle levels (Fig. 6D). These findings confirm that BNRF1 is required for primary infection of B-cells, and suggests that the Daxx interaction domain of BNRF1 is important for this function.
Other herpesvirus tegument proteins that interact with Daxx and ATRX have been shown to function in the transcription activation of viral genes during primary infection [25]. To investigate the role of BNRF1 on viral gene transcription early after primary infection, we infected human B-lymphocytes purified from PBMCs with the ΔBNRF1 virus complemented with empty FLAG-vector, WT BNRF1 or d26 mutant BNRF1 (Fig. 7A). Viral gene expression in these newly infected cells was assayed at four days post infection using Reverse Transcription qPCR (RT-qPCR). We found that WT BNRF1 complementation induced an up-regulation of EBNA1, EBNA2 and BZLF1 mRNA expression compared with non-complemented virus or the d26-mutant complementation. Interestingly, background levels of BZLF1 expression were detectable in non-complemented and d26 mutant infections, suggesting that BNRF1 may only partly enhance BZLF1 expression, which can occur at low levels independently of BNRF1.
To investigate the potential mechanism of BNRF1 in viral gene regulation, we first tested the effect of BNRF1 on reporter plasmids using transient transfection assays, but found no consistent effect on candidate viral promoters (data not shown). We reasoned that reporter plasmids may lack essential BNRF1 target elements or chromatin assembly, and therefore assayed BNRF1 activity on EBV bacmid genomes after transfection into 293 cells (Fig. 7B). EBV bacmid DNA (Bac36) and either empty FLAG-vector, WT BNRF1, or the d26 mutant BNRF1 were co-transfected into 293 cells and assayed 3 days post transfection for viral gene expression using RT-qPCR. We found that WT BNRF1 promoted a robust expression of BZLF1 transcripts (∼20 fold), which was not observed in vector control or the d26 mutant (Fig. 7B). BNRF1 also increased EBNA2 mRNA (∼3 fold) relative to vector control, but this was not significantly increased relative to that of the d26 mutant. These studies suggest that BNRF1 can activate the expression of the EBV immediate early gene BZLF1 in the context of the viral genome, and in the absence of other virion-delivered tegument proteins.
The previous experiments suggest that BNRF1 can function during tegument delivery in early infection, as well as after de novo synthesis, perhaps regulating the transition from latent to lytic infection. To explore the role of Daxx and ATRX in the context of EBV latent to lytic gene regulation, we test the effects of Daxx and ATRX knockdown on viral lytic gene expression in Mutu I cells, an EBV-latently infected Burkitt's lymphoma cell line (Fig. 8). Mutu I cells were transduced with puromycin resistant lentivirus carrying either non-targeting shRNA (shNeg), shRNA against Daxx (shDaxx), ATRX (shATRX), or ZEB1 (shZEB1.1) which acts as a positive control for reactivation. ZEB1 has been shown to repress Zta expression, and shRNA depletion of ZEB1 can reactivate lytic gene expression in several cell types [48], [49], [50]. Mutu I cells were harvested 9 days after shRNA transduction and selection, and then tested for viral reactivation by Western blot and FACS (Fig 8). Western blot analysis of whole cell lysates (Fig. 8A and B) revealed that knockdown of either Daxx or ATRX induced a reactivation of EBV early antigens, as shown by increased band intensities of both the immediate early gene Zta (2-fold) and the lytic early antigen EA-D (3-fold). These induction levels are comparable to that observed with the shZEB1.1 positive control. The efficiency of shRNA-mediated knockdown was confirmed by the loss of Daxx, ATRX and ZEB1 bands in the corresponding lanes (Fig. 8A). We also verified reactivation by flow cytometry quantification of the EBV viral capsid antigen VCA on cells from three independent shRNA-treatments (Fig. 8C), where we observed an approximately 6-to-10-fold induction by either Daxx or ATRX depletion. These findings indicate that the depletion of either Daxx or ATRX can promote viral lytic gene expression from latently infected B-cells, and suggest that BNRF1 disruption of the Daxx-ATRX complex contributes to viral gene control during early infection and reactivation.
The specific class of antiviral defense dubbed the intrinsic immunity [51], [52] plays a broad and general role in restricting viral infection. PML-NBs and its associated proteins such as PML, Sp100, Daxx and ATRX, have been extensively studied as cellular defenses against herpesviruses, specifically with the alphaherpesvirus HSV-1 and betaherpesvirus HCMV. Upon the early stages of infection right after cell entry, HSV-1 and HCMV utilize viral proteins that effectively disrupt the structure and disable the function of the PML-NBs in restricting viral gene expression and replication. However, the gammaherpesvirus EBV has been relatively less studied in terms of how it counteracts these cellular resistances upon primary infection or reactivation. We show here that the major tegument protein of EBV, BNRF1, interacts with Daxx (Figs. 1–3, S1–2) and disrupts its ability to form a complex with ATRX or recruit ATRX to PML-NBs (Figs. 3–4, S1–2). Moreover, we show that BNRF1 functionally promotes viral early gene expression with a preference for the activation of the immediate early gene BZLF1, and to a lesser extent the latent activator EBNA2 (Figs. 6–7). These findings indicate that EBV, like its relatives HSV1 and HCMV, encodes a viral tegument protein that targets PML-NB components to promote viral gene expression.
Daxx is a prominent PML-NB component, but is also associated with a diverse, yet non-mutually exclusive variety of cellular functions, including the regulation of apoptosis, chromatin remodeling, gene repression, and antiviral resistance [47], [53]. Daxx is a primary target of the HCMV pp71 protein, which both binds and induces the degradation of Daxx [30]. Like HCMV pp71, BNRF1 binds Daxx and prevents the Daxx-interaction partner, ATRX, from associating with Daxx and localizing to PML-NBs. BNRF1 and pp71 are both tegument proteins, whose pre-made nature likely provides them with a temporal advantage to disarm cellular repression machinery without the prior need of viral gene transcription. However, unlike HCMV pp71, BNRF1 does not induce Daxx degradation, which remains prominently associated with PML-NBs when BNRF1 is expressed (Figs 2–5). BNRF1 and pp71 share no obvious amino acid sequence similarity, and the Daxx interaction domains of these two proteins vary significantly in amino acid composition and size of the interaction domains. These findings suggest that BNRF1 is a functional homologue of pp71, but utilizes a distinct mechanism for the dissociation of ATRX from PML-NBs.
Herpesvirus tegument proteins have been implicated in the determination of viral lytic or latent gene expression programs. Restriction of tegument protein entry into the nucleus, as has been shown for HCMV pp71 and HSV VP16, correlates with the establishment of latency [24], suggesting that tegument proteins may play a critical role in determining lytic or latent gene expression programs. Interestingly, we found that the Daxx-interaction deficient BNRF1 mutant d26, which fails to interact with Daxx, showed a weak diffuse subcellular distribution instead of the punctate nuclear dots of WT BNRF1 (Fig. 4 and S1). Similarly, biochemical fractionation studies (Fig. 5B) suggest that while BNRF1 can localize to both the cytoplasm and nucleus, it may require the Daxx-interaction domain to efficiently accumulate in the nucleus. Potentially related is the observation that HCMV pp71 translocation to PML-NBs is also dependent on its interaction with Daxx [54]. Selective cytoplasmic retention of several herpesvirus tegument proteins, including pp71 and VP16, may play a critical role in determining lytic or latent gene expression programs [55]. We suspect that BNRF1 might be subject to similar regulation through its PML-NB localization.
Daxx and ATRX are known to play a global role in the control of cellular and viral gene expression and chromosomal structure. Daxx itself has been shown to associate with HDACs and to function as a global repressor of transcription [20], [21]. The Daxx-ATRX complex has in vitro chromatin remodeling activities [26] and can function as a histone H3.3 chaperone [56]. Recent studies suggest also that ATRX interacts with G-rich repeat chromatin regions [57], and in collaboration with Daxx load histone variant H3.3 onto pericentromeric and telomeric chromatin [56], [58], [59]. H3.3 is generally associated with open chromatin and active transcription when loaded by the histone chaperone HIRA [60]. However, the Daxx-ATRX complex loaded H3.3 has been found to facilitate transcription from pericentromeric regions [58] but repress transcription from telomeric regions [59]. Interestingly, HIRA-loaded H3.3 can facilitate the lytic replication of HSV-1 during the early steps of infection [61]. Furthermore, the Daxx-degrading pp71 blocks the establishment of heterochromatin on the HCMV Major Immediate Early Promoter (MIEP) region [25]. These findings underscore the importance of host chromatin regulatory mechanisms in the control of herpesvirus infection. We suspect that the viral gene activation function of BNRF1 (Fig. 7) is likely to be mediated by chromatin-dependent processes since we failed to observe consistent transcription activation when assayed in transient plasmid-based reporter assays using EBV promoters for BZLF1 (Zp) or EBNA2 (Cp or Wp) (data not shown). We propose that BNRF1 stimulates EBV early gene activity through de-repression of the Daxx-ATRX mediated chromatin repression mechanism, perhaps similar to that of pp71 de-repression of the HCMV MIEP locus. However, the precise molecular mechanism through which BNRF1 activates early gene transcription through the disruption of ATRX-Daxx interaction remains to be investigated.
While not explored yet, it is also not known if the FGARAT enzyme-homology domain of BNRF1 has any function in the context of supporting viral infection. This enzyme homology is conserved among all gammaherpesvirus orthologues of BNRF1, including the KSHV and MHV68 ORF75 family members. Despite significant sequence similarity with BNRF1, KSHV and MHV68 ORF75 proteins do not appear to interact with Daxx (data not shown). However, MHV68 ORF75c targets PML-NBs through the degradation of PML [45], [46], an activity that we did not observe with BNRF1. Thus, while these tegument family members share the FGARAT homology regions, and may similarly target components of the PML-NBs, they appear to target different proteins and utilize distinct mechanisms. It is also important to note that the disruption of ATRX by BNRF1 was partially dependent on the FGARAT domain, since the DID alone, which binds Daxx efficiently, only partially disrupt ATRX binding in IP assays (Figs 2 and 3) while not causing any significant ATRX dispersion from PML-NBs in IF assays (Fig. 4). Also, deletions within the FGARAT domain (d3 and d4) resulted in a mutant BNRF1 that co-precipitated with ATRX, creating a gain of function not seen with WT BNRF1. All of this suggests that the FGARAT domain may play a regulatory role in BNRF1 interactions with Daxx and ATRX.
In conclusion, our data demonstrates a novel example of herpesvirus tegument protein interacting with components of the cellular antiviral resistance. BNRF1 interaction with Daxx may provide several functions, including the establishment of a chromatin structure conducive to viral early gene activity. Our findings demonstrate that EBV, like other herpesviruses, confront the PML-NB associated intrinsic defenses through a viral factor that is available and active upon the early stages of infection, and shed light into the critical control mechanisms that govern the early events of EBV infection before the establishment of latency.
Human B-lymphocytes were obtained from the Wistar Institute phlebotomy lab. All samples were from anonymous adult donors and approved by the Wistar Institute Institutional Review Board. Written informed consent was provided by study participants.
Hep2 and 293T cells were grown in Dulbecco's modified Eagle medium (DMEM) supplemented with 10% fetal bovine serum (FBS), 20 mM GlutaMAX (Gibco), 100 U/ml penicillin and 100 µl/ml streptomycin. 293HEK cells were grown in minimum essential medium Eagle (MEM), supplemented with 10% FBS and 20 mM GlutaMAX (Gibco). DG75 and Mutu I cells are EBV negative and positive (respectively) Burkitt's lymphoma cell lines, grown in RPMI 1640 medium supplemented with 10% FBS, 100 U/ml penicillin and 100 µl/ml streptomycin. Peripheral blood mononuclear cells (PBMCs) were isolated from fresh donated human blood by density gradient centrifugation with Ficoll-Paque Plus purchased from GE healthcare. Primary B cells were then isolated from PBMCs using Dynabeads Untouched Human B Cell isolation kit (Invitrogen). All cells were grown in a 5% CO2 incubator at 37°C. Stable 293 cell lines expressing FLAG-BNRF1 (clone 3 and clone 9) and empty FLAG vector (clone C) were grown in DMEM as described for 293T cells above, supplemented with 2.5 µg/ml Puromycin for selection.
Viruses were produced using chloramphenicol and hygromycin resistant bacmids containing the EBV genome and the gene coding for green fluorescence protein (GFP). 293/EBV-wt cells (a gift from H. J. Delecluse) are 293HEK cells stably transfected with the wild type EBV bacmid [62]. 293/△BNRF1 cells (a gift from H. J. Delecluse) are 293HEK cells stably transfected with an EBV bacmid with the BNRF1 gene deleted [44]. 293/EBV-wt and 293/△BNRF1 cells were grown in RPMI 1640 medium supplemented with 10% FBS and 100 µg/ml hygromycin.
All restriction enzymes, T4 DNA ligase and associated buffers were purchased from New England Biolabs. Monoclonal mouse anti-FLAG antibody (F1804), Polyclonal rabbit anti-FLAG anibody (F7425), Polyclonal rabbit anti-Daxx antibody (F7810), Monoclonal Anti-α-Tubulin antibody (T5168), Monoclonal mouse anti-β-Actin-Peroxidase antibody (A3854), and Anti-mouse IgG R-Phycoerythrin (PE) conjugated antibody (P8547) were purchased from Sigma-Aldrich. Monoclonal mice anti-PML (PG-M3, sc-966), and polyclonal rabbit anti-ATRX (H-300, sc-15408), and polyclonal rabbit anti-ZEB1 (sc25388) were purchased from Santa Cruz Biotechnology. Polyclonal rabbit anti-PARP1 antibody (ALX-210-895-R100) was purchased from Enzo Life Sciences. Mouse anti-EA-D antibody was purchased from Millipore. Anti-EBV-VCA (0231) antibody was purchased from Pierce Thermo Scientific.
BNRF1 was cloned into the HindIII-SalI sites of the p3xFLAG-Myc-CMV-24 Expression Vector (Sigma-Aldrich), using the PCR primers: gcgaagcttgaagagaggggcagggaaacgcaa and gcggtcgactcactcggaggggcgaccgtgcctg. BNRF1 deletion mutants were generated as follows. PCR Primers (Table S2) were designed so that the front and rear halves of the DNA oligo each binds the 5′ or 3′ regions flanking the targeted deletion site on the BNRF1 template. PCR reactions were setup using iProof High-Fidelity DNA polymerase 2x master mix (Bio-Rad), with primers at 1 µM concentration, and the FLAG-BNRF1 expression plasmid as the template at a concentration of 50 ng DNA in a 25 µl reaction setup. PCR was done with a Bio-Rad C1000 thermal cycler, thermal cycles setup according to DNA polymerase mix manufacturer suggested conditions. To clear out the wild type BNRF1 template, 15 µl of the PCR product were treated with 30 U DpnI (New England Biolabs) in a 20 µl reaction for 2 hours to over night at 37°C. 2 µl of DpnI-treated DNA were then transformed into 50 µl of Library Efficiency DH5α competent cells (Invitrogen). Colonies were screened for the deletion by enzyme digestion analysis of miniprep DNA, and then confirmed by DNA sequencing of the expected deletion site.
BNRF1 expression plasmids were transfected using Lipofectamine 2000 (Invitrogen) according to manufacturer instructions. Cells were harvested 2 days post transfection by washing cells off the plate with PBS. Harvested cells were further washed 3 times with cold PBS, and then subject to lysis with freshly prepared NET lysis buffer (50 mM Tris-HCl pH 7.5, 150 mM NaCl, 5 mM EDTA, 0.5% NP-40, and 0.1% mammalian protease inhibitor cocktail mix (P8340, Sigma-Aldrich), at 1 ml NET per IP pull-down. Cell lysates were homogenized by doing 10 strokes in a Dounce homogenizer. 60 µl of each lysate were isolated after this step as input control. The remaining lysates were incubated at 4°C rotating for 30 mins to fully solubilize proteins. Lysates were then spun at 13000 rpm 5 mins to remove insoluble cell debris, then antibodies were added (5 µl of each antibody per IP) to the cleared lysates, and left rotating over night. 100 µl of 50% slurry of Protein A sepharose beads (GE healthcare) in NET buffer was added to each IP with rotating at 4°C for 2–3 hours, then washed three times with NET for 10 mins (rotating at 4°C) per wash. Pulled down proteins were released by adding 50 µl 2x Laemmli buffer (100 mM Tris-Cl pH 6.8, 4% SDS, 0.2% Bromophenol Blue, 20% Glycerol), and boiling for 10 mins at 100°C. The resulting samples (excluding beads) were then loaded directly into protein gels and subject to Western blot analysis.
For mass spectrometry identification of BNRF1 associated proteins, FLAG-BNRF1 expressing and FLAG-vector control stable cell lines were generated as mentioned above. Nuclear extracts from 5×107 cells were subject to immunopurification with anti-FLAG Sepharose beads (A2220, Sigma-Aldrich) followed extensive washing with NET buffer, and FLAG peptide elution. Eluted protein was subject to precipitation with 10% trichloroacetic acid (TCA) followed by SDS-PAGE and colloidal blue staining. Sections of the gel with enriched polypeptides were subject to LC/MS/MS at the Wistar Proteomics Facility.
Hep2 cells were transfected with BNRF1 expression plasmids using Lipofectamine 2000 (Invitrogen) according to the manufacturer's instructions, transfected cells were then reseeded at 2.7×104 cells/well in 24 well plates containing microscope coverslips 5 hours post transfection. 2 days post transfection, coverslips with cells attached were harvested, fixed with 1% paraformaldehyde at room temperature for 15 mins, then permeablized with 0.3% Triton-X 100. Coverslips were then stained with the first antibodies over night at 4°C. First antibody dilutions used were as follows: mouse anti-FLAG at 1∶20000, rabbit anti-Daxx at 1∶5000, rabbit anti-FLAG at 1∶5000, mouse anti-PML at 1∶250, rabbit anti-ATRX at 1∶250, all antibodies diluted in PBS. Second antibody stainings were carried out for 1 hour at room temperature with the red-fluorescent Alexafluor594 goat anti-rabbit antibody and green Alexafluor488 goat anti-mouse antibody (both from Invitrogen) each at 1/800 dilution in PBS. Coverslips were washed twice in PBS for 5 mins between each of the above treatments. Cell nuclei were stained briefly with DAPI (diluted to a final concentration of 0.167 µg/ml in PBS) for 2 mins, then washed with PBS, 70% EtOH, then 100% EtOH to wash out residual salts. Coverslips were air-dried briefly, and then mounted onto microscope slides with Vectasheld mounting media (Vector Laboratories). Mounted slides were examined under a Nikon E600 upright microscope with a 100x oil objective. Photos for nuclear body quantification were took using a 40x objective to maximize the number of cells in each photo while retaining a clear view of PML bodies.
Microscopy photos were analyzed using ImagePro Plus 6.2 software (Media Cybernetics). Photos were pre-processed by subtracting out the background intensity using the operation function (with a value of −30), and passing through a flatten filter (a value of 10). A morphological ‘top hat’ filter was then applied to emphasize points or grains brighter then the background. The number of nuclear bodies in each cell nucleus was counted by quantifying the object numbers after applying the signal intensity threshold/segmentation tool to select the nuclear bodies as objects. Cell boundaries were defined by the outline from DAPI channel photos of the same field, while omitting all cells on the border of the image border. Resulting quantification numbers were then analyzed using Prism 4 software (Graph Pad Software), statistical analysis did by Mann-Whitney U non-parametric, unpaired t test.
293T cells were transfected in 10 cm plates with 2 µg expression plasmids of either empty FLAG vector, WT-BNRF1, BNRF1-DID, or 6 µg of BNRF1-d26. Transfection was carried out using 10 µl Lipofectamine 2000 (Invitrogen) per transfection, following manufacturer instructions. Cells were harvested 24 hrs post transfection. 1/6 of cells isolated as input control. The rest of the cell pellets were fractionated with the Fermentas ProteoJET Cytoplasmic and Nuclear Protein Extraction Kit (K0311). The resulting cytoplasmic and nuclear fractions, along with the input samples, were analyzed by Western blot.
To induce lytic virus production, 293/EBV-wt and 293/△BNRF1 cells were transfected in 10 cm plates with expression plasmids of 1.75 µg BALF4, 3.25 µg BZLF1 or cDNA3 empty vector, and 3 µg of either empty FLAG vector or 3 µg BNRF1 or 7.5 µg BNRF1-d26. Transfection was carried out using 15 µl Lipofectamine 2000 (Invitrogen) per transfection, following manufacturer instructions. The media of virus production cells were harvested 3 days post transfection, filtered through 0.45 µm filters, and added directly to freshly isolated primary B cells. B cells in virus containing media were centrifuged for 1200 rpm 90 mins at 25°C to enhance infection.
For measuring infection by GFP levels, infected B cells were treated with 1 mM Sodium Butyrate and 20 ng/ml TPA 3 days post transfection to enhance GFP expression, and the number of GFP positive cells in each well were counted manually under a Nikon TE2000 microscope using a 20x objective.
For measuring virus gene expression in infected B cells, cells were collected 4 days post transfection, and total RNA was purified using Trizol (Invitrogen). The resulting RNA was then subject to DNase 1 treatment at 2 U/50 µl, 1 hour at 37°C, then DNase was heat inactivated by adding a final concentration of 5 mM EDTA and incubated at 70°C for 10 mins. cDNA was synthesized using the Super Script III first strand synthesis system reverse-transcription kit (Invitrogen). The resulting cDNA was then subject to real time PCR analysis by ΔCt method and normalized to viral titers, measured as described below. Real time PCR primers used are listed in Table S3.
To measure the amount of complemented BNRF1 protein that were packaged into virions, △BNRF1 virions complemented with WT-BNRF1, BNRF1-d26, or un-complemented, were produced as mentioned above. 100 µl of the harvested and filtered virus-containing media were isolated for viral titer measurement as described below. The rest of the virus-containing media were concentrated by loading the media above a 5 ml layer of 22.5% sucrose in PBS, then centrifuged at 27000 rpm (∼100,000 g) 4°C for 1 hour in a SureSpin 630 Rotor (Thermo Scientific) with a Sorvall WX 100 Ultra ultracentrifuge. The resulting virus pellet was then resuspended in PBS, and analyzed by Western blot. Protein gel loading volumes were normalized according to viral titers to ensure equal amounts of virion protein in each well.
Viral DNA in media was extracted as described by C. Busse et al. [63]. Virus-containing media were treated with 5 U/50 µl of DNase I (New England Biolabs) for 1 hour at 37°C. DNase was then deactivated by adding EDTA to a final concentration of 5 mM, followed by 10 mins heat inactivation at 70°C. Samples were then mixed 1∶1 with 0.1 mg/ml of proteinase K in water, and incubated at 50°C for 1 hour, followed by 20 mins of heat inactivation at 75°C. The released viral DNA was measured by real time PCR analysis, using a serial dilution series of Namalwa cell lysate as the standard curve, which contain two copies of integrated EBV genome per Namalwa cell. EBV genomes were detected using primers specific to the OriLyt region: 5′- CGTCTTACTGCCCAGCCTACTC-3′ (OriLyt-fwd), 5′- AGTGGGAGGGCAGGAAATG-3′ (OriLyt-rev).
Wild type EBV genome bacmids were prepared from 2.5 mls overnight LB culture using the Bacmax DNA purification kit (Epicentre). 293HEK cells were seeded (2.3 million cells per plate) the previous day in 10 cm plates, and transfected with 1.5 µg freshly prepared bacmids along with 0.5 µg of either empty FLAG vector, BNRF1, or BNRF1-d26 mutant. Transfection was carried out using Effectene transfection reagents (Qiagen), following manufacturer instructions. Cells were harvested three days post transfection, total RNA was purified using Trizol (Invitrogen), and then subject to DNase 1 treatment at 2 U/50 µl, 1 hour at 37°C. DNase was heat inactivated by adding a final concentration of 5 mM EDTA and incubated at 70°C for 10 mins. cDNA was then synthesized using the Super Script III first strand synthesis system reverse-transcription kit (Invitrogen). The resulting cDNA was then subject to real time PCR analysis by ΔCt method. Real time PCR primers used are listed in Table S3.
shNeg (pLKO-shNeg), shDaxx (pLKO-shDaxx-2) and shATRX (pLKO-shATRX90) constructs in lentivirus production plasmid backbones were generous gifts from Roger Everett. shNeg (sequence TTATCGCGCATATCACGCG) was designed to poorly target the E.coli DNA polymerase and extensively screened to ensure that it does not affect human nor viral transcripts. Use of shDaxx and shATRX was previously described else where [30], [31]. shZEB1.1 was obtained from the TRC library (Sigma, Inc), with targeting sequence GCAACAATACAAGAGGTTAAACTCGAGTTTAACCTCTTGTATTGTTGC). Mutu I cells were infected with lentiviruses carrying pLKO.1-puro vectors by spin-infection at 400 g for 45 minutes at room temperature. The pellets were resuspended in fresh medium and left growing overnight. The RPMI medium was replaced each day, with 2.5 ug/ml Puromycin added for selection for lentivirus transduced cells. The cells were collected after 9 days of puromycin selection, and subject to Flow cytometry quantification of EBV viral capsid antigen positive cells, and Western blot analysis.
|
10.1371/journal.pgen.1004844 | A Drosophila ABC Transporter Regulates Lifespan | MRP4 (multidrug resistance-associated protein 4) is a member of the MRP/ABCC subfamily of ATP-binding cassette (ABC) transporters that are essential for many cellular processes requiring the transport of substrates across cell membranes. Although MRP4 has been implicated as a detoxification protein by transport of structurally diverse endogenous and xenobiotic compounds, including antivirus and anticancer drugs, that usually induce oxidative stress in cells, its in vivo biological function remains unknown. In this study, we investigate the biological functions of a Drosophila homolog of human MRP4, dMRP4. We show that dMRP4 expression is elevated in response to oxidative stress (paraquat, hydrogen peroxide and hyperoxia) in Drosophila. Flies lacking dMRP4 have a shortened lifespan under both oxidative and normal conditions. Overexpression of dMRP4, on the other hand, is sufficient to increase oxidative stress resistance and extend lifespan. By genetic manipulations, we demonstrate that dMRP4 is required for JNK (c-Jun NH2-terminal kinase) activation during paraquat challenge and for basal transcription of some JNK target genes under normal condition. We show that impaired JNK signaling is an important cause for major defects associated with dMRP4 mutations, suggesting that dMRP4 regulates lifespan by modulating the expression of a set of genes related to both oxidative resistance and aging, at least in part, through JNK signaling.
| The drug transporters are often known for their ability to transport different physiological-related compounds across cell membranes. Although the abnormal up-regulation of some these transporters is believed to be the common cause of the clinic problem called drug resistance, the biological functions of these transporters remain largely unknown. Here we show that a Drosophila homolog of the mammalian drug transporter plays a role in lifespan regulation. Mutations of this gene increase the sensitivity to oxidative stress and reduce lifespan, while overexpression of this gene increases resistance to oxidative stress and extends lifespan. By molecular and genetic analyses, we have linked functions of this gene to a key signaling transduction pathway that has been known to be important in lifespan regulation.
| In Drosophila, one important feature of the aging process appears to be the similarity between the changes in gene expression that occur during aging and oxidative stress response [1], [2], [3]. For instance, the up-regulation of genes encoding for some chaperones and/or detoxification agents in response to oxidative stress has been found to highly correlate with the aging process [1], [2], [3]. Hsp proteins may promote longevity by facilitating the clearance of damaged proteins that accumulate during aging [4]. Another example is the JNK signaling pathway which can be triggered by a variety of insults, including oxidative stress, and has been shown to be a genetic determinant of aging in Drosophila [5]. Mutations in the JNK cascade increase stress sensitivity and lead to shortened lifespan. Conversely, flies with increased JNK activity can sustain oxidative stress and live longer [6]. Although genome-wide surveys [1] are powerful and have linked a set of genes between stress response and aging, the majority of them have not been tested experimentally for lifespan; some genes involved in both processes may still be missing by genome-wide surveys. Here we report that a new gene, namely dMRP4, which has not been reported on the survey list [1], clearly plays a role in both aging process and oxidative stress.
The multidrug resistance-associated protein 4 (MRP4) belongs to the subfamily C (also known as ABCC) of the ATP-binding cassette (ABC) transporter protein family. It has been classified as a detoxification protein that is implicated in transport of structurally diverse endogenous and xenobiotic compounds, including antivirus and anticancer drugs that usually induce oxidative stress in cells and lead to toxicity [7], [8], [9]. MRP4 mRNA and protein are widely expressed in many tissues of mammals including humans [10], suggesting that this transporter may be involved in different physiological processes. However, several recent studies have shown that mammalian MRP4 is not essential for development, since MRP4-knockout mice are viable and do not reveal any abnormalities [11], [12], [13], [14]. Therefore, the biological function of MRP4 remains largely unknown.
MRP-associated drug resistance has represented an important clinical problem in the treatment of cancers. Some cancer cells seem to adopt a survival strategy to protect against chemotherapy-induced oxidative stress by increasing transport of chemotherapeutics out of cells, as a result of induction of MRP, including MRP4 [15], [16], [17], [18], [19]. Indeed, up-regulation of MRP4 expression has been linked to a variety of human cancers [20], [21], [22], [23], [24]. The induction of hepatic MRP4 by oxidative stress has also been observed in mammalian liver injury after chemical treatments and this response appears to be regulated primarily at a transcriptional level [25], [26]. However, oxidative stress-inducing agents do not always induce MRP4 [27], [28], [29], [30], raising the possibility that the induction of MRP4 expression during oxidative stress may be agent-dependent and/or cell type-specific. Furthermore, no study has attempted to address whether MRP4 is required for general oxidative stress resistance at a whole organismal level.
We have previously identified the Drosophila homolog of mammalian MRP4, called dMRP4, during an unbiased screen for genes whose overexpression causes an abnormal response to hypoxia in adult flies [31]. dMRP4 encodes a protein sharing 43% overall amino acid identity and 63% similarity with the human MRP4 [32], [33]. In this study, we have investigated the possible involvement of dMRP4 in resistance to oxidative stress. By genetic manipulation, we present evidence that dMRP4 is associated with changes in lifespan under both oxidative stress and normal conditions, likely through a mechanism that is linked to JNK signaling in Drosophila.
To test our hypothesis that the expression of dMRP4 may be regulated by oxidative stress in Drosophila, we first analyzed dMRP4 transcriptional activity in response to oxidative stimuli by feeding flies with paraquat, which generates superoxide in mitochondria [34] and has been widely used as an oxidative stress inducer in vivo. The expression of dMRP4 was strongly induced in wild-type flies fed with 10 mM paraquat for 12 hours (Fig. 1A). Similar induction patterns were observed in parallel with two known oxidative stress-responsive genes [3], [6], [35], puc (puckered) and gstD1 (glutathione s transferase D1). To test whether dMRP4 responds to other oxidative stressors, we analyzed its transcriptional changes in flies treated with hydrogen peroxide as well as hyperoxia. Up-regualtion of dMRP4 was clearly observed after hydrogen peroxide or hyperoxia treatment, in parallel with two known up-regulated markers, gstD1 and hsp22, under these conditions [1] (Fig. 1B–C). These results indicate that Drosophila dMRP4 is a bona fide oxidative stress-responsive gene.
To test whether dMRP4 indeed might play a role in oxidative stress resistance, we generated two mutations by excision of two independent EP elements near the dMRP4 gene (Fig. 2A). Analysis of the dMRP4 expression by RT-PCR indicated that dMRP4 RNA was undetectable in these mutants (Fig. 2B). However, the more sensitive assay with qt-PCR revealed about 8% dMRP4 mRNA retaining in both homozygous mutations (Fig. 2C). Currently it is not clear if this transcript residual was resulted from splice forms of the predicted full length mRNA or from an alternative transcription start site of the remaining dMRP4 transcript after the truncation. Nevertheless, these results indicate that the two dMRP4 alleles represent strong loss-of-function mutations. In addition, flies homozygous for both mutations were viable and fertile, suggesting that dMRP4 may not be an essential gene for development. However, it cannot be ruled out that the remaining residual in these mutations might still retain some vital function during development.
To address whether induction of dMRP4 is required for defense against oxidative stress, we monitored the survival of adult flies treated with three most commonly used oxidative stressors: paraquat, hydrogen peroxide, or hyperoxia. In each condition the two dMRP4 alleles or their transheterozygous combination displayed similar and reproducible phenotypes: flies lacking dMRP4 reduced profoundly their viability under oxidative stress relative to controls (Fig. 2D–F, Log-rank test, p<0.001). These results demonstrate that wild-type dMRP4 is required for oxidative stress resistance in Drosophila.
Oxidative stress is known to activate a protective program involving induction of a number of stress-responsive genes in cells [3], [6], [35], [36]. JNK signaling is activated in response to oxidative stress and is a major genetic factor in control of oxidative stress tolerance and aging process [3], [6], [35], [36], [37]. Since puc (a phosphatase inhibitor of JNK) is often used as a marker for activation of the JNK pathway [3], [6], [35], [36], [38], we tested whether there were any differential expression changes of JNK signaling by examining puc induction in dMRP4 mutant flies fed with paraquat. Compared to the pattern in wild-type flies, puc expression was completely diminished in dMRP4 mutant flies under oxidative stress (Fig. 3A). To further evaluate whether dMRP4 might play a general role in JNK signaling, induction of other JNK-mediated marker genes, such as gstD1 [6], hsp68 and Jafrac1, was also examined. Although expression of all these marker genes was induced in wild-type flies after paraquat feeding, their induction, with exception for gstD1, was significantly reduced in the dMRP4 mutant flies (Fig. 3C–D), indicating that activation of JNK signaling by oxidative stress requires a wild-type dMRP4 function. Because flies deficient for JNK signaling become more susceptible to stress [6], a phenotype resembling what we have observed with flies deficient for dMRP4, impairment of JNK signaling in dMRP4 mutants may be an important cause for increased lethality when animals face oxidative insults. There was also a possibility that dMRP4 itself may be a component of the JNK pathway.
To test whether dMRP4 might be a component of the JNK pathway, we examined dMRP4 response in flies with reduced activities of JNK signaling by the expression of a dominant negative form of Bsk (BskDN) (Basket, a Drosophila homolog of JNK). BskDN can mimic bsk mutant phenotypes in flies and cells [39]. In this experiment, BskDN expression was induced in adult flies by actin-GeneSwitch-Gal4 (actGS-Gal4), a RU486-mediated system [40] that drives ubiquitous expression in whole fly. In the presence of drug RU486, BskDN expression was activated from the UAS driven transgene. The relative mRNA levels from RU486-fed flies were compared to control flies carrying the same induction system (actGS>dMRP4) without drug feeding. Inhibition of JNK activity by BskDN, as shown by puc expression, did not repress dMRP4 induction in response to paraquat (Fig. 3E), indicating that JNK signaling is not required for dMRP4 induction under this stress. Next we asked whether stimulation of JNK signaling might influence dMRP4 induction. This was achieved by conditionally expressing an activated version of Hep (HepAct) (hemipterus, a Drosophila homolog of JNKK). HepAct has been shown to be a JNK gain-of-function mutant [39]. Constitutive activation of JNK signaling by HepAct did not change dMRP4 expression in paraquat-fed flies relative to controls (Fig. 3F). These results indicate that unlike those direct targets of JNK, dMRP4 induction by paraquat is independent of JNK activity, and therefore dMRP4 is not a direct component, but instead acts in parallel on a signaling that perhaps only regulates expression of some downstream effectors, of the JNK pathway.
If dMRP4 is essential for oxidative resistance in Drosophila, an increased dMRP4 expression may increase oxidative resistance in wild-type flies. To test this hypothesis, we used the RU486-system to test the role of dMRP4 overexpressing in paraquat resistance. Adult flies carrying tub5GS>dMRP4, after being fed with RU486 for dMRP4 induction (Fig. 4I), significantly improved survival rates following acute treatment with paraquat (30 mM) compared to control flies (Fig. 4A). Importantly, RU486 feeding itself had no effect on survival under the same condition (Fig. 4B). These experiments underline the protective role of dMRP4 from paraquat challenge. It also implies that this protection does not need dMRP4 to be elevated before reaching adulthood.
Because mammalian MRP4 has been implicated in protecting the liver from oxidative stress [25], [26], we sought to investigate whether it was also the case in Drosophila. Drosophila fat body is an analogous tissue to mammalian liver and white adipose tissue [41], [42]. yolk-Gal4 is expressed specifically in the female fat body [43]. We tested whether overexpression of dMRP4 in the fat body could provide overall protection against oxidative damage to the whole fly. Induction of dMRP4 in female fat body by yolk-Gal4 led 4-fold increase in the dMRP4 transcript (Fig. 4H) and rendered flies much more tolerant to paraquat treatment as compared to controls (yolk-Gal4/+ or dMRP4/+) (Fig. 4C, Log-rank test, p<0.01). Similarly, overexpression of dMRP4 by S106-Gal4, an inducible driver expressed predominantly in adult fat body [40], [44], [45], significantly increased survival of paraquat-fed flies in the presence of RU486 (Fig. 4D). Again, RU486 treatment showed dose-dependent induction of dMRP4 expression (Fig. 4J) but played no role in mortality under the same condition (Fig. 4E). Thus, the Drosophila fat body appears to be an important tissue for dMRP4 to sustain paraquat-induced oxidative stress. Furthermore, the protective role of dMRP4 under paraquat challenge is applicable for both sexes.
The anti-oxidative effect of dMRP4 on lifespan was further tested by exposing flies to hyperoxia. Flies overexpressing dMRP4 by RU484 induction clearly lived longer under 90% oxygen environment compared to controls (Fig. 4F, Log-rank test, p<0.001). We conclude that wild-type dMRP4 function is to promote resistance to oxidative stress in Drosophila.
Aging shares many features with oxidative stress [1]. The free radical theory has proposed a link between aging and oxidative stress [46], [47]. Recent studies from genetic manipulation of many genes in Drosophila have presented evidence that resistance to oxidative stress genetic often correlate with increased lifespan [6], [48], [49], [50]. Since manipulation of dMRP4 can influence lifespan under oxidative stress, it would be important to examine whether dMRP4 regulates lifespan under non-stress conditions. We observed that mutations in dMRP4 dramatically caused a shortened normal adult lifespan (Fig. 5A, Log-rank test, p<0.0001). In particular, dMRP4M2/M2 flies had a mean lifespan (as measured by 50% survival) of 45 days and a maximum lifespan (as measured by the 90 percent survival) of 60 days. Compared to wild-type controls, dMRP4M2/M2 flies had a major reduction in the mean lifespan of about 47% and a decrease in maximum lifespan of 24% (Fig. 5A). Similar results were observed with dMRP4M1/M1 flies (Fig. 5A). The overall mortality rates of these groups were compared using Partial Slopes Rank-Sum Test [51] over the linear portion of the increase in mortality. Despite an apparent initiation of early mortality before day 30 in survival of dMRP4 mutants, there was no significant difference in slopes between the mutants and wild type (Fig. 5B), indicating that loss of dMRP4 decreased lifespan by lowing the whole mortality trajectory, but not the rate of increase in mortality with age. Thus, although dMRP4 is not required for normal development, it is required for normal lifespan under non-stress conditions.
Since flies overexpressing dMRP4 were more resistant to oxidative stress, we tested whether overexpressing dMRP4 would be sufficient to extend lifespan. RU486-mediated overexpression was used to minimize the influence of genetic background on lifespan assays. RU486-fed tub5GS>dMRP4 flies lived significantly longer than their siblings without RU486 feeding (Fig. 5C, Log-rank test, p<0.0001). The lifespan extension by tub5GS>dMRP4 expression appeared to be correlated with the dose of RU 486. In one case, the mean lifespan was extended to 16% and the maximum lifespan to 8% (Fig. 5C, RU486 100 ug/ml). In the other case, when flies were fed with 20 ug/ml RU486, this group of flies showed only about 9% of increase in the mean lifespan and 5% of increase in the maximum lifespan, even though their overall lifespan appeared to significantly increase (Fig. 5C, Log-rank test, p<0.0001). Increased lifespan was not due to chronic RU486 treatment because no significant difference in lifespan was seen between treated or untreated tub5GS-Gal4 groups (Fig. 5D, Log-rank test, p = 0.3). We conclude that another dMRP4 function is to promote normal lifespan in Drosophila.
In these experiments the lifespan extension clearly correlated with increased expression of dMRP4, but it remained unclear whether tissue-specific dMRP4 overexpression was sufficient to extend lifespan and whether the overall levels and/or timing of such expression would be critical. Interestingly, S106>dMRP4 flies treated with RU486 did not live longer (Fig. 5E, Log-rank test, p = 0.37) even though the fat body-specific expression of dMRP4 did show resistance to paraquat, suggesting that there might be different requirements between resistance to oxidative stress and lifespan extension. Again, RU486 treatment showed no difference between parallel controls (Fig. 5F, Log-rank test, p = 0.09). Moreover, high levels of ubiquitous dMRP4 expression by da-Gal4 throughout development were not beneficial and instead, there was a negative correlation with lifespan (Fig. 5G, Log-rank test, p<0.0001). These observations suggest that in order for dMRP4 overexpression to be beneficial for lifespan extension, the spatial and temporal such expression with proper levels have to be tightly controlled.
In order to learn the molecular mechanism by which dMRP4 regulates lifespan, we selectively studied transcription profiling of several genes whose expression changes have been linked to both aging and stress [1]. Among five hsp (heat shock protein) genes examined, expression of three genes, hsp68, hsp70 and l(2)efl (lethal (2) essential for life, a small hsp gene) was severely down-regulated in dMRP4 mutant flies (Fig. 6A), while they were significantly up-regulated when dMRP4 was overexpressed (Fig. 6B). Overexpression of dMRP4 was also sufficient to increase expression of other two hsp genes, hsp22 and hsp83 (Fig. 6B). Since l(2)efl is a known target of dFOXO (Drosophila forkhead transcription factor) in lifespan regulation [52], it raised the possibility that dMRP4 might regulate expression of other dFOXO-dependent genes. Indeed, expression of the dFOXO target gene thor, which encodes 4E-BP (eIF4E binding protein), was also greatly enhanced when dMRP4 was overexpressed. Since both thor and hsp68 are target genes of JNK signaling [6], [52], we further examined expression patterns of several other JNK targets (Fig. 3A–D). Like hsp68, basal expression of puc and gstD1 was down-regulated in dMRP4 mutant flies and was up-regulated with dMRP4 overexpression (Fig. 6A–B). Furthermore, basal expression of Jafrac1 was increased when dMRP4 was overexpressed, even though its expression was not affected by dMRP4 mutation under normal condition. Thus, in addition to regulating the JNK-dependent gene expression under oxidative stress, dMRP4 also regulates the basal transcription of such genes under normal conditions.
Increased expression of hsp22 [53], hsp68 [6], [54], hsp70 [55], l(2)efl [52], Jafrac1 [54], [56], has been reported to increase Drosophila lifespan. We hence suggest that increased expression of these genes by elevated dMRP4 expression may account for, at least in part, the dMRP4-mediated lifespan extension.
Increasing age is accompanied with physiological decline. The locomotor decline is one of prominent physiological changes as they grow older. The climbing ability, measured by negative geotaxis, of adult fly reflects a function of age in Drosophila [57], [58]. To determine whether the onset of aging associated with dMRP4, we performed a negative geotaxis test for flies with different ages. Although there was no difference in negative geotaxis behavior between 5-days old dMRP4 mutant and wild-type adults, the age-associated functional decline became visible in dMRP4 mutant flies already at day 10 of adulthood, at a time when no mortality was seen regardless of mutant or wild-type controls (Fig. 7C). By age 40 days, although there was a progressive functional decline in the control group, it was clearly worse in dMRP4 mutant groups (w1118 vs dMRP4M2/M1, Fig. 7C). Thus, the functional decline as they aged was faster in dMRP4 mutants than in controls.
Activation of JNK signaling can increase stress resistance and extend lifespan in both Drosophila [6], [52], [59], and C.elegans [60]. Our observations (Fig. 3A–D, Fig. 6A–B) suggest that the deficiency in basal transcription and stress response of JNK signaling may be an important cause for loss of stress tolerance and normal lifespan with dMRP4 mutant flies. If this were the case, increasing JNK signaling might be expected to correct dMRP4 deficiency. We tested this hypothesis by recombination of a pucE69 chromosome into the dMRP4 mutant background. pucE69/+ flies were more resistant to paraquat and lived longer under normal conditions [6] (Fig. 6A and B). When dMRP4 mutant flies also heterozygous for pucE69 were challenged with paraquat, they behaved like pucE69/+ flies alone: they lived significantly longer not only than dMRP4 mutant flies, but also longer than wild-type controls (Fig. 7A, p<0.01). Consistent with a previous report [6], pucE69/+ flies extended normal lifespan (27% mean lifespan and 24% maximum lifespan) of control flies (dMRP4M2/+) under non-stress conditions (Fig. 7B, p<0.0001). More strikingly, the puc, dMRP4 double mutant flies remarkably extended the mutant mean lifespan by 61% (dMRP4M2/M1 vs pucE69/+, dMRP4M2/M1) and maximum lifespan by 42% (Fig. 7B, p<0.0001). These results demonstrate that dMRP4 deficiency in stress resistance and lifespan regulation is correlated with a defect in JNK signaling. These results also place puc genetically in epistatic interaction with dMRP4 in both stress resistance and lifespan regulation.
We tested whether the functional decline with age might also be associated with JNK activity by comparing the climbing ability between wild-type and pucE69/+ flies. Increased JNK signaling did not appear to benefit wild-type flies before 30 days of age, as climbing tests did not reveal a significant difference in locomotor function between wild-type and pucE69/+ flies (Fig. 7C). However, after 40 days of age, increased JNK activity indeed improved climbing ability, and therefore functional aging in wild-type flies (w1118 vs pucE69/+ in the 40 d group, Fig. 7C), suggesting that JNK activity is required for fitness of older flies. We then tested whether the age-associated functional decline of dMRP4 mutants could be caused by impaired JNK signaling as well. The climbing ability of puc, dMRP4 double mutant flies was restored to the level comparable to that of wild-type flies in the first 30 days of age. Therefore, early functional decline of dMRP4 mutants is possibly associated with a decline of JNK signaling (Fig. 7C). Furthermore, by age of 40 days, puc, dMRP4 double mutant flies behaved like pucE69/+ flies, showing better climbing performance even over wild-type flies (Fig. 7C). Thus, the JNK activity can seemingly rescue all defects that are associated with dMRP4 phenotypes. We conclude that dMRP4 plays a critical role in regulation of JNK-mediated oxidative resistance and aging process.
The MRP4 subfamily and its homologs have not been reported in any lifespan-related studies including genome-wide surveys. In this study, we have investigated the physiological function of dMRP4 gene in Drosophila. A main finding from our work is that dMRP4 regulates lifespan under both normal conditions and oxidative stress, concomitantly with changes of JNK activity in vivo. Our main finding is based on several observations: First, dMRP4 is required for induction of some JNK-dependent genes in response to paraquat-induced oxidative stress. Second, elevated dMRP4 expression stimulates basal transcription of some JNK-dependent genes downstream of JNK signaling. Third, increased JNK activity in dMRP4 mutant background can rescue dMRP4-related phenotypes identified in this work, supporting our hypothesis that dMRP4 may regulate oxidative resistance and lifespan, at least in part, through JNK signaling.
The finding that dMRP4 has a role in lifespan is particularly intriguing because we are able to show for the first time that a drug transporter like MRP4 is involved in lifespan regulation. Like Drosophila dMRP4, MRP4 KO mice show no visible phenotype [11], [12], [13], [14], and mrp-4 knockdown in C.elegans with RNAi results in no observed phenotype either [61], [62]. These observations together suggest that MRP4 and its homologs across species do not contribute to normal development in the animal world. However, unlike in other species, we found that the Drosophila dMRP4 is required for adult lifespan. Flies deficient for dMRP4 live significantly shorter, under both stressful and normal conditions. Subsequently, our work reveals that dMRP4 acts as a modulator of a network of gene expression since loss- or gain-of dMRP4 function leads to major changes in the transcriptional profiling of a number of genes that may contribute to lifespan regulation. Therefore we suggest that gene expression changes mediated by dMRP4 may represent a molecular mechanism by which dMRP4 regulates lifespan. For instance, hsp genes have been implicated in regulation of both stress resistance and lifespan extension [4], [63], and are among the best-known biomarkers of aging in C.elegans [63], [64], in Drosophila [1], [65], and perhaps even in humans [66]. Given the fact that the expression of hsp reporters in young individual flies has been observed to be partially predictive of remaining lifespan [65], down-regulation of several hsp gene expression (i.e. hsp68, hsp70, l(2)efl) in dMRP4 mutant background could explain the shorter lifespan of these flies, while their up-regulation (i.e. hsp22, hsp68, hsp70, hsp83, l(2)efl) at a young age by dMRP4 overexpression may help protect against oxidative stress and extend lifespan of wild-type flies. This scenario is consistent with previous notions that genes are involved in stress responses generally share similar involvement with aging [1].
In addition to hsp genes, the interaction of dMRP4 with JNK signaling may provide an alternative mechanism to explain dMRP4 functions. Because the JNK pathway is known to be crucial in stress resistance and aging, impairment of JNK signaling in dMRP4 mutant flies, indicated by transcriptional down-regulation of several known JNK-related effecters, could result in dMRP4-associated phenotypes. The acute phenotype is seen particularly when the animal faces stressors such as paraquat-induced oxidative stress, which recapitulates the phenotype shown by mutations in the JNK pathway [6]. The effect of the JNK pathway on lifespan has also been observed during aging under normal conditions. Flies with reduced JNK activity have a shorter lifespan [6], a phenotype similar to that seen in dMRP4 mutant flies. Furthermore, some downstream effectors in the JNK pathway also exhibit phenotypes that are reminiscent of dMRP4. For instance, loss of Jafrac1 function leads to an exaggerated sensitivity to paraquat-induced oxidative stress and a shortened lifespan, while overexpression of Jafrac1 increases oxidative resistance and extends lifespan [56]. Interestingly, expression of Jafrac1 transcription is down-regulated in the dMRP4 mutant in response to oxidative stress (Fig. 2D) and is up-regulated by dMRP4 overexpression (Fig. 6B). How dMRP4 regulates Jafrac1 remains to be investigated. One possible scenario is that dMRP4 executes its functions through interacting with JNK signaling to modulate the expression of downstream effectors such as Jafrac1 especially that the expression of Jafrac1 itself is regulated by JNK signaling [56]. After all, the most compelling evidence for the relationship between dMRP4 and JNK signaling comes from our genetic epistatic assays. When JNK signaling is enhanced in dMRP4 mutant background, all dMRP4-related defects are restored, and puc, dMRP4 double mutant flies now phenocopy pucE69/+ flies, clearly proving that JNK signaling plays a central role in realizing dMRP4 functions. Our work also suggests that promoting lifespan by increasing JNK signaling may be a result of its ability to antagonize oxidation on macromolecules, thereby postponing aging. Compared to JNK signaling, the effect of increased dMRP4 expression on lifespan extension seems less dramatic. Yet this phenotype, together with the results showing that loss- or gain-of JNK function does not alter dMRP4 expression, indicates that dMRP4 functions as a modulator of, but not a component within, JNK signaling. Furthermore, if dMRP4 is one of upstream modulators of JNK/Puc signaling, it is conceivable that its overexpression cannot entirely recapitulate the effect of JNK/Puc activation and consequently, it may not be as effective as a direct manipulation of JNK/Puc signaling with respect to lifespan.
Together our results, we propose a working model to summarize how dMRP4 executes its functions in conjunction with JNK signaling (Fig. 7D). Future work needs to explore how a transmembrane protein such as dMRP4 could integrate its signal into the JNK pathway under both stress and normal conditions.
Although in human and mammalian models of cholestasis, MRP4 has been implicated in providing protection against oxidative stress, the genetic basis for this resistance has not yet been addressed. Therefore, the connection between tissue oxidative stress, survival of the animal, and the physiological function of MRP4, has been lacking. In this work we show that overexpression of dMRP4 in Drosophila fat body, the equivalent tissue of mammalian liver and white adipose tissue, can confer oxidative resistance to the whole animal, suggesting a functional importance of dMRP4 in the fat body in the protection of Drosophila against oxidative stress. Drosophila fat body has recently been reported as a primary site of lipid oxidative damage after paraquat treatment [67]. dFOXO, whose expression is predominately restricted to the fat body, appears to regulate sensitivity of paraquat-induced oxidative damage and age-associated degeneration of behavioral rhythms through this tissue [68]. Furthermore, overexpression of dFOXO in the adult fat body can increase stress resistance and retard aging process [44], [69], supporting the physiological role of fat body in stress defense for the whole organism. Strikingly, we show in this work that expression of two targets of dFOXO, l()efl and thor, are greatly induced when dMRP4 is overexpressed, raising the intriguing possibility that dMRP4 may promote stress resistance and lifespan extension by activation of dFOXO, for instance through JNK signaling [52]. However, unlike the finding that global induction of dMRP4 can promote lifespan, we have not observed a significant lifespan extension when dMRP4 overexpression is restricted in fat body. This observation suggests that the ability of stress resistance may not be an absolute factor associated with longevity in a particular tissue. It is also possible that in order for dMRP4 to benefit for longer life, more tissues with its elevated expression need to be involved. Our studies in fact have not ruled out the roles of dMRP4 in tissues other than the fat body to survival even under oxidative stress.
The main function of MRP4 family is known for their ability to transport a variety of diverse endogenous and xenobiotic compounds. An interesting speculation could be raised as to whether dMRP4 might function simply as a transport in paraquat resistance. In this scenario, flies deficient in dMRP4 might not be able to efficiently exclude paraquat out of cells, thereby leading to substrate-related toxic effects. However, this assumption would hardly explain why flies deficient in dMRP4 lose their resistance to hydrogen peroxide and hyperoxia. In addition, there is no report for paraquat as a potential substrate of any MRP4 members thus far.
The deteriorate influence by da>dMRP4 overexpression is notable because this phenotype has not been seen in overexpression studies of mammalian MRP4. Although use of the whole animal in this study clearly differs from use of cultured cells in mammalian researches, it is more likely that high levels of dMRP4 expression may interfere with normal development, resulting in a pleotropic impact on later assays. An early report did observed that overexpression of two EP lines, which all targeted dMRP4, in larvae caused neuromuscular phenotypes [70].
Given the considerable conservation of pathways between Drosophila and mammals, it will be interesting to test if manipulating MRP4 in mammalian liver cells could confer resistance to the liver, or even to the whole animal subjected to chemotherapy-induced oxidative stress. Finally, our proposed mechanism that interactions between dMRP4 and JNK signaling may shed new light on the clinic problems for long-lived cancer cells with drug resistance due to elevated expression of MRP including MRP4 proteins.
EP3177 and EP3655 were described previously [31]. Other stocks: w1118, w; TM3,Sb,Ser/TM6B,Tb, w; Sco/CyO; MKRS/TM6B,Tb, daughterless (da)-Gal4, S106-Gal4, pucE69, UAS-BskDN and UAS-HepAct strains were obtained from Bloomington stock center. These strains have been backcrossed to w1118 for 8–10 times before experiments. yolk-Gal4 [43] was kindly provided by Norbert Perrimon and was backcrossed into w1118 background for 8 times. Actin-GeneSwitch-Gal4 (actGS-Gal4, [71], [72]) was a gift from Dirk Bohmann. tublin5-GeneSwitch-Gal4 (tub5GS-Gal4, [73]) was a gift from Scott Pletcher. These Gal4 strains have been backcrossed into w1118 background for 6 times before use. Flies were raised on standard Drosophila food (per liter: 17.3 g of yeast, 73.1 g of cornmeal, 10 g of soy flour, 77 ml of light corn syrup, 4.8 ml of propionic acid, and 5.7 g of agar).
To generate dMRP4 mutant flies, two independent EP lines, EP3177 and EP3655 were first backcrossed into w1118 background for 8 times. EP males were crossing to w1118; Δ2-3 Sb/TM3 females that provides with transposase. Males with mosaic color eyes were excised and subsequently balanced with w1118; TM3,Sb,Ser/TM6B,Tb strain. The balanced excisions were then repeatedly backcrossed via the balancer strain for 8 times to establish excision stocks. They were identified by loss of the expression of the mini-white gene. The genomic deletions were determined by sequencing with specific primers spanning the EP insertion region. Two deletions obtained had truncated the 5′-end of putative dMRP4 transcript, which was designated as dMRP4 mutation 1 (w1118; dMRP4M1) and dMRP4 mutation 2 (w1118; dMRP4M2) (Fig. 2A). dMRP4M1 was excised from EP3655, which inserted at 47 bp from the transcription start site of the predicted gene CG14709, resulting a 2.7 kb deletion that removed 1179 bp upstream of dMRP4 transcript and a 1521 bp region including 585 bp of the entire exon 1 encoding the first 25 amino acids of the protein, as well as 936 bp of the intron 1. dMRP4M2 was resulted from an excision of EP3177, which inserted at 88 bp from the transcription start site of the predicted gene CG14709. This led to a 3 kb deletion that has removed 2117 bp upstream of dMRP4 transcript and an 883 bp region spanning the entire exon 1 and part of intron 1.
The pucE69, dMRPM2 recombination strain was generated by recombination of pucE69 and dMRPM2 onto the same 3rd chromosome. Both the balanced pucE69 and dMRPM2 were repeatedly backcrossed via w1118; TM3,Sb,Ser/TM6B,Tb for 8 times before the recombination experiments. The presence of both mutations after meiotic recombination was verified by genetic cross and by PCR with specific primers. Resultant pucE69/+, dMRPM2/M2 double mutants were then continuously backcrossed via w1118; TM3,Sb,Ser/TM6B,Tb for more than 10 times and were kept with the balancer as a parent stock.
To induce dMRP4 overexpression, adult flies carrying different Gal4 drives were crossed to homozygous EP3177 lines. For RU486 induction, a 25 mg/ml RU486 (mifepristone, Sigma) stock solution made in 100% ethanol was diluted with water for desired concentrations. 250 ul of diluted RU486 solution was added onto the surface of standard fly food. This “on food” method has been shown to be simple and effective over other RU486 supply methods [74]. The vials were allowed to dry for 24 hours before use. The same solution without RU486 was added to fly food for control experiments.
In most experiments, three to four day-old males, grouped with 20 flies per vial, were fed on a 3 mm Whatmann paper soaked with 10 mM paraquat (N,N′-dimethyl-4,4′-bipyridinium dichloride, Sigma) in 5% sucrose/PBS. Flies of different genotypes were also fed only with 5% sucrose/PBS as experimental controls. Under this condition all flies can live up for 10 days perfectly. Scores were done every 12 hours for the number of dead flies. Fresh paraquat was added daily. All tests were performed at 25°C. Flies were not starved before adding paraquat in this test to avoid unnecessary stress. Survival comparisons were analyzed by Kaplan–Meier Log-rank Test using Graph Pad Prism4. p<0.05 was considered statistically significant.
In RU486-induced experiments, 20 adult males (2–4 days old) per vial were fed with different concentrations of RU486 for 4–6 days. They were then transferred on a 3 mm Whatmann paper soaked with 30 mM paraquat in 5% sucrose for acute survival test, or with 10 mM paraquat in 5% sucrose for mRNA induction at 24 hours. Control flies were from the same collection and were treated in parallel.
For RNA, all samples were collected at the end of treatments and were immediately frozen in dry ice for RNA preparations.
Eight day-old males were fed with different concentrations of hydrogen peroxide (v/v, Sigma) in 5% sucrose/PBS. Control flies were fed with 5% sucrose/PBS only. RNA for qt-PCR was extracted from these flies after 24 hours treatment. For survival tests, ten day-old males with different genotypes were fed with 3% hydrogen peroxide. Fresh hydrogen peroxide was added every day. Scoring and analysis were done essentially as described in paraquat treatment.
Eight day-old males, grouped with 20 flies per vial on regular food, were exposed to a steady flow of 95% or 90% oxygen bubbled through water in a sealed chamber. RNA for qt-PCR was extracted from these flies after indicated time points. For survival tests, twelve day-old males with different genotypes were treated with 90% oxygen as above. For RU486 induction, flies from the same breeding were divided into two groups, one group fed on food containing RU486 (150 ug/ml) and the other on normal food through the experiments. Flies were transferred to fresh vials every 2–3 days. Scoring was done every day.
Flies were collected within 24 hours of eclosion and grouped into 20 males per vial. Tests were performed at 25°C. For each experiment, at least 200 flies of each genotype were tested.
For GeneSwitch experiments, males of genotypes w1118; tub5GS-Gal4, w1118; actGS-Gal4, or w1118; S106-Gal4 were crossed to w1118; EP3177 or w1118 females, respectively. Male progeny from these crosses were aged for 3 days after eclosion, and then were divided into 20 flies per vial, with or without indicated concentrations of RU486 in food. Flies were transferred to fresh vials with or without RU486 every other day and dead flies were scored at the time of transfer.
All experiments were conducted at least two times from independent biological breeding. The maximum lifespan was the mean lifespan of last 10% of survival animals in each cohort.
10–20 male flies, ages from 5–40 days at 25°C, were transferred to a clean plastic vial, rested for 3 min, and then measured for bang-induced vertical climbing distance at room temperature (20–21°C). The performance was scored as percentage of flies crossing 7 cm within 10 seconds in a single vial, which was expressed as average of 5 repeated tests for a single vial. 80–100 flies were tested for each genotype at each time point.
Total RNA was isolated from whole flies using RNeasy Mini Kit (Qiagen, Maryland, USA) according to the manufacturer's instructions. cDNA synthesis was performed with oligo-dT and random primers using SuperScript III first-strand synthesis system (Invitrogen, Carlsbad, CA). Semiquantitative PCR was performed as described [31]. Real-time PCR was performed in duplicate using SYBR Green on an ABI 7900HT Real-Time PCR system (Applied Biosystems) according to the manufacture's protocol. All samples were analyzed from at least 3 independent of experiments. Data was normalized first to the level of the rp49 mRNA prior to quantifying the relative levels of mRNA between controls and experimentally treated samples. All detailed primers are available upon request.
All survival data were analyzed by Kaplan–Meier Log-rank Test for overall survival and by the Student's t-test for mean and maximum lifespan using Graph Pad Prism4. The log mortality was determined by OASIS program [51]. Treated data were then plotted using Graph Pad Prism4. Other comparisons were determined either by Student's t-test or One way ANOVA followed by post hoc t-test. p<0.05 was considered statistically significant.
|
10.1371/journal.pgen.1003457 | Reference-Free Population Genomics from Next-Generation Transcriptome Data and the Vertebrate–Invertebrate Gap | In animals, the population genomic literature is dominated by two taxa, namely mammals and drosophilids, in which fully sequenced, well-annotated genomes have been available for years. Data from other metazoan phyla are scarce, probably because the vast majority of living species still lack a closely related reference genome. Here we achieve de novo, reference-free population genomic analysis from wild samples in five non-model animal species, based on next-generation sequencing transcriptome data. We introduce a pipe-line for cDNA assembly, read mapping, SNP/genotype calling, and data cleaning, with specific focus on the issue of hidden paralogy detection. In two species for which a reference genome is available, similar results were obtained whether the reference was used or not, demonstrating the robustness of our de novo inferences. The population genomic profile of a hare, a turtle, an oyster, a tunicate, and a termite were found to be intermediate between those of human and Drosophila, indicating that the discordant genomic diversity patterns that have been reported between these two species do not reflect a generalized vertebrate versus invertebrate gap. The genomic average diversity was generally higher in invertebrates than in vertebrates (with the notable exception of termite), in agreement with the notion that population size tends to be larger in the former than in the latter. The non-synonymous to synonymous ratio, however, did not differ significantly between vertebrates and invertebrates, even though it was negatively correlated with genetic diversity within each of the two groups. This study opens promising perspective regarding genome-wide population analyses of non-model organisms and the influence of population size on non-synonymous versus synonymous diversity.
| The analysis of genomic variation between individuals of a given species has so far been restricted to a small number of model organisms, such as human and fruitfly, for which a fully sequenced, well-annotated reference genome was available. Here we show that, thanks to next-generation high-throughput sequencing technologies and appropriate genotype-calling methods, de novo population genomic analysis is possible in absence of a reference genome. We characterize the genomic level of neutral and selected polymorphism in five non-model animal species, two vertebrates and three invertebrates, paying particular attention to the treatment of multi-copy genes. The analyses demonstrate the influence of population size on genetic diversity in animals, the two vertebrates (hare, turtle) and the social insect (termite) being less polymorphic than the two marine invertebrates (oyster, tunicate) in our sample. Interestingly, genomic indicators of the efficiency of natural selection, both purifying and adaptive, did not vary in a simple, predictable way across organisms. These results prove the value of a diversified sampling of species when it comes to understand the determinants of genome evolutionary dynamics.
| Population genomics, the analysis of within-species, genome-wide patterns of molecular variation, is a promising area of research, both applied and fundamental [1]. So far such studies have essentially been restricted to model organisms such as yeast [2] and Arabidopsis [3], in which a well-annotated, completely sequenced genome is available. In animals, the population genomic literature has long been dominated by drosophila and human (e.g. [4], [5]). Interestingly, these two species yielded very different patterns of genome variation. The per-site average synonymous nucleotide heterozygosity (πS), for instance, is roughly twenty times as high in Drosophila melanogaster (πS∼0.02 [6]) as in Homo sapiens (πS∼0.001 [7]) coding sequences. The ratio of non-synonymous to synonymous polymorphisms (πN/πS) is substantially lower, and the estimated proportion of adaptive amino-acid evolution (α) substantially higher, in D. melanogaster than in H. sapiens [8]–[12]. These distinctive patterns are interpreted as reflecting differences in effective population size (Ne) between human, a large vertebrate, and drosophila, a tiny invertebrate. A small Ne in human would explain the relatively low level of genetic diversity in this species, as well as a reduced efficacy of natural selection due to enhanced genetic drift, which would increase the probability of segregation of slightly deleterious mutations (hence the higher πN/πS), and decrease the probability of fixation of adaptive ones (hence the lower α [13], [14]).
The human-drosophila contrast, however instructive it has been for molecular evolutionary research, is a comparison between just two species, out of the millions of existing animals. It is unclear whether the same picture would have been reached if a distinct vertebrate and a distinct invertebrate species had been sampled. Population genomic statistics in D. simulans were found to be essentially similar to those of D. melanogaster [15], and the central chimpanzee (Pan troglodytes), although genetically more diverse than H. sapiens, showed genomic patterns consistent with a relatively low-Ne species [16]. These are knowledgeable corroborations, but from species very closely related to D. melanogaster or H. sapiens. A very high amount of synonymous diversity and a very low πN/πS ratio were reported in the tunicate Ciona intestinalis B [17]. This was interpreted as reflecting both a high mutation rate and large population size in this marine invertebrate species. Based on a small number of markers but many species, it was found that the average nuclear genetic diversity is higher in invertebrates than in vertebrates, and in marine than in terrestrial species [18], even though the difference is lower than expected from the neutral theory [19]. The influence of Ne was also invoked to explain the variations in non-synonymous to synonymous substitution rate between species of mammals [20], [21], and between populations of mice [22] and sunflower [23].
A recent population genomic study of the European rabbit (Oryctolagus cuniculus), however, revealed large amounts of genetic diversity, and a πN/πS ratio similar to those measured in Drosophila [24]. Although perhaps abundant, rabbits, being vertebrates, are among the 5% largest living animal species. Observing a very low πN/πS ratio in this species is somehow surprising according to the population size hypothesis, knowing that density and body mass tend to be negatively correlated across species (e.g. [25]). Still in mammals, relatively high levels of genomic polymorphism in endangered primate species were recently reported [26], again questioning the link between current abundance and population genomic patterns. It should be noted that what matters regarding molecular evolution is the long-term Ne, averaged over thousands to millions of generations. It is therefore perhaps not so surprising that the Ne effect in mammals is not correctly predicted by species conservation status, as discussed in reference [26]. At any rate, the sample of metazoan species for which population genomic data are available is still quite small, and highly biased towards mammals. Genome-wide studies of additional species from various phyla appear needed to confirm or infirm the role of Ne in animal molecular evolution, and to explore variations of within-species genomic diversity across the phylogenetic and ecological dimensions.
Next-Generation Sequencing (NGS) technologies potentially offer the opportunity to gather population genomic data in non-model organisms, in the absence of prior knowledge, at affordable cost. Genomes in animals can be large, highly repetitive and, consequently, difficult to assemble. The transcriptome appears as a valuable alternative target [26]. Transcriptomics gives access to large numbers of genes at relatively low cost, plus information about gene expression levels [27]–[29], with potential applications for SNP discovery and speciation genomics [30]–[32]. However, unlike PCR-based techniques, NGS does not return alleles or genotypes at well-defined loci, but rather large amounts of mixed, noisy, anonymous sequence reads. Extracting proper population genetic information from such data is a challenge, both conceptually and computationally. Starting from raw NGS transcriptomic data, one must assemble predicted cDNA, map reads, call single nucleotide polymorphisms (SNPs) and genotypes, and calculate population genetics statistics. Each of these steps requires appropriate methods and data-cleaning strategies to cope with paralogous gene copies, unequal expression level across genes, alternative splicing, transcription errors, sequencing errors and missing data, among other problems. Obviously, the whole task is especially difficult in the absence of a well-assembled reference genome.
Here we introduce a pipeline for de novo transcriptome-based NGS population genomics, which is applied to newly-generated data from five animal species – two vertebrates and three invertebrates. Based on samples of eight to ten individuals caught in the wild, we identify between ∼4,500 and ∼17,000 SNPs per species, from ∼2000–3500 distinct nuclear protein-coding genes. For each species, we separate synonymous versus non-synonymous variants, and estimate the level of genetic polymorphism, the amount of divergence to a closely-related outgroup, site-frequency spectra, and adaptive evolutionary rates. We assess the robustness of these statistics to various SNP-calling and data cleaning options, and to the presence/absence of a reference genome, paying specific attention to the removal of spurious SNPs due to hidden paralogy. Then we focus on the between-species variation in the average synonymous and non-synonymous levels of within-species diversity. Our expectation is that small-Ne species should show a lower πN, a lower πS, and a higher πN/πS ratio than large-Ne species. This is because genetic drift, which is enhanced in small populations, is expected to reduce the neutral and selected levels of genomic diversity, but to increase the relative probability of slightly deleterious, non-synonymous mutations (relatively to neutral, synonymous mutations) segregating at observable frequency. Our analyses suggest that the vertebrate versus invertebrate contrast is not an obvious predictor of Ne from a molecular evolutionary viewpoint.
Table 1 lists the five species studied in this work. The urochordate Ciona intestinalis is a model organism for evo-devo research [33]. The existence of two cryptic species, called A and B, has recently been discovered [34], [35]. C. intestinalis A, which occupies the Pacific Ocean and the Mediterranean Sea, was taken as the focal species in this study. The flat oyster Ostrea edulis is a marine bivalve of economic interest, which lives in the Eastern Atlantic coasts. C. intestinalis and O. edulis belong to two phyla, tunicates and bivalves, in which very high levels of within-species genetic diversity have been reported [17]–[19], [36]–[38]. The Iberian hare Lepus granatensis has attracted the attention as a model taxon for phylogeographic analysis and the study of speciation and reticulate evolution [39]. Its geographic range is limited to Iberia. The European pond turtle Emys orbicularis occurs in freshwater environments in Europe [40]. Both L. granatensis and E. orbicularis are terrestrial, medium-sized vertebrates, for which a relatively low Ne can be expected. The subterranean termite Reticulitermes grassei, finally, is a eusocial termite species occurring in Spain and south-west France, feeding on wood, and causing damage to human habitations. R. grassei is a small invertebrate, by far the smallest of the five species analyzed here. However, its effective population size is presumably highly reduced by eusociality – few individuals per colony contribute to reproduction. In the rest of the article, these five species will be designated as ciona, oyster, hare, turtle and termite, respectively.
A reference genome and transcriptome is available for two species of our panel, namely ciona, which was fully sequenced [41], and hare, which is closely related (∼5% divergence) to the fully-sequenced rabbit, O. cuniculus [24]. For these two species, reference-free population genomic inferences were compared to reference-based ones. For each of the five focal species, a closely-related outgroup was included in the study in order to perform divergence analyses. The outgroup was taken from the same genus as the focal species, except for the turtle, in which the outgroup was the pond slider Trachemys scripta (Table 1).
Table 1 describes the NGS data sets generated in this analysis. Nine to ten individuals per focal species and two to eight individuals per outgroup species were analysed. An average 7.85 millions single-ended illumina reads of mean length 89 were obtained per individual. In oyster, termite, hare, and turtle, 454 analysis of one or a pool of individuals provided an additional ∼500,000 reads of average length 306. Roughly 50% of the data were newly generated for this study. The other 50%, i.e., eight individuals each of ciona (B species), oyster, hare and turtle, were previously used to investigate various cDNA assembling strategies [42].
The data analysis pipeline is illustrated by Figure 1, and fully described in the Material & Methods section. Depending on the species, between 28,000 and 85,000 contigs were generated by a combination of Abyss and Cap3. Illumina reads were mapped onto the predicted cDNAs using BWA. Genotypes were called using program reads2snps, which implements the maximum likelihood framework introduced by Tsagkogeorga et al. [17], in which the per-contig error rate is estimated assuming a multinomial distribution of read counts and the Hardy-Weinberg equilibrium. When the posterior probability of the best-supported genotype (either homozygote or heterozygote) was below 0.95, the position was coded as missing data. Classical population genomic statistics were calculated based on these predicted genotypes, after various data cleaning steps, using custom-witten C++ programs. The number of contigs available for population genomic analyses – i.e., contigs which passed the coverage and ORF length filters – varied among species from 1978 to 3661. Note that the 454 reads were only used at the assembly step, not for individual genotyping.
In the genotype-calling procedure described above, we assume that all the reads that map to a given position correspond to a single locus. It might be, however, that reads from distinct loci map to the same place. This is expected to occur in cases of undetected paralogy, copy number variation, and repetitive genomes. In such cases, variation between paralogues might result in spurious heterozygous genotype calls. We introduced a new test to detect and clean these spurious heterozygotes. Briefly, the rationale is to compare the likelihood of a model assuming one bi-allelic locus with the likelihood of a model assuming two bi-allelic loci, both carrying the same two alleles (see Material and methods and Text S1 for details). Among the sites at which at least one heterozygous genotype was called, those for which the paralogy test was significant (p-val<0.001) were discarded. Depending on the species, between 7% (ciona) and 37% (hare) of SNPs were detected as potential paralogues.
Our major analyses involve comparison of population genetic statistics between species, and so it is important to be sure that these differences are due to real biological differences and not methodological artefacts. We first analysed the variations and impact of sequencing coverage across samples and genes. The average coverage of the analysed contigs varied from 5X to 15X across individuals and species after removal of potential PCR duplicates (Figure S1), oyster being slightly less covered, on average, than the other four species. The observed heterozygosity (i.e., the proportion of predicted heterozygous sites) was calculated for all individuals. Its relative level of variation among individuals was minimal in hare (0.0013–0.0018), and maximal in turtle (0.0003–0.0017). Importantly, this value was not correlated with the average sequencing depth in any of the five species – individuals for which large amounts of data were obtained were not more (or less) heterozygous, on average, than other individuals (Figure S1). The correlation coefficient of sequencing coverage across genes was typically above 0.9 for individuals from the same species, and declined when individuals from distinct species were compared, consistent with reference [26]. No correlation was found across species between the between-individual variance in sequencing depth and the mean or between-individual variance in heterozygosity (result not shown).
Then, in all five species, the contig containing the cox1 mitochondrial gene was identified by BLAST and individually analysed. Cox1 is a highly-expressed, haploid locus for which homozygous genotypes should be recovered if nuclear-encoded paralogs (the so-called “numt”) have been correctly filtered, and contamination between samples avoided. In turtle, ciona, oyster and termite, cox1 genealogies revealed monophyletic species, and amounts of within-species mitochondrial diversity below 1% (Figure S2). Examining the predicted SNPs, we found a single (in oyster) predicted heterozygous genotype out of the ∼40,000 genotyped positions. The average proportion of heterozygous genotypes across individuals and positions in these four species was 4.10−5, i.e., very low.
In hare, the cox1 tree revealed two divergent groups of L. granatensis haplotypes, of which one was more closely related to the arctic hare Lepus timidus. This is consistent with the documented introgression of L. timidus mitochondrial DNA into northern iberian populations of L. granatensis [39], [43]. A closer examination of the cox1 contig analysed here revealed that it was a complex chimera, i.e., a concatenation of fragments from the granatensis and timidus haplotypes, which are ∼10% divergent from each other. Six positions in this alignment contained unexpected heterozygous genotypes. Five of them were located close to (<30 bp away from) the boundary between a granatensis and a timidus fragment. The heterozygous genotypes correspond to low-coverage positions/individuals, which occurred when most reads from a specific individual had mapped to a distinct contig – the hare assembly included several other highly-covered contigs homologous to cox1, of length 200–460 bp. When a minimal coverage of 30X per individual, instead of 10X per individual, was required to call a genotype (our “high-coverage control”, see below), all the unexpected heterozygotes disappeared. We note that such a situation – two divergent, highly-expressed alleles coexisting in the population, with each individual carrying a single copy – is presumably very uncommon. The results of our main analyses were qualitatively unchanged when the three introgressed individuals were removed from the hare data set. To summarize, our analysis of the Cox1 gene were consistent with previous knowledge regarding mtDNA evolution in the five target species, and revealed a satisfying behaviour of our genotype-calling procedure, in its basic or high-coverage version.
Finally, we investigated the geographic patterns of genetic variation the five analysed species by plotting between-individual genetic versus geographic distance (Figure S3). A clear isolation-by-distance pattern was detected in ciona, in which the Mediterranean and Californian samples were differentiated, and in turtle, in which some population substructure associated with Pleistocene glacial refugia is detected. The relationship was much weaker in oyster, and absent in hare and termite. These patterns are essentially consistent with the phylogeographic literature in these five species [40], [44]–[47], which is typically based on fewer loci but many more individuals than the current study. The concordance between these two sources of data provides additional corroboration for our inferred SNPs and genotypes.
For each species, population genomic statistics were calculated and averaged across loci (Table 2, row A). Their robustness to various data cleaning/SNP calling options was examined in two species, ciona and hare, for which a full genome and a reference transcriptome are available.
Estimates of πN and, especially, πS were reasonably robust to the high-coverage control, even though fewer SNPs were called with the increased coverage/quality requirement (Table 2, row B). This is because requiring a higher quality decreases not only the number of predicted SNPs, but also the number of predicted homozygous positions. The slightly lower πN/πS ratio obtained from the high-coverage control might reflect a biological effect, i.e., stronger selective constraint on highly-expressed genes [48]. High levels of robustness were also obtained with respect to our “high-quality”, “threshold-free” and “clip-ends” controls (Table S2, row F, G, H).
Importantly, results were only weakly affected when reads were mapped on existing genomic references, rather than on predicted contigs (Table 2, row C). In ciona, both πN and πS were reduced by <10% in the reference-based control. In hare, the situation was a bit worse, with πN being reduced by ∼30% when reads were mapped to the rabbit transcriptome, while πS was unchanged. Note that in the case of hare, the reference is ∼5% divergent from our focal species, which might bias the sample towards evolutionarily conserved genes in the reference-based control. Taken together, the reference-based controls suggest that the uncertainty in cDNA prediction [42] does not impede de novo population genomic analysis from NGS transcriptomic data.
When potentially spurious SNPs due to undetected paralogy were not filtered out, the total number of analysed SNPs increased, as could have been expected (Table 2, row D). This change did not dramatically affect πS and πN, but a lower (i.e., more negative) FIS was obtained when the paralog filter was off. Negative FIS denotes an excess of heterozygotes, as compared to the Hardy-Weinberg expectation. This is unexpected from natural population samples, in which population structure and inbreeding typically result in a deficiency, rather than an excess, of heterozygotes. The observed decrease in FIS when the paralog filter was switched off suggests that erroneous SNPs/genotypes due to mapping problems are common, and that filtering them out is necessary. The slightly negative FIS measured in our main ciona and hare analysis suggest that the filter does not entirely solve the problem.
Our results were compared to an entirely different data analysis pipeline based on samtools [49] (Table 2, row E). The two approaches yielded similar results in ciona, but in hare πS was slightly decreased, and πN/πS substantially increased, when samtools was used. The same trend was observed in oyster, termite and turtle, to various extents (Table 2). To investigate further the causes of this discrepancy, we computed site frequency spectra (SFS) from the genotypes predicted by samtools versus reads2snps (our main analysis). Figure 2 displays the folded synonymous and non-synonymous SFS in hare. As far as reads2snps predictions were concerned, the proportion of low-frequency variants was higher in non-synonymous SNPs than in synonymous SNPs, as previously reported in human [13] and drosophila [50]. This is expected under the hypothesis of a prevalent influence of purifying selection on non-synonymous mutations. Such a pattern was not observed with the samtools-predicted SNPs, in which the synonymous and non-synonymous SFS were similar to each other, and similar to the SFS expected in a neutrally evolving, panmictic, Wright-Fisher population (Figure 2, left), in which the probability of observing a SNP at a derived allele frequency of k is proportional to 1/k [51]. The inferred SFS for the other four species are displayed in Figure S4. A pattern similar to the hare was observed in turtle and termite. In ciona and oyster, the contrast between the synonymous and non-synonymous spectra was weaker.
The samtools and reads2snps genotype callers differ in two main aspects. First, reads2snps does not make use of sequence quality data, and, instead, estimates the error rate, assumed to be constant across positions in a contig, from the data. When the analysis was restricted to high-quality reads only, reads2snps-based SFS were essentially unchanged (results not shown), which does not suggest that the treatment of sequencing errors is an issue here. Secondly, reads2snps places no explicit prior on the SFS, whereas the samtools caller uses a Wright-Fisher prior (equation 20 in [52]). This could explain the difference between reads2snps-predicted and samtools-predicted SFS, and especially the higher similarity of samtools-predicted SFS, both synonymous and non-synonymous, to the Wright-Fisher expectation, as reflected in Tajima's D values that are closer to zero (Figure 2, Figure S4).
Sequences from outgroup species were added to within-species alignments. Contigs showing extreme levels of synonymous divergence between focal and outgroup species (i.e., genes that exceeded the median dS by two standard deviations or more) were considered as dubious and discarded. Outgroup inclusion resulted in a strong decrease in number of analysed contigs,and a slight reduction in estimated πN/πS ratio (Table S2, row I). This presumably reflects a more accurate prediction of ORFs when data from two distinct species are available, and/or an increased level of selective constraint on the subset of genes for which orthology search was successful.
We examined the robustness of our results to individual sampling. We generated random sub-samples of five to nine individuals (all combinations), and re-called SNPs and genotypes. Figure 3 shows the distribution of πS and πN across sub-samples, as a function of sub-sample size, in turtle (green) and ciona (blue). In turtle, no sampling bias was detected: the average estimated πS and πN did not vary with sub-sample size. The standard deviation across all sub-samples was 5% of the πS estimate, and 7% of the πN estimate. In ciona, no bias was detected for πS, but the estimated πN slightly declined as sub-sample size decreased. The median πN across sub-samples of five individuals was 23% lower than the estimate obtained from all ten individuals. The coefficient of variation was still relatively low for both πS (8%) and πN (12%). The hare pattern was similar to turtle, and the oyster and termite patterns similar to ciona. The reasons for a decline of πN with sub-sample size in three species are unclear. The occurrence of this pattern does not appear related to the existence of population substructure (Figure S3). At any rate, this analysis indicates that our estimates of within-species synonymous and non-synonymous diversity are reasonably robust to sampling size, and that the sampling variance is well below the reported between-species differences.
Table 3 summarizes the population genomic statistics, calculated using our main settings, in the five species analysed in this study, with outgroup. The two vertebrates, hare and turtle, were less polymorphic than the three invertebrates, as could have been expected from intuition about population sizes. Ciona was the most polymorphic species of our panel. This is in line with the analysis of Tsagkogeorga et al, who reported an extremely high πS in the congeneric C. intestinalis B [17]. Oyster, perhaps surprisingly, was not much more polymorphic than the two vertebrates as far as synonymous sites were concerned. A similar πS estimate (0.07) was obtained by E. Harrang (personal communication) based on 37 loci Sanger-sequenced in a sample of 20 flat oysters. Termite, finally, was the least polymorphic species of the panel, consistent with the expectation of a reduced population size associated to eusociality.
Figure 4a plots genomic average πN against genomic average πS across 19 animal species for which such estimates are available from the literature ([10], [15]–[17], [24], [26], estimates obtained from at least four individuals caught in the wild and 1000 genes). This figure shows that the five species sampled here (closed circles) are intermediate between human and drosophila in terms of within-species diversity. Vertebrates (in blue), here represented by thirteen mammals (among which nine primates) and one turtle, showed an average πS below 0.01, and an average πN below 0.0006. More variance was detected within the group of invertebrate species, in which termite was a clear outlier. Both πS and πN reached in invertebrates values well above the maximal records of mammals and turtle. So a vertebrate versus invertebrate gap in genomic diversity is still apparent in Figure 4a, even though the contrast is not as sharp as suggested by the sole human versus drosophila comparison – and please note that the vertebrate taxon sampling is still highly biased towards mammals.
In Figure 4b, the πN/πS ratio was plotted as a function of πS. A significant negative relationship was recovered both in vertebrates (r2 = 0.43, p-val<10−5, n = 14) and invertebrates (r2 = 0.86, p-val = 0.002, n = 5), in agreement with the hypothesis of a population size effect on the efficiency of purifying selection. However, the average πN/πS ratio was not significantly higher in invertebrates than in vertebrates, and the correlation coefficient computed across all 19 species (r2 = 0.18) was not significantly different from zero. This is an intriguing result, which does not seem to accommodate well the idea of a Ne-dependent πN/πS ratio. Figure 4b was unchanged when the average πS was calculated from one half of the contigs, and the average πN/πS from the other half, thus removing any intrinsic dependence between the two variables (not shown). The ratio of non-synonymous to synonymous divergence, dN/dS, was also negatively correlated to πS, again in agreement with the hypothesis of a more efficient purifying selection in large populations (Figure S5).
The proportion of adaptive amino-acid substitutions, α, was estimated using two distinct methods based on the McDonald-Kreitman principle [8], and the (per synonymous substitution) rate of adaptive non-synonymous substitution, ωa, was computed too. Estimates of α varied from 0 to 0.9 among species and methods. In hare, the DoFE program returned a highly negative, aberrant value for α when the method of reference [53] was used. These estimates showed no obvious correlation with variations in effective population size. Neither α nor ωa were found to be higher in invertebrates than in vertebrates when low-frequency variants were appropriately handled (Figure S5). Our data, therefore, do not bring support to the hypothesis of a higher adaptive rate in large-Ne species, in contrast with several recent reports [22], [23], [54], [55]. We note that theoretical predictions are equivocal regarding the α/ωa/Ne relationships: the adaptive rate itself appears to be strongly limited by linkage and hardly influenced by Ne (assuming large enough populations and a constant supply of advantageous mutations [56], [57], and under purifying selection alone the α/Ne relationship can be complex [58].
Here we show that population genomics is possible in absence of a reference genome, thanks to an appropriate treatment of NGS data. Based on de novo assembled contigs, predicted ORF, empirical estimation of sequencing/mapping error rate and statistical filtering of potential paralogs, we recovered estimates of the major population genomic statistics that were reasonably similar to the ones obtained using published genomic annotations. Our estimates were robust to various methodological options, including constraints on sequence quality and coverage, threshold-based versus threshold-free genotype calling, and sub-sampling of contigs or individuals. Our results are consistent with a larger amount of within-species genetic diversity in invertebrates than in vertebrates (with exceptions), but question the relevance of Ne as a determinant of the πN/πS ratio and the adaptive substitution rate, which did not differ between vertebrates and invertebrates in our analysis.
From the several control steps we implemented, the most problematic issue we faced in this analysis was due to hidden paralogy, which manifested itself through spurious polymorphic positions at which many individuals, if not all, were heterozygous, and shared a common highly-expressed (and a common lowly expressed) allelic state. Dou et al. [59] recently highlighted this problem, and proposed a method to overcome it, based on the idea that sequencing coverage is expected to be higher in repeated than in unique genomic regions. This approach does not apply to transcriptomic data, in which coverage primarily reflects the level of gene expression, which is not only determined by gene copy number. We introduce a novel filtering method based on explicit modelling of the single versus multiple copy cases. Our analyses indicate that this method removes a large fraction of hidden paralogy instances, as suggested by the substantial reduction in heterozygote excess in ciona and hare. We presume that hidden paralogy will be identified as the major caveat of de novo population genomics in future research, as suggested by the relatively large amount of dubious SNPs that were filtered out in this analysis. Besides the paralogy issue, our results were quite robust to the several methodological options we tried. In particular, both πS and individual heterozygosity were unrelated to sequencing depth (Table 2, high-coverage control and Figure S1) – a desirable property of NGS-based population genomic studies.
The two SNP-calling approaches we used yielded correlated (across species) but distinct results, with samtools predicting a lower SNP density than our reads2snps method. The two approaches differ in several aspects, including quality-based versus sequence-based estimation of the error rate, and whether a Wright-Fisher prior was used. Obviously, even slight differences in methodological design can have detectable consequences on the predicted genotypes, as suggested by the comparison between samtools-predicted and reads2snps-predicted site frequency spectra (Figure 2). These results highlight the need for an empirical assessment of the relative merits of the various SNP-calling methods that were published during the last two or three years (reviewed in [60]). Importantly, the two approaches used in this study yielded results reasonably consistent across species, so that the biological conclusions to be drawn (see below) are probably not method-dependent.
The major part of the existing population genomic literature in animals is restricted to drosophila and apes. These two groups of species show contrasting patterns of within-species genetic variation, with drosophila being ∼20 times as polymorphic as humans, showing more efficient purifying selection, and higher rates adaptive evolution. Here we uncovered the population genomic profile of five new non-model species – two vertebrates and three invertebrates. These five new species appear intermediate between human and drosophila in terms of genomic diversity (Figure 4). This suggests that the typical vertebrate versus invertebrate contrast is perhaps not as sharp as suggested by the human versus drosophila comparison. So far a single species, C. intestinalis B, has been documented to be more polymorphic than drosophila ([17], right-most circle in Figure 4), and a single one, aye-aye, as less polymorphic than human (based on just two individuals [26]). Still, the vertebrate versus invertebrate divide is apparent in Figure 4, in which all the vertebrate species show a per-site synonymous heterozygosity below 1%, and a per-site non-synonymous heterozygosity below 6‰. This is also true of the turtle E. orbicularis, the single non-mammalian vertebrate included in this figure. This result appears consistent with the hypothesis that effective population size (Ne) is generally higher in invertebrates than in invertebrates. The termite pattern is also quite consistent with intuitive expectations about population size: a colony of termites is comparable to many vertebrate species in terms of mass and life-history traits. Our report in termite of a significant deficit in heterogygotes (FIS>0.1) but no population structure (Figure S3D) is indicative of high levels of inbreeding, consistent with previous analyses in subterranean termites [61]. This tends to further reduce the effective population size in this species.
Species biology and ecology, however, does not explain every aspect of our data analysis. Hare, for instance, shows a lower πS and a much higher πN/πS ratio than rabbit, even though the two species are closely related, both phylogenetically and ecologically. The difference in πN/πS between the two species is even stronger when our samtools-based hare estimates are considered – i.e., the very data analysis pipeline used in rabbit [24]. Similarly, C. intestinalis A shows evidence for a smaller population size than its sister species C. intestinalis B – πS in A is four times as low as in B, and πN/πS twice as high – even though the two taxa are morphologically and ecologically indistinguishable. Finally, an unexpectedly low, vertebrate-like πS value is reported in flat oyster, despite the abundance of these marine animals in European Atlantic coasts
Most intriguingly, no significant difference was detected between vertebrates and invertebrates regarding the πN/πS ratio, even though πS and πN/πS were found to be negatively correlated across vertebrates, and across invertebrates. This is paradoxical: if a population size effect indeed accounted for the negative slopes within vertebrates and within invertebrates, then why not across the whole data set? Several explanations can be suggested. First, it must be recalled that the data points in Figure 4 were taken from several distinct studies, based on distinct gene samples, and distinct data analysis methods. Perry et al. [26], for instance, only selected SNPs covered at 30X or more, equivalently to our “high-coverage” control, which yielded a slightly reduced πN/πS ratio in ciona and hare as compared to our main analysis. It would be good to confirm the pattern of Figure 4b using a larger number species, especially non-mammals, and a common analysis strategy. Another potential methodological issue comes from our across-loci πN/πS averaging procedure, in which mean(πN/πS) is estimated as mean(πN)/mean(πS) (see Material and Methods), which might create a downward bias of unequal magnitude among species [12].
Alternatively, the distinctive behaviour of vertebrates and invertebrates in Figure 4b might reflect a true biological difference between these two groups of species. Differences in mutation rate, hereafter noted μ, could be invoked. The πN/πS ratio is independent of μ, whereas πS is essentially proportional to μ. So if μ was generally higher in invertebrates than in vertebrates, then a higher πS would be expected in the former than in the latter, for a given πN/πS ratio. However, let us recall that what matters regarding πS is the per-generation mutation rate. Published estimates of the per-generation μ indicate that this parameter is lower, not higher, in D. melanogaster and in the nematode Caenorhabditis elegans than it is in human and mouse [62], [63]. So, even though a potential influence of μ on the pattern of Figure 4b cannot be formally ruled out, current knowledge on across-species mutation rate variations would tend to even reinforce the paradox.
Selection on synonymous positions might also be a confounding factor. The genes used in this transcriptome-based study are the most highly expressed ones, i.e., prone to selection on codon usage for translation efficiency. Selected codon usage, which is documented in Drosophila but not in human [64], leads to a reduction in πS, and therefore an increase in πN/πS, irrespective of functional constraint on amino-acids. In mammals, synonymous positions are affected by GC-biased gene conversion [65], a neutral process that mimics natural selection, and is also expected to result in a decrease in πS. Substantial selective contraints on synonymous sites for efficient splicing of mRNA and nucleosome positioning are also documented, especially in mammals [66]. However, we note that such effects should affect both the X-axis (πS) and the Y-axis (πN/πS) of Figure 4b, so that a non-neutral behaviour of synonymous sites, if any, should essentially result in a re-scaling of the axes, not a shift upward of a subset of data points.
Another potential explanation to this unexpected pattern would invoke a difference in the selective regime between vertebrates and invertebrates. For a given Ne, the πN/πS ratio is expected to increase as the distribution of selection coefficients, s, of non-synonymous deleterious mutations becomes more leptokurtic [67]. One could imagine, for instance, that metabolic and protein interaction networks are more complex in vertebrates than in invertebrates [68], [69], so that the average amino-acid position is involved in a higher number of physical interactions, reducing the proportion of effectively neutral sites in vertebrates. This is consistent with the theoretical prediction of an increased variance in the distribution of deleterious selection coefficients as mutational pleiotropy increases [70]. Between-species differences in the distribution of deleterious selection coefficients are documented, with animals (drosophila and caenorhabditis) showing a higher average effect and a lower skewness as compared to micro-organisms [71].
Finally, it might be that vertebrates and invertebrates differ in their biology in such a way that the neutral and the selected levels of diversity do not respond similarly to demographic variations in the two groups. The invertebrates of this study are high-fecundity species: very large numbers of propagules (eggs, larvae, alates) are released every generation, each with a very small probability of survival to adulthood. This life cycle results in a highly skewed distribution of offspring, in which a minority of progenitors contributes to the next generation [72]. This departure from the Wright-Fisher model distinctively affects the fate of neutral [73]–[75] and selected [76] mutations, so that πS and πN/πS might respond non-linearly. At any rate, our results revivify old questions raised at the onset of experimental population genetics [77] that have been left unsolved during the long time-lag required to be able to conduct population genomics in non-model species [78].
In this study, we showed that de novo population genomics in non-model taxa can be achieved based on transcriptome data. Our analysis demonstrates the contrast between vertebrates and invertebrates regarding πN and πS, with exceptions (termites), but detects no significant difference as far as πN/πS is concerned, questioning the hypothesis that neutral and selected levels of diversity are uniquely determined by the variations of a one-dimensional variable – i.e., Ne – across organisms. The methods developed in this study will be worth applying to additional animal species to explore further the influence of species ecology on population genomics, and the role/meaning of effective population size in molecular evolution.
Nine or ten individuals per focal species, and one to eight individuals per outgroup species, were sampled from three to ten localities across the species range. Details on sampling dates and locations are available from Table S1. Tissues were preserved from RNA degradation using liquid nitrogen, RNAlater buffer or Guanidinium thiocyanate-Phenol solution (Trizol and TriReagent BD ) was used for termites, hares and ciona. Silica membrane - SM kits (RNEasy, Qiagen) was used for hares and ciona. We previously developed a third RNA isolation method using combined GTPC and SM [79], used here for oysters and turtles. RNA quantity and quality (purity and degradation) was assessed using NanoDrop spectrophotometry, agarose gel electrophoresis and Agilent bioanalyzer 2100 system before external sequencing (GATC, Konstanz Germany). See Table S1 and reference [79] for additional details.
Five µg of total RNA of each sample were used to build 3′-primed, non-normalized cDNA libraries, sequenced using Hiseq2000 or Genome Analyzer II (Illumina) with 8 and 5 libraries pooled per lane, respectively. Fifty bp (termite) or 100 bp (other four species) single-end reads were produced. In hare, turtle and oyster, 25 µg of total RNA of one individual per focal species was used to build a random-primed normalized cDNA library. The latter was sequenced for half a run with GS FLX Titanium (Roche ). Low quality bases, adaptors and primers were removed using the SeqClean program (http://compbio.dfci.harvard.edu/tgi/).
Figure 1 summarizes the main data analysis strategy used in this study. For each focal species, 454 and Illumina reads were assembled in contigs – i.e., predicted cDNAs – using the Abyss and Cap3 programs [80], [81], according to method D in [42]. In this approach, 454 and Illumina reads are separately assembled then merged in a mixed assembly thanks to an additional Cap3 run. Illumina reads were mapped to the contigs using BWA [82]. For each contig, average coverage was defined as the total length of mapped reads divided by contig length. Contigs less covered than an average 2.5 X per individual were immediately discarded. Open reading frames (ORF) were predicted the program transcripts_to_best_scoring_ORFs.pl, which is part of the Trinity package (http://trinityrnaseq.sf.net, courtesy of Brian Haas). This program makes use of hexanucleotide frequencies, learnt from a first pass on the data, to annotate coding sequence boundaries.
For each position of each contig and each individual, genotypes were called using the method introduced by Tsagkogeorga et al. [17] (M1 model), specifically designed to handle transcriptome-based NGS data, and implemented in the home-made program reads2snps. Briefly, this method first estimates the error rate (assumed to be shared across positions) in the maximum likelihood framework, then calculates the posterior probability of each of the 16 possible genotypes knowing the error rate, assuming Hardy-Weinberg equilibrium. When one genotype, either homozygous or heterozygous, had a posterior probability above 0.95, it was validated. Otherwise, the genotype was coded as missing data. In contrast with “variant calling” approaches (in which a homozygote is called in case of insufficient power to detect a heterozygote), no coverage-associated bias in heterozygosity prediction is expected with this method. Positions in which no more than 10 reads were available for a specific individual were also considered as missing. Prior to SNP/genotype calling, potential PCR duplicates were removed by collapsing sets of identical reads into a single read.
Paralogous gene copies are a potential source of spurious SNPs: if two distinct genes were merged in a single contig at the assembly step, then between-copy variations might be mistaken for heterozygosity. To cope with this problem, the detected SNPs were filtered for potential paralogy thanks to a newly-developed likelihood ratio test. Briefly, for a given SNP, the probability of the observed data (read counts for A, C, G and T in every individual) was calculated under the one-locus model used for SNP calling [17], on one hand, and under a two-locus model, on the other hand. The two-locus model assumes that two paralogous loci contribute reads to this SNP, with locus 1 contributing a proportion p of the reads. The two-locus model predicts an excess of heterozygotes (assuming that every individual carries and expresses the two loci), and correlated read count asymmetry across individuals (assuming that the relative contribution p of locus 1 is constant among individuals). SNPs were validated when the two-locus model did not significantly improve the fit, as compared to the one-locus model. In this test, potential departure from the 50%/50% expectation for read counts in heterozygotes was taken into account by assuming a Dirichlet-multinomial distribution of read counts, instead of a standard multinomial. Such an overdispersion of read counts is expected in case of allele-specific expression bias [83], and because of the stochasticity of allele amplification during library preparation [84]–[85]. Details of the method and simulations are provided in Text S1. The reads2snps SNP-caller and paralogue filter can be downloaded from http://kimura.univ-montp2.fr/PopPhyl/resources/tools/reads2snp.tar.gz.
Outgroup sequences were added to these alignments, when available. To achieve this aim, Illumina reads from the outgroup species were assembled using Abyss and Cap3, following method B in reference [42], and ORF were predicted as above. Orthologous pairs of coding sequences from the focal and the outgroup species were identified using reciprocal best BLAST hit, a hit being considered as valid when alignment length was above 130 bp, sequence similarity above 80%, and e-value below e−50. Outgroup sequences were added to within-focal species alignments using a profile-alignment version of MACSE [86], a program dedicated to the alignment of coding sequences and the detection of frameshifts. Contigs were only retained if no frameshift was identified by MACSE, and if the predicted ORF in the focal species was longer than 100 codons.
Codon sites showing a proportion of missing data above 50% were discarded. Then focal species sequences showing a proportion of missing data above 50% were removed. Alignments made of less than 10 codon sites after cleaning were removed. For each contig, the following statistics were calculated using the Bio++ library [87]: per-site synonymous (πS) and non-synonymous (πN) diversity in focal species, heterozygote deficiency (FIS), number of synonymous (pS) and non-synonymous (pN) segregating sites in focal species, number of synonymous (dS) and non-synonymous (dN) fixed differences between focal and outgroup species, neutrality index NI = (pN/pS)/(dN/dS) [88], and neutrality index calculated after removing SNPs for which the minor allele frequency was below 0.2 (NI0.2). These statistics were computed from complete, biallelic sites only – i.e., sites showing no missing data after alignment cleaning, and no more than two distinct states. The per-individual heterozygosity (proportion of heterozygote positions) was also calculated.
For each species, statistics were averaged across contigs weighting by contig length, thus giving equal weight to every SNP. Confidence intervals around estimates were obtained by bootstrapping contigs. Averaging population genomic statistics across loci can be problematic when ratios have to be calculated. The ratio of mean(πN) to mean(πS), for instance, is a biased estimate of the mean(πN/πS) if selective constraint on non-synonymous sites and neutral diversity are correlated across genes [12]. A correction for this bias was proposed [89], which is valid only if the number of synonymous SNPs per contig is large enough. This correction is not applicable to our data set, in which a majority of contigs are relatively short, and therefore include small numbers of synonymous SNPs.
The synonymous and non-synonymous site frequency spectra (SFS, i.e., the distribution of minor allele counts across SNPs) were computed based on predicted genotypes. To cope with the variable sample size across SNPs, we applied a hypergeometric projection of the observed SFS into a subsample of n = 12 sequences [90], SNPs sampled in less than n sequences being discarded. The synonymous and non-synonymous SFS were used to calculate Tajima's D [91], and to estimate the proportion of adaptive amino acid substitutions according to the method of Eyre-Walker and Keightley [53] using the DoFE program (http://www.lifesci.sussex.ac.uk/home/Adam_Eyre-Walker/Website/Software) – an estimate we call αEWK. This proportion was also estimated as α0.2 = 1−NI0.2 [13]. We finally calculated the (per synonymous substitution) rate of adaptive non-synonymous substitution, ωa = α dN/dS [54].
Several aspects of the pipeline described above were modified in order to assess the robustness of population genetics estimates to methodological options. Here are the main alternative strategies that were explored.
|
10.1371/journal.pcbi.1005857 | Automated visualization of rule-based models | Frameworks such as BioNetGen, Kappa and Simmune use “reaction rules” to specify biochemical interactions compactly, where each rule specifies a mechanism such as binding or phosphorylation and its structural requirements. Current rule-based models of signaling pathways have tens to hundreds of rules, and these numbers are expected to increase as more molecule types and pathways are added. Visual representations are critical for conveying rule-based models, but current approaches to show rules and interactions between rules scale poorly with model size. Also, inferring design motifs that emerge from biochemical interactions is an open problem, so current approaches to visualize model architecture rely on manual interpretation of the model. Here, we present three new visualization tools that constitute an automated visualization framework for rule-based models: (i) a compact rule visualization that efficiently displays each rule, (ii) the atom-rule graph that conveys regulatory interactions in the model as a bipartite network, and (iii) a tunable compression pipeline that incorporates expert knowledge and produces compact diagrams of model architecture when applied to the atom-rule graph. The compressed graphs convey network motifs and architectural features useful for understanding both small and large rule-based models, as we show by application to specific examples. Our tools also produce more readable diagrams than current approaches, as we show by comparing visualizations of 27 published models using standard graph metrics. We provide an implementation in the open source and freely available BioNetGen framework, but the underlying methods are general and can be applied to rule-based models from the Kappa and Simmune frameworks also. We expect that these tools will promote communication and analysis of rule-based models and their eventual integration into comprehensive whole-cell models.
| Signaling in living cells is mediated through a complex network of chemical interactions. Current predictive models of signal pathways have hundreds of reaction rules that specify chemical interactions, and a comprehensive model of a stem cell or cancer cell would be expected to have many more. Visualizations of rules and their interactions are needed to navigate, organize, communicate and analyze large signaling models. In this work, we have developed: (i) a novel visualization for individual rules that compactly conveys what each rule does, (ii) a comprehensive visualization of a set of rules as a network of regulatory interactions called an atom-rule (AR) graph, and (iii) a set of procedures for compressing the AR graph into a pathway diagram that highlights underlying signaling motifs such as feedback and feed-forward loops. We show that these visualizations are compact and informative across models of widely varying sizes. The methods developed here not only improve the understandability of current models, but also establish principles for organizing the much larger models of the future.
| Rule-based frameworks such as BioNetGen [1–3], Kappa [4–6] and Simmune [7,8] have been used to build detailed kinetic models of signaling pathways (e.g., FcεRI [9–11], TCR [12], EGFR [13,14], and p53 [15]). A rule-based model is composed of multiple “reaction rules”, where each rule specifies a reaction mechanism and its structural requirements, e.g., a phosphorylation rule would specify the set of binding interactions that bring the kinase into contact with substrate and the specific site on the substrate that is phosphorylated. Current models range in size from tens to hundreds of reaction rules, but these numbers are expected to increase as rule-based models are collectively organized in databases of kinetic interactions [10,12,14,16] and eventually integrated into whole cell models [17]. Large models, whether rule-based or otherwise, are difficult to understand or communicate without good visualization methods. Currently, the size of rule-based model that can be simulated far exceeds the size of model for which useful visualizations can be constructed automatically. In particular, we do not have visualizations that can present the regulatory interactions embedded in a model as a network diagram of signal flows. Also, other than using manual approaches, we do not have an effective approach to build compact pathway diagrams to communicate the model. Solving the automated diagramming problem is necessary to make the leap from opaque machine-readable model descriptions that can only be understood through manual annotation to transparent models that can be understood and explored by the wider community.
Why is it challenging to visualize rule-based models? Tools that formally visualize the model tend to focus on a single type of information, such as what molecular structures are being modeled (contact map [6]), what rules have been defined on those structures (Simmune [8], Virtual Cell [18,19], BioUML [20]), and how various rules interact with each other (rule influence diagram [21], Kappa story [22]). To communicate the architecture of the model at a global level, these different types of information have to be integrated into a single diagram, but current approaches such as the Extended Contact Map (ECM) [23], the Systems Biology Graphical Notation: Entity Relationship Diagram (SBGN:ER) [24] and the Molecular Interaction Map (MIM) [25] rely on human interpretation, which decouples the diagram from the executable model. Methods to automate generation of diagrams include the Simmune Network Viewer [26], which uses an interactive approach to visualization, and the rxncon regulatory graph [27], which has a simplified representation of rule-based models that is more amenable for visualization than standard rules. In Fig 1, we apply a contact map, a conventional rule visualization approach, a rule influence diagram and an extended contact map to a previously published model of immunoreceptor signaling [9], and below, we discuss the issues raised by each type of information displayed in those diagrams. We also present more detailed comparisons to the remaining tools in Discussion.
The contact map (Fig 1A) conveys the structural composition of a model (e.g. in [28–30] and others) by showing what types of molecular structures are available to compose reaction rules [6]. This includes structured objects called molecules, components, states and bonds, which we explain in more detail in the Methods section.
Conventional rule visualizations (Fig 1B) show reaction rules as reactant to product transformations. The reactant side includes not just the structures that are to be modified in the rule, but also the structural requirements that need to be matched for the rule to be triggered. To determine the action of a rule, the reader has to compare reactants to products, which can be challenging for complex rules that have a number of structural requirements (e.g., rules in the center column of Fig 1B). Nevertheless, this is the standard approach to show rules (e.g., in [9,13,28] and others), whether using manually drawn diagrams such as Fig 1B or automated diagrams generated by various software (Simmune [8], Virtual Cell [18,19], BioUML [20]).
The rule influence diagram (Fig 1C) represents each rule with a single node and each computed interaction between rules as a directed edge [21,31]. Each rule interacts with other rules through shared structures, e.g., a binding rule that produces a kinase-bound configuration regulates a phosphorylation rule that requires the same configuration. However, it is difficult to understand regulatory interactions from just the rule influence diagram because it does not show structures interacting with rules. Also, even moderate-sized models produce unreadably dense diagrams such as Fig 1C, and the computation of influences is quadratic in the number of rules, which is limiting for large models. Both BioNetGen and Kappa frameworks can generate rule influence diagrams, with the Kappa version allowing for different levels of precision [31].
The extended contact map (Fig 1D) is an expert-curated diagram that highlights functional roles of various structures and mechanisms as well as emergent regulatory architectures such as feedbacks and cascades [23]. It uses standard diagramming conventions to convey function (e.g., round arrowhead to indicate phosphorylation), annotation to relate diagram to model (e.g., edge label 2 pointing to rule number 2), and secondary documentation to convey biological significance (e.g., an attached model guide that indexes and describes each rule). Each of these components is constructed manually, which is also true for related methods such as SBGN:ER [24] and MIM [25] (see Discussion). Several recent models make use of the ECM ([10,12,14,32] and others).
In this work, we introduce three new methods that together constitute a new visualization framework for rule-based models. First, we introduce a novel compact rule visualization, which is more concise than conventional representations of rules and does not require visual comparison to convey the action of the rule. Second, we develop the atom-rule (AR) graph for showing regulatory interactions that can be efficiently derived from rules without pairwise comparisons. The bipartite AR graph displays a global view of how rules interact through the structures present in a model. Finally, because the raw AR graph is too dense for many applications, we present an AR graph compression pipeline that integrates expert knowledge and generates more readable diagrams. These methods are compatible with rules from the three widely-used frameworks of BioNetGen [1–3], Kappa [4–6] and Simmune [7,8] and also with the proposed interchange format SBML-multi [33]. We have provided an implementation in BioNetGen 2.2 [3], which is already available to users and to frameworks that incorporate BioNetGen, such as PySB [34] and Virtual Cell [18,19].
The remainder of the paper is organized as follows. In Methods, we briefly describe the new visualization methods and apply them to simple examples. In Results, we apply the methods to larger and more complex models, including a test set of 27 rule-based models from the literature. We use standard measures of graph readability to show that our methods produce more readable diagrams than current alternatives. In Discussion, we present additional comparisons with existing tools and discuss the potential benefits of the new tools for analysis of rule-based models.
The frameworks of BioNetGen [1–3], Kappa [4–6] and Simmune [7,8] share similar rule-based representations for which several formal treatments have been presented in the literature (BioNetGen [35–37], Kappa [4,5,22]). The visualization tools developed in this work have been implemented in BioNetGen, but operate on features of rule-based modeling common to all three frameworks. We recommend Chylek et al. [38] for a recent review of rule-based modeling, Sekar et al. [39] for a BioNetGen tutorial, and Hogg et al. (Supplement) [36] for a description of the BioNetGen formalism. In this section, we use a simple rule-based model to introduce reaction rule syntax and semantics, then demonstrate our new visualization approaches, namely compact rule visualizations and atom-rule graphs. S1 Appendix provides a more detailed theoretical foundation as well as specifications for algorithms and rendering conventions. S2 Appendix provides a step-by-step tutorial for applying methods to a complex signaling model from Suderman and Deeds [40].
In a rule-based model, molecules are structured objects composed of components. Fig 2A shows the BioNetGen language (BNGL) specification of molecules Enz and Sub representing enzyme and substrate respectively, along with corresponding visualizations. Enz has component sub and Sub has components enz, p1 and p2. By convention, a component with a binding function is named after the molecule that it binds. So, sub on enzyme and enz on substrate represent binding sites for substrate and enzyme respectively. Components p1 and p2 represent phosphorylation sites.
A component may have one or more modifications available to it, called internal states. For example, components p1 and p2 may be in the unphosphorylated state Y or phosphorylated state pY. Bonds can occur between pairs of components. Here, component sub on an Enz molecule can bind component enz on a Sub molecule to form an enzyme-substrate complex. Patterns, which are constructed from molecules, components, internal states and bonds, specify the reactants and products of a reaction rule. In Fig 2B, we show the BNGL specification of a simple enzyme-substrate system. Each rule requires a rate constant, with reversible rules, such as R1, requiring rate constants for both forward and reverse directions. In Fig 2C, we visualize the rules using a conventional approach.
Each reaction rule explicitly encodes model assumptions about a reaction mechanism. Structural features specified on the reactant side and modified on the product side constitute the reaction center. In rule R1 and its reverse, the sub-enz bond is formed in the forward direction and removed in the reverse direction, which indicates that R1 models reversible enzyme-substrate binding. In rule R2, the unphosphorylated state of p1 is transformed to the phosphorylated state, which indicates that R2 models phosphorylation of component p1. Analogously, rule R3 models phosphorylation of component p2.
Features that remain the same on both sides of a rule constitute reaction context, which describes the local conditions necessary for the mechanism to occur. In rules R2 and R3, the sub-enz bond is present on both sides of the rule, which indicates that the respective phosphorylation mechanisms require the enzyme-substrate binding interaction. Features omitted on both sides of the rule are assumed not to affect the reaction mechanism. Components p1 and p2 are omitted in rule R1 and its reverse, which specifies that binding and unbinding mechanisms are independent of p1 and p2. Similarly, rules R2 and R3 specify that phosphorylation at p1 is independent of p2 and vice versa.
The site graph is a nested graph used to represent patterns [22], such as the reactants and products in Fig 2C. In this work, we use site graph to refer to the visualization scheme where nodes representing molecules, components and internal states are nested hierarchically and bonds are shown as edges between components.
In conventional rule visualization, as shown in Fig 2C, each reactant and product pattern is drawn separately as a site graph. To distinguish reaction center and reaction context, e.g., to identify that rule rule R2 transforms the internal state of p1 and requires the sub-enz bond, the viewer has to visually compare the graphs from each side of the rule. This imposes a high mental load for complex rules, especially when a large amount of context obscures a much smaller reaction center.
In this work, we introduce compact rule visualization (Fig 2D), which does not require visual graph comparison and avoids drawing reaction context twice. We describe its derivation in S1 Appendix. Briefly, we identify and merge structures common to both sides of the rule, then use special nodes called graph operation nodes to represent the modifications performed. The directions of edges on the graph operation node indicate whether a structure is consumed or produced by that operation. In Fig 2D, each rule is shown with the respective operation node, namely AddBond (R1), DeleteBond (_reverse_R1), and ChangeState (R2, R3) respectively. BioNetGen also supports creating and deleting molecules (AddMol, DeleteMol) and multiple operations per rule (S1 Fig).
To interpret compact rule visualization, the viewer looks for graph operation nodes, which are visually distinguishable from molecule, component and internal state nodes. The structures adjacent to the graph operation nodes constitute the reaction center, whereas the remaining structures constitute reaction context.
In this work, we introduce atoms and atom-rule graphs, which enable visualizing the regulatory architecture represented by a set of reaction rules.
Atoms are elementary structural features found in patterns. In Fig 3A, using BioNetGen syntax as well as site graph visuals, we show instances of various types of atoms present in the product pattern of rule R2. They include:
The Atom-Rule (AR) graph indicates the relationship of a rule with various atoms, which can be reactant, product and/or context. We describe its derivation in detail in S1 Appendix. Briefly, a reactant or product edge is drawn if an instance of the atom is present in the reaction center, on the left or right side of the rule respectively. A context edge is drawn if an instance is present in the reaction context. In Fig 3D, we show AR graphs derived from the rules in Fig 2C, with atomic node labels in BioNetGen syntax. For convenience, the molecule atoms are omitted if there are no molecules added or deleted in the rule.
To interpret the AR graph, one views each atom as a class of actionable sites present in the model. For example, Sub(p1~Y) represents the class of unphosphorylated states on p1 components that can potentially be acted upon by phosphorylation mechanisms. Then, one interprets each edge as an interaction between a mechanism and a class of sites. A reactant or product edge respectively indicates that a mechanism has a consumption or production effect on that particular class of sites. A context edge indicates that the mechanism requires that particular class of sites as a local condition. For example, from the AR graph of rule R2 in Fig 3B, we infer that R2 consumes unphosphorylated p1, produces phosphorylated p1, and requires that p1 be unbound and that enzyme be bound to substrate.
The model AR graph, as in Fig 3C, is a bipartite graph between rules and atoms that is constructed by merging AR graphs of individual rules. Paths on the model AR graph that alternate between rules and atoms represent signal flows. A particular set of rules will always produce the same AR graph, which is a complete representation of signal flow in that rule set between atoms and rules.
To build compact pathway diagrams that convey function, we provide a pipeline for reducing the complexity of the model AR graph (Fig 4A) while preserving relevant regulatory features. Briefly, it involves:
The output of this pipeline is the compressed model AR graph.
To decide which atoms and rules to remove (Step 1) as well as which atoms belong together as groups (Step 2), we take a semi-automated approach. An automated heuristic grounded in commonly encountered biological scenarios makes a first pass through the full AR graph and outputs a template file containing the choices made by the heuristic. To account for nuances of individual systems, the user can edit this template to make alternate choices and import it back into the visualization tool (see tutorial in S2 Appendix for a demonstration). Following this, an automated procedure examines each rule on the graph, the edges incident on the rule and the atom groups adjacent to the rule, then groups rules that share the same edge signature (Step 3). Currently, we support two types of edge signature: strict, which examines all three edge types, and permissive, which examines only reactant and product edges. Finally, an automated procedure replaces each group of nodes with a single representative node (Step 4). Edges incident on individual nodes are merged onto the representative node. A particular set of pipeline inputs (edge signature, template) will generate the same compressed AR graph, but these inputs can be tuned to produce different compressed AR graphs.
Each step in the pipeline has a specific interpretation. Atoms and rules that are removed represent structures and mechanisms with low functional priority, which are typically free binding sites, unphosphorylated states, unbinding rules and dephosphorylation rules. Atom groups represent functional categories of biological structures, e.g., the set of phosphorylation sites on a receptor. Rule groups represent categories of similarly acting mechanisms, e.g., phosphorylation mechanisms active at a particular group of sites. Merging groups is equivalent to reducing the resolution of the graph from individual sites and processes to broad categories of those elements. Permissive grouping also introduces a weaker semantic for the context edge on the compressed graph: a merged group node with a context edge implies that at least one of its members prior to merging had the same context edge.
An implementation of the methods described here is freely available as part of the open source BioNetGen distribution at http://bionetgen.org. A typical procedure involves calling a “visualize()” method from the BioNetGen model file with arguments for user input as well as a template file with edits, if applicable. The default template file can also be automatically generated as a text file. The typical visualization output is a file in GML format (graph modeling language) [41,42] encoding nodes, node labels, edges, edge directions and style attributes of nodes and edges such as color and shape. To lay out the graph, i.e., assign specific coordinates to nodes, we recommend using a third party application such as the yEd graph editor (http://yworks.com/yed), which was also used for the graphs shown in this paper. The tutorial in S2 Appendix provides a detailed walkthrough of the visualization tools using the model from Suderman and Deeds [40] as an example.
We compiled a list of 27 rule-based models from the literature, which we list in S1 Table and attach in S1 Dataset. The models had 2239 rules in total, with the number of rules per model ranging from 6 to 625. We applied to these models a suite of nine visualization tools: contact map, conventional rule visualization, compact rule visualization, Simmune Network Viewer, rule influence diagram and atom-rule graphs at various steps in the complexity reduction pipeline: full model AR graph, AR graph with background removed, AR graph compressed using a strict edge signature, and AR graph compressed using a permissive edge signature. The compression pipeline was applied automatically by making default choices for prioritizing and grouping nodes. On the output graphs, we computed number of nodes (n) and number of edges per node (e/n), counting hierarchical relationships between nodes also as edges. We present these statistics in the Results section.
Pseudocode for the algorithms underlying the tools as well as a detailed accounting of computational costs is available in S1 Appendix. Briefly, for compact rule visualization, the rate-limiting step is building a correspondence map between left and right sides of the rule. Given a maximum finite rule size, the cost can be considered as O(1) per rule. Examining the rule with the correspondence map to synthesize the AR graph is also O(1) per rule. Merging AR graphs of individual rules, grouping rules and merging groups are all O(n), where n is the number of rules.
Visualizing individual rules promotes understanding the structural and kinetic assumptions encoded in a model. Unlike conventional rule diagrams, which require visually comparing reactant and product sides of a rule, compact rule visualization explicitly indicates which modification is performed on which set of structures. Specifically, it allows us to distinguish reaction center, the site of action of a rule, from reaction context, the structural requirements that need to be matched for the rule to fire. In Fig 5A, we show compact rule visualizations of four reaction rules from the immunoreceptor signaling model of Fig 1. Rules R3 and R6 have AddBond operations and represent two distinct binding modes of Lyn kinase to the β domain of FcεRI receptor. In R3, the U domain of Lyn binds the unphosphorylated β domain (constitutive binding), whereas in R6, the SH2 domain of Lyn binds the phosphorylated β domain (activated binding). Rules R4 and R7 have ChangeState operations and represent phosphorylation of the β domain in receptor dimers, with the active kinase being Lyn recruited through constitutive and activated modes respectively.
To understand a model, it is important to know how rules interact with each other and whether they form common motifs such as feedback or feedforward loops. For example, the rules in Fig 5A constitute a positive feedback loop: phosphorylation of β domain (R4, R7) activates Lyn binding (R6), which in turn promotes β phosphorylation (R7), but this is not obvious from conventional and compact rule visualizations. Current methods identify regulatory interactions between pairs of rules through graph comparison [21], simulation [6,22], or manual interpretation [23]. In contrast, the atom-rule graph, which is a bipartite graph showing regulatory interactions between rules and elementary structural features called atoms (see Methods), is constructed efficiently by examining each rule’s reaction center and reaction context. In Fig 5B, we show an AR graph constructed from rules R3, R4, R6 and R7, and the feedback loop is visible as a path on this graph.
The model AR graph for the full immunoreceptor model (Fig 6A) is a complete representation of signal flow in the model, encompassing all 24 rules. The compression pipeline (described in Methods) extracts the essential features of signal flow from the model AR graph and displays them as a compact pathway diagram. The steps of the pipeline, which we apply to the model AR graph in Fig 6A, include:
In Step 1, we remove unphosphorylated states, dissociation rules and dephosphorylation rules from Fig 6A, producing the graph in Fig 6B. In Step 2, we group bonds that link the same molecules (Lig|Rec, Lyn|Rec, Rec|Syk) and phosphorylation sites on molecules (Rec_pY, Syk_pY), producing the atom groups shown in S2A Fig. In Step 3, grouping rules that share similar reaction centers and contexts produces the rule groups shown in S2B Fig, whereas dropping the context similarity requirement produces the more inclusive rule groups shown in S2C Fig. In Step 4, merging groups shown in S2B and S2C Fig produces the compressed AR graphs in Fig 6C and 6D respectively.
Unlike the full AR graph, the compressed graphs are compact and easier to understand. It is also easier to trace specific signal flows on the compressed graphs, such as the feedback between Lyn-receptor binding and receptor phosphorylation (edges marked x in Fig 6A–6D). Under default settings, the whole pipeline is automated, but the resolution of the compressed graphs and the quality of the output diagram can be tuned by providing user input, which includes customizing the heuristics for Steps 1 & 2 and choosing the grouping strategy for Step 3. The strict grouping used in Fig 6C resolves three variants of Syk phosphorylation under various contexts (nodes 1–3) and constitutive and phospho-activated Lyn|Rec binding modes (nodes 4–5), whereas the permissive grouping in Fig 6D merges variants of the same process and represents them with a single node (nodes 6,7). A specific set of pipeline inputs reproducibly generates the same compressed graph from the model and serves as diagram documentation.
To test the scaling of our approach to the growing set of large rule-based models [10,12,14,16], we applied the AR graph compression pipeline to two extensive models of receptor signaling: the FcεRI rule library constructed by Chylek et al. [10] (17 molecule types, 178 rules), and the ErbB signaling model constructed by Creamer et al. [14] (19 molecule types, 625 rules). The compressed graphs for these libraries are shown in Figs 7 and 8 respectively. Unlike the manually constructed Extended Contact Maps (ECMs) [23] that were published with these models, the graphs we show are pathway diagrams that were generated directly from the model specification. In S2 Appendix, we provide a tutorial on generating similar diagrams using the yeast pheromone signaling model of Suderman and Deeds [40] (26 molecule types, 272 rules) as an example.
The modeler can customize pipeline inputs to capture specific biochemical features in the model as well as strike a balance between compression and resolution on the output graph. For example, the default heuristic assumes that co-occurring phosphorylation sites can be grouped together, but for the FcεRI model, we wanted to distinguish between co-occurring phosphorylation sites with opposing functions, specifically those on Src family kinases Lyn and Fyn (SFKs). So, during atom grouping, we grouped functionally similar sites across molecules, e.g., the group SFK_Act_p contains activation-related phosphorylation sites on both Lyn and Fyn. As a result, the output graph (Fig 7) resolves the regulatory interactions of a generic SFK rather than Lyn and Fyn individually. Similarly, for the much larger ErBb model, creating functional groups such as ligands, receptors, and receptor dimers caused a dramatic reduction in complexity, with the output graph (Fig 8) showing signaling interactions of a generic ErbB receptor. Alternatively, grouping Lyn sites separately from Fyn or EGFR and ErbB2 receptors separately from ErbB3 and ErbB4 will produce graphs larger than those shown in Figs 7 & 8, with regulatory interactions resolved in more detail.
The compressed AR graph offers a convenient venue for analysis and exploration of a rule-based model. For example, on the FcεRI and ErbB graphs, we were able to identify well-known pathways such as MAPK (transparent overlays in Figs 7 & 8) using a combination of node clustering and visual inspection. Also, on the FcεRI graph, we were able to trace network motifs encoded in the model (Fig 9) by Chylek et al. [10]. Without the compressed AR graph, the same analyses would have required examining hundreds of complex rules in various combinations, which would have required significant effort. Thus, the compressed AR graph offers a useful proxy for the rule-based model that is more amenable to analysis.
To assess the readability of various visualization tools, we examined the joint distribution of graph size n and edge density e/n for each visualization when applied to 27 published rule-based models (see Methods), where n and e refer to number of nodes and edges respectively. In S3 Fig, we report these distributions for 9 visualization methods, and in Fig 10, we show their geometric means. The choice of metrics follows from Ghoniem et al. [43], who determined that user performance on visual graph analysis tasks decays with increasing graph size and edge density. Ghoniem et al. used much denser graphs than the ones in our test set, so we replaced their edge density metric √(e/n2) with e/n, which has a higher coefficient of variation for the graphs in our test set (2.54 vs 1.03), and therefore higher discriminatory power.
The results in Fig 10 confirm our qualitative observations on the readability of current visualizations and the improvements present in our new ones. Contact maps are generally compact with sparse edges as they only show structural composition and do not show individual mechanisms or signal flow. Rule visualizations, both conventional and compact, produce large graphs with sparse edges as they show the patterns encoded in each rule. However, compact rule visualizations are smaller than conventional ones as they make use of graph operation nodes. Diagrams showing interactions of rules are typically dense, such as rule influence diagrams and full AR graphs. However, full AR graphs have much lower edge density than rule influence diagrams as they use atoms to mediate interactions between rules. When the compression pipeline is applied, AR graphs’ size and edge density can be reduced to approach that of contact maps. This makes compressed AR graphs as readable as contact maps, while conveying substantially more information about the signaling architecture. The Simmune Network Viewer, which is intermediate between rule visualizations and full AR graphs, is discussed in detail below.
In this work we have developed new visualization approaches for rule-based models. The novel compact rule visualization conveys the mechanism underlying individual rules more effectively than conventional visualizations. The atom-rule (AR) graph conveys interactions between rules more efficiently than rule influence diagrams. A compression pipeline for the AR graph flexibly accounts for nuances of specific biological systems and reproducibly generates compact pathway diagrams even for models with hundreds of complex rules. These tools open the door for new forms of analysis for rule-based models such as network motif identification. In Supplementary Material, we show the theoretical foundation for these tools (S1 Appendix) as well as a tutorial for how to apply them to a large rule-based model (S2 Appendix).
Edward R. Tufte, a pioneer of modern data visualization and analytic design, argues that “universal cognitive tasks” underlie how humans perceive information and motivates that “cognitive tasks should be turned into design principles” [44]. In the biochemical literature, diagrams and text employ a number of such cognitive tasks, and our automated methods recapitulate some of these. For example, one often describes a biochemical process using an action verb such as “binds” or “phosphorylates”. Graph operation nodes in compact rule visualization (Fig 2D) play a similar role in conveying the action of a rule. Similarly, one uses “site” to denote a molecular part that behaves distinctly or is targeted by a specific process. Atoms used in the atom-rule graph (Fig 3A) have a similar interpretation as types of actionable sites. Literature descriptions and diagrams also selectively emphasize active states over ground states and signal-activated processes over processes that attenuate the signal or occur in the background, which allows the reader to filter redundant information. Removing low priority nodes on the AR graph follows a similar principle (Fig 4B). Text descriptions routinely categorize molecules and sites using principles such as homology and functional similarity [45–48], and use broad terms to summarize information about specific molecules and sites. Grouping atoms and rules using the described heuristics (Fig 4C–4D) and compressing the AR graph (Fig 4E) recapitulates this approach.
Whenever compression is applied to data, there exists a many-to-one relationship between the uncompressed and compressed representations. In the context of visualization, a rule-based model will generate the same conventional and compact rule visualizations and vice versa, but different models can generate the same contact map, rule influence diagram and AR graph. Therefore, one should use each tool at the resolution for which it is designed to be used. Compact rule visualization should be used to show the mechanism underlying each rule. The AR graph is less useful for this purpose, as it approximates each rule as a bipartite graph. Instead, it should be used to infer interactions between rules through formal or informal approaches. When applying the compression pipeline to the AR graph, one should verify that the choice of inputs is biologically reasonable. If this is the case, then the compressed AR graph is useful for both communicating the model to others as well as graph analysis.
In addition to the approaches discussed in Introduction (Fig 1A–1D) and Methods (Fig 2C), we show examples of other currently available tools (Fig 11) and how they compare with compact rule visualizations and atom-rule graphs.
The SBGN Process Description (Fig 11A) [24] is a visualization standard for reacting entities. It has the same limitation as conventional rule visualization, namely the need for visual graph comparison.
The Kappa story (Fig 11B) [22] shows the causal order in which rules can be applied to generate specific outputs, and these are derived by analysis of model simulation trajectories. It is complementary to the statically derived AR graph for showing interactions between rules, but it does not show the structures that mediate these interactions nor does it provide a mechanism for grouping rules. Integrating Kappa stories with AR graphs is an interesting area for future work.
The Simmune Network Viewer (Fig 11C) [26] compresses the representation of rules differently from the AR graph: it merges patterns that have the same molecules and bonds, but differ in internal states. Like the AR graph, it shows both structures and rules, and it produces diagrams with much lower density (‘sim’ in Fig 10), but it obscures causal dependencies on internal states (S4 Fig).
The SBGN Entity Relationship diagram (Fig 11D) [24] and the Molecular Interaction Map (Fig 11E) [25], like the Extended Contact Map [23], are diagrams of model architecture that rely on manual analysis.
The rxncon regulatory graph (Fig 11F) visualizes the rxncon model format [27], which uses atoms (called elemental states in rxncon) to specify contextual influences on processes. This approach, which is also followed in Process Interaction Model[49], is less expressive than the graph transformation approach used in BioNetGen, Kappa and Simmune (S5 Fig). The AR graph we have developed generalizes the regulatory graph visualization so it can be derived from arbitrary types of rules found in BioNetGen, Kappa and Simmune models.
The AR graph offers many advantages over existing methods, but there are a number of ways in which it could be improved or generalized. There are alternate ways to show the content of the AR graph, for example, as a two-dimensional matrix [43]. The compression algorithms can be extended to identify more complex relationships, for example, treating the consumption of an active state as an ‘inhibits’ relationship, grouping enzyme-binding and catalytic processes together as a Michaelis-Menten mechanism, etc. In the immediate future, we plan to add support for other features present in the BioNetGen model specification, such as compartmental states, transport rules and dependencies encoded in rate laws [50,51].
Additionally, the AR graph opens up rule-based models to a wide variety of analysis and visualization tools, as it transforms a complex rule-based model into a simple bipartite graph. For example, simulation fluxes can be conveniently visualized on a bipartite graph by mapping numeric values to node size or edge thickness [52]. Also, as rxncon developers have shown, one can perform stochastic Boolean simulations on a bipartite graph [53]. Model reduction approaches developed for rule-based models have previously used information on interactions between structures and rules [54] that can now be obtained directly from the AR graph. The AR graph also serves as a rich source of information that could be mined using formal approaches. Potential areas where new methods can be developed include identifying model subsystems (as in Figs 7 and 8) by graph partitioning [55], identifying network motifs (as in Fig 9) by cycle detection [56], dynamically grouping atoms and rules using graph structure discovery [57,58], etc. Thus, adoption of the AR graph could pave the way for novel applications of graph analysis, data mining and machine learning to rule-based models.
A natural future direction for signaling models is to explore the effects of complex input stimuli and crosstalk between pathways [59,60] on a comprehensive scale. This would require integrating rules from multiple sources, such as databases constructed in tandem by different groups (e.g. [10,12,14,34]). The recently published whole cell model of Mycoplasma genitalium [17] makes effective use of databases to organize and visualize kinetic information [61–63] and provides proof-of-concept of a database-oriented approach. Currently, models of signaling from various receptors have as many as hundreds of rules [10,12,14] and this number is expected to increase by an order of magnitude to cover more molecule types, receptors and signal pathways. We expect that AR graphs will play a role in the construction, navigation and visualization of the rule-based databases of the future, similar to approaches deployed on other biological data (VisANT [64], ChiBE [65]). The AR graph will also be useful for frameworks that implement rule-based data structures (SBML-Multi [33], BioPax Level 3 [66]) or integrate rules with higher-order model composition (Virtual Cell [18,19], PySB [34]). Thus, in addition to the immediate benefit of visualizing and understanding large models, the AR graph is expected to be useful in developing the comprehensive cell models of the future.
|
10.1371/journal.pgen.1000344 | Organised Genome Dynamics in the Escherichia coli Species Results in Highly Diverse Adaptive Paths | The Escherichia coli species represents one of the best-studied model organisms, but also encompasses a variety of commensal and pathogenic strains that diversify by high rates of genetic change. We uniformly (re-) annotated the genomes of 20 commensal and pathogenic E. coli strains and one strain of E. fergusonii (the closest E. coli related species), including seven that we sequenced to completion. Within the ∼18,000 families of orthologous genes, we found ∼2,000 common to all strains. Although recombination rates are much higher than mutation rates, we show, both theoretically and using phylogenetic inference, that this does not obscure the phylogenetic signal, which places the B2 phylogenetic group and one group D strain at the basal position. Based on this phylogeny, we inferred past evolutionary events of gain and loss of genes, identifying functional classes under opposite selection pressures. We found an important adaptive role for metabolism diversification within group B2 and Shigella strains, but identified few or no extraintestinal virulence-specific genes, which could render difficult the development of a vaccine against extraintestinal infections. Genome flux in E. coli is confined to a small number of conserved positions in the chromosome, which most often are not associated with integrases or tRNA genes. Core genes flanking some of these regions show higher rates of recombination, suggesting that a gene, once acquired by a strain, spreads within the species by homologous recombination at the flanking genes. Finally, the genome's long-scale structure of recombination indicates lower recombination rates, but not higher mutation rates, at the terminus of replication. The ensuing effect of background selection and biased gene conversion may thus explain why this region is A+T-rich and shows high sequence divergence but low sequence polymorphism. Overall, despite a very high gene flow, genes co-exist in an organised genome.
| Although abundant knowledge has been accumulated regarding the E. coli laboratory strain K-12, little is known about the evolutionary trajectories that have driven the high diversity observed among natural isolates of the species, which encompass both commensal and highly virulent intestinal and extraintestinal pathogenic strains. We have annotated or re-annotated the genomes of 20 commensal and pathogenic E. coli strains and one strain of E. fergusonii (the closest E. coli related species), including seven that we sequenced to completion. Although recombination rates are much higher than mutation rates, we were able to reconstruct a robust phylogeny based on the ∼2,000 genes common to all strains. Based on this phylogeny, we established the evolutionary scenario of gains and losses of thousands of specific genes, identifying functional classes under opposite selection pressures. This genome flux is confined to very few positions in the chromosome, which are the same for every genome. Notably, we identified few or no extraintestinal virulence-specific genes. We also defined a long-scale structure of recombination in the genome with lower recombination rates at the terminus of replication. These findings demonstrate that, despite a very high gene flow, genes can co-exist in an organised genome.
| Escherichia coli was brought into laboratories almost a century ago to become one of the most important model organisms and by far the best-studied prokaryote. Major findings in phage genetics, bacterial conjugation, recombination, genetic regulation and chromosome replication involved the use of E. coli, especially laboratory derivatives of the K-12 strain, originally isolated from the faeces of a convalescent diphtheria patient in Palo Alto in 1922 [1]. However, K-12 derivatives are far from representing the whole E. coli species [2]. The primary habitat of E. coli is the lower intestinal tract of humans and other vertebrates, with which it typically establishes commensal associations. Healthy humans typically carry more than a billion E. coli cells in their intestine. It has been estimated that half of the living E. coli cells are outside their host, in their secondary habitat [3]. Beside these habitats, certain strains have the potential to cause a wide spectrum of intestinal and extra-intestinal diseases such as urinary tract infection, septicaemia, meningitis, and pneumonia in humans and animals [4]. Furthermore, Shigella, which have been elevated to the genus order with four species (dysenteriae, flexneri, boydii, sonnei) based on their capacity to generate a specific mucosal invasive diarrhoea strictly in humans and their biochemical characteristics, in fact belong to the E. coli species [5]–[7]. Of note, Shigella and enteroinvasive E. coli are considered the only obligate pathogens of the species, whereas other strains are facultative pathogens with a broad host range. Thus, natural isolates of E. coli/Shigella live in conditions quite different from those in the laboratory and must cope with very diverse environments that provide stresses ranging from immune system attack and protozoal grazing to starvation, low temperatures, and, more recently, antibiotic therapy.
With its large range of pathologies, E. coli is a major cause of human morbidity and mortality around the world. Each year E. coli causes more than two million deaths due to infant diarrhoea [8],[9] and extraintestinal infections (mainly septicaemia derived from urinary tract infection) [10], and is also responsible for approximately 150 million cases of uncomplicated cystitis [10]. Since humans and food animals carry so many E. coli cells that may establish commensal or antagonistic interactions with their hosts it is mandatory to define the genetic and population determinants that drive commensal strains to adopt a pathogenic behaviour.
Population genetic studies based on both multi-locus enzyme electrophoresis [11]–[13] and various DNA markers [14]–[18] have identified four major phylogenetic groups (A, B1, D and B2) and a potential fifth group (E) among E. coli strains. Strains of these groups differ in their phenotypic characteristics, including the ability to use certain sugars, antibiotic resistance profiles and growth rate–temperature relationships [19]. The distribution (presence/absence) of a range of virulence factors thought to be involved in the ability of a strain to cause diverse diseases also varies among strains of these phylogenetic groups [20]–[22], indicating a role of the genetic background in the expression of virulence [23]. Consequently, these groups are differently associated with certain ecological niches, life-history characteristics and propensity to cause disease. For example, group B2 and D strains are less frequently isolated from the environment [24], but more frequently recovered from extra-intestinal body sites [23]. While B2 strains represent 30 to 50% of the strains isolated from the faeces of healthy humans living in industrialised countries, they account for less than 5% in French Guyana Amerindians [25]–[26].
The clear clustering of E. coli strains into monophyletically meaningful groups has long been used as an argument favouring clonality within the species. However, analysis of gene sequences shows pervasive recombination, matching the well-known efficiency of conjugation and transduction of the species [17],[27]. Hence, it remains controversial whether such frequent recombination obliterates the phylogenetic signal. E. coli genomes show evidence of widespread acquisition of functions by horizontal gene transfer, concomitant with similar amounts of gene deletion [28]–[29]. While less than 3% of nucleotide divergence is found among conserved genes, the gene content between pairs of E. coli genomes may diverge by more than 30% [30]. Such diversification of gene content due to horizontal gene transfer contributes greatly to the diversity of the strains' phenotypes and can be accurately quantified only by the sequencing of a large number of strains to completion and closure.
Until now, sequencing efforts in E. coli have been focused mainly on pathogenic strains, particularly on diarrhoeal and group B2 extraintestinal pathogenic strains (see Table 1), precluding an unbiased assessment of the diversity of the species. Therefore, we have sequenced with high coverage and up to completion the genomes of 6 human-source E. coli strains. The E. coli strains were chosen to complement the available sequences and other ongoing sequencing projects (http://msc.jcvi.org/e_coli_and_shigella/index.shtml, http://www.sanger.ac.uk/Projects/Escherichia_Shigella/). They encompass two commensal strains of phylogenetic groups B1 and B2, a group B1 enteroaggregative strain, two group D urinary tract infection strains and a group B2 newborn meningitis strain (Table 1). We also sequenced the type strain of the closest E. coli relative, i.e., E. fergusonii [31], as an outgroup to permit accurate and meaningful evolutionary analyses with the 6 new E. coli genomes and the 14 other currently available E. coli/Shigella genomes. To statistically substantiate the identification of extraintestinal virulence-associated genes, we also applied a mouse lethality assay to the strains [32] to quantify the intrinsic virulence of the strain, excluding host variability and other potential confounding factors (Table 1). Our goal was to take the outstanding opportunity provided by the availability of many genomes of a single bacterial species, regarding which a considerable amount of knowledge has been accumulated over the years, to answer to the following questions. (i) Is there genome-wide evidence of frequent recombination and does it vary with genome location? (ii) If so, can one nonetheless infer an intra-specific bacterial phylogeny? (iii) How do the different factors of genome dynamics (mutation, horizontal gene transfer with or without recombination) result together in strain diversification? (iv) Is genome dynamics in conflict with genome organisation? (v) How does the commensalism/pathogenicity duality evolve?
We fully sequenced the chromosomes and the plasmids, if any, of 6 strains of E. coli and the reference type strain of E. fergusonii. The general features of these replicons are listed in Tables 2 and 3. Genomes were sequenced at an average of 12-fold coverage and were then finished. The 6 newly sequenced E. coli chromosomes contain between 4.7 Mb and 5.2 Mb each, corresponding to between 4627 and 5129 protein coding genes, slightly above the average value within the 20 genomes that we analyzed (∼4700 genes, ranging from 4068 to 5379). The chromosome of E. fergusonii is slightly smaller with ∼4.6 Mb and ∼4500 protein coding genes. The G+C content is very similar among the 6 strains and close to the E. coli K-12 MG1655 value (∼50.8%). The G+C content of E. fergusonii is lower at 49.9%. These chromosomes have similar densities of coding genes and numbers of stable RNA genes. By contrast, the number of pseudogenes varies more widely, from 22 in E. fergusonii to 95 in strain ED1a (Table 2). The list of pseudogenes is available in Table S1.
The variation in the number of pseudogenes is uncorrelated with the number of transposable elements and phage-associated genes, which vary in the range 42–224 and 201–517 respectively. While some phage-associated genes are scattered throughout the chromosomes, the majority are concentrated in well-defined prophage regions. Analyses of the prophages suggest that many may still be functional. These prophages often carry at their extremity some unrelated cargo genes that probably arose from genomes of previously infected bacteria, as found in Salmonella [33]. We sequenced a total of 6 plasmids, varying in size from 34 to 134 kbp: four strains possess one plasmid each whereas one strain has 2 plasmids (Table 3). As frequently noted, the plasmids have a lower gene density (84%, vs. 87% for chromosomes), lower G+C content (47.4%, vs. 50.7% for chromosomes) and more pseudogenes (2.7%, vs. 1.5% for chromosomes). The percentage of orphan proteins (i.e., having no detectable homolog in other organisms) is also high on plasmids (6.5 to 52.2%), while it ranges between 1–3% on the chromosomes.
A manual expert annotation of the new E. coli strains was performed on genes and regions not found in E. coli K-12 MG1655 (about 10 000 genes in total; Table S2A). This allowed the re-annotation of orthologs in the previously available Escherichia and Shigella genomes (see Materials and Methods). The annotation data, together with the results of the comparative analysis were stored in a relational database called ColiScope, which is publicly available using the MaGe Web-based interface at http://www.genoscope.cns.fr/agc/mage. This re-annotation process revealed extensive variations in the number of the newly predicted genes (Table S2B). For example, between the two strains of E. coli O157:H7 we found twice as many newly predicted genes in one strain as in the other. In some genomes important genes were missing. For example, in E. coli APEC O1 several subunits of the ribosome, DNA polymerase III, and ATP synthase were missing in the original annotation (Table S3, E. coli APEC sheet). In other genomes, the re-annotation allowed us to standardise the definition and identification of pseudogenes. For example, in S. sonnei Ss 046 most of the newly annotated genes correspond to insertion sequences (ISs) and small fragments of incompletely annotated pseudogenes (Table S3, S. sonnei sheet). As a result of this effort, the present ColiScope database contains a complete and consistent set of annotations for the 7 newly sequenced genomes and the 14 available Escherichia and Shigella genomes. These data were the starting point of the work presented here.
We analyzed gene order conservation within the 21 genomes (Table S4). More than half of the genomes have exactly the gene order of E. coli K-12 MG1655, which we inferred as ancestral. Thus, the organisation of the core genome is stable in most strains. Three genomes show 1 or 2 rearrangements. Seven genomes show more than 10 blocks of synteny: 6 of these genomes are from Shigella, the high rearrangement rates of which resulted in up to 65 blocks of synteny in S. dysenteriae. These genomes have a large number of ISs, ranging from 549 to 1155 in S. flexneri and S. dysenteriae, respectively, which are well known to shuffle genomes. E. fergusonii also shows a large number of rearrangements relative to the ancestral organization of the E. coli genome. Since the organisation of some strains of the more distantly related Salmonella enterica closely resembles that of E. coli K-12 MG1655, many rearrangements must have taken place in the branch leading to E. fergusonii.
Figure S1 provides the classical concentric circle representation for the 7 genomes we sequenced, showing GC skews, G+C variation, and a description of the presence of genes in ever-increasing clades within the genus, relative to the inferred ancestral genome. The first position of the sequences was chosen to match the orthologous region in the E. coli K-12 MG1655 genome and corresponds to the intergenic region between lasT and thrL. Origins and termini of replication were identified by GC skews and homology with the respective E. coli K-12 MG1655 regions. These figures show that divergence from the average G+C content often occurs in genomic regions absent in the other strains. They also reveal the highly mosaic structure of these genomes, comprising the core genes and the accessory genes, which we then set out to quantify.
The analysis of the first E. coli genomes changed our views about the evolution of gene repertoires in bacteria. Genomes within the species vary in size by more than 1 Mb, i.e., by more than 1000 genes, and even the gene repertoires of similarly sized genomes differ widely [30],[34]. We have thus taken advantage of the unprecedented availability of 20 completely sequenced genomes of the same species to analyse the evolution of the gene repertoire. We first identified the core and pan-genomes of E. coli, i.e., the genes present in all genomes and the full set of non-orthologous genes among all genomes. In our data set, the average E. coli genome contains 4721 genes, the core genome contains 1976 genes, and the pan-genome contains 17 838 genes. The random sampling of one gene within a randomly selected E. coli genome has a probability of only ∼42% of revealing a ubiquitous gene. On the other hand, the full sequencing of an E. coli strain allows observation of only one-fourth of the observed pan-genome. This implies that although some fundamental functions can be well studied by using a model strain, no single strain can be regarded as highly representative of the species.
Further sampling of E. coli genomes is unlikely to change significantly the estimate of the core genome, however, the pan-genome is far from being fully uncovered (Figure 1). Annotation and sequencing artefacts may affect the estimations of core and pan-genome sizes, e.g. by spurious annotation of small genes or pseudogenes. We hope to have minimised such problems by using a coherent set of annotations. Still, we found that 40 genes deemed essential in E. coli K-12 W3110 [35] were missing in the core genome. Among these, 17 correspond to genes with conflicting reports of essentiality, or contextually essential genes such as prophage repressors, and are absent in most genomes. The other 23 genes have orthologs in most genomes and 19 are missing in a single genome where they can be found as pseudogenes interrupted by a single-nucleotide frameshift. While “pseudogenisation” does often start with such frameshifts [36], these genes correspond to core housekeeping functions, so the reported frameshifts probably represent sequencing errors. For example, it is hard to see how S. boydii could replicate without the catalytic α-subunit of the DNA polymerase III or how E. coli 536 could survive without a tyrosine tRNA synthetase. We found some comfort in verifying that none of the 23 genes was absent from the 7 genomes we sequenced. If one assumes that these essential genes cannot be deleted and that no special care has been taken to check for sequencing errors at these loci, then our estimation of the core genome should be increased by a factor of 260/(260-23) to 2167 genes. This still makes the core genome less than half of the average E. coli genome (∼46%). Importantly, no gene of the core genome, nor any operon ubiquitous in E. coli, was unique to the species, i.e., we could always find a homolog in at least one of the other fully sequenced bacterial genomes.
Some elements recently amplified in the genome, such as transposable elements, create multiple copies that are not orthologs sensu strictu, even though they probably have the same function. They will thus inflate the size of the pan-genome by increasing the number of strain-specific genes. We therefore made two complementary analyses. First, we classed together all paralogs with more than 80% sequence similarity. This led to 11 432 genes of a functionally diverse pan-genome (Figure 1). Second, we removed all transposable elements and prophages, but not their cargo genes, from the pan-genome to obtain a set of 10 131 genes. These analyses still lead to a vast pan-genome for the species and show that its large size is not a simple consequence of the presence of selfish genes or recent amplifications of genetic material. They also show that further sampling of E. coli genomes is likely to uncover a significant number of currently unrecognised genes that may confer lasting adaptive value for the diversification of the species.
Progressive sampling of E. coli genomes will tend to reduce the core genome to the list of essential genes because only instantaneously lethal deletions will never be found in natural populations of living cells. Hence, it is more relevant to quantify the relative frequency of each gene of the pan-genome among extant genomes (Figure 2). Of the genes in an average E. coli genome, approximately 62% are present in at least 18 genomes, and thus might be called the persistent genes [37], while 26% exist in 4 or fewer genomes, and thus might be called the volatile genes. Thus, most genes of the pan-genome exist in very few (≤20%) or almost all (≥90%) of the genomes, leaving only a small subset of genes that are present in around half of the genomes. The functional pattern of these groups of genes varies. Genes of known function are strongly over-represented among persistent genes, whereas genes of unknown function and especially selfish DNA, such as transposable and prophage elements, are over-represented among strain-specific (volatile) genes (Figure 2). Although some of these strain-specific genes may confer adaptive functions that allow the exploration of new niches (see below the section on the genome repertoire dynamics), the volatility of this set and the functions thereby over-represented suggest that most such genes are non-adaptive.
We assessed how different was E. fergusonii from the strains of E. coli. We computed the core genome of the 21 genomes (20 E. coli+1 E. fergusonii), which contained 1878 genes. We then made experiments in which we computed the core genome of all combinations of 20 genomes and then added the 21st at the end. We ranked the genomes in terms of which led to the highest decrease in the core genome size. S. dysenteriae (174 genes) led to the greatest reduction in the core genome, followed by E. fergusonii (98 genes). We then repeated the experiment with the pan-genome. In this analysis, we also found that the most contributory 21st genome was S. dysenteriae (1434 genes), followed by E. fergusonii (984 genes). However, this results from the large number of ISs in the former strain. When we computed the pan-genome while merging together paralogs that are more than 80% identical, we found that E. fergusonii ranks first (709 genes), well ahead of the second place strain (E.coli CFT073 with 462 genes). This latter difference matches the phylogenetic distance of E. fergusonii, but the overall analysis shows that crossing the E. coli species barrier does not lead to dramatic changes in the core and pan-genome.
Horizontal transfer of new genes necessarily entails different phylogenies for these genes, but has few implications for the inference of phylogeny in the core genome. However, a considerable fraction of the large amounts of DNA that seemingly enter E. coli cells is expected to arise from consepecifics or closely related species. Such DNA can integrate into the chromosome by homologous recombination and thus lead to allelic replacements that obscure the phylogenetic signal. To address this question, we first estimated the rate of recombination in the genomes, then tested whether such a rate could affect the phylogenetic reconstruction.
Using methods based on the coalescent framework, it is possible to estimate the ratio of recombination to mutation rates, i.e., to compare the probability of a recombination being initiated at a particular nucleotide with the probability of a mutation occurring at that same nucleotide. We analyzed each core gene with LDHat, a coalescent-based estimator of recombination [38], and estimated an average ratio of recombination to mutation close to 1.0 (data not shown). Classical population genetics models, such as the one used in LDHat, assume that recombination occurs through reciprocal exchange of DNA with a single crossover. In prokaryotes, incoming DNA sequences are short and the recombination process is akin to gene conversion, whereby linkage between two close regions may be weaker than between two distant ones if one of the former has engaged in conversion with incoming DNA. Bacterial genetic exchange does not always imply mechanisms strictly analogous to those involved in eukaryotic gene conversion. However, since we are concerned more with the signature of gene conversion in linkage disequilibrium than with the underlying molecular mechanisms, we will use the term gene conversion hereafter to refer generically to bacterial genetic exchanges. We took advantage of the peculiar signature of gene conversion on linkage disequilibrium [39] to estimate the per-base rates of mutation (theta) and gene conversion (Cgc), as well as the average tract length (Lgc) (assuming a geometrical distribution), with Approximate Bayesian Computation method [40],[41] (see model in Materials and Methods). We applied the method to individual genes of the core genome and to 3 kbp sliding windows along the whole genome multiple alignment (see Materials and Methods, Figure S2).
Both analyses provided similar average values, but since the genes differ widely in size, we preferred to use the genome alignment for the rest of the analyses. The average ratio of gene conversion to mutation (Cgc/theta) was 2.47±0.05. The average tract length was very short: 50 bp on average, lower than our previous estimate of 120 bp based on multi-locus sequence typing (MLST) data [42], and lower than expected based on experimental data [43]. Contrary to expectations based on random experiments (see Materials and Methods), we observed a strong negative correlation (Pearson r = −0.55, p<0.001) between the ratio of recombination to mutation and the length of the conversion fragments. This may be explained by the overlap of gene conversion fragments in regions of high rate of exchange, which results in artificially low values of Lgc, lending further support to the existence of high conversion rates in the population. In any case, these tract lengths should not necessarily be equated with the size of incoming DNA fragments.
Our model assumes a homogenous population. However, in the gut of a vertebrate, the most likely neighbour for a cell probably is another cell from the same clone, since mucus provides a structured environment within which sister cells are likely to stay together for some time. Transfers between such closely related strains are less affected by restriction [43] or divergence [44]. Every time such a transfer overlaps with a previous transfer from a distant clone it will effectively remove some trace of recombination and, thus, lead to a lower observed tract length. In spite of such limitations we find that a gene conversion event is twice as likely as a mutation to occur at a given position. Therefore, taking into account the estimated tract length (50 bp), a base is 100 times more likely to be involved in a gene conversion than to be involved in a mutation. This is twice as large as the classical estimate [27].
Is such a rate of gene conversion compatible with a meaningful phylogeny? If we do not consider the specificities of bacterial genetic exchange, the answer is no. The estimates provided under a simple crossing-over model are incompatible with any phylogenetic approach (data not shown). However the answer might be different if one considers that exchange in bacteria results in gene conversion. To test this idea quantitatively, we made coalescent simulations in which we used the parameters estimated previously (theta = 0.014 and Lgc = 50) and various rates of gene conversion to mutation (100 experiments for each value) to simulate the evolution of 25 kbp sequences (see Materials and Methods). We then compared the tree inferred by maximum likelihood with the tree derived directly from the simulated history, which reflects the history of the chromosomal background. We compared the tree topologies with Robinson and Foulds distances [45] and the SH, KH and ELW tests (see Materials and Methods). The average distance between the topologies of the pair of trees only starts to increase for gene conversion to mutation ratios (Cgc/theta) much higher than the observed value (Figure 3). Hence, surprisingly, the substantial level of gene conversion in E. coli is not expected to blur the phylogenetic signal, and a meaningful and robust tree topology can be extracted from the sequences.
The foregoing analysis suggests that phylogenetic approaches can be used to analyse genome evolution even within highly non-clonal prokaryotic species. We therefore characterised the phylogenetic relationships among the 20 fully sequenced strains and the outgroup, using a maximum likelihood approach on all 1878 genes of the Escherichia core genome (i.e., the genes present in all 20 E. coli/Shigella and E. fergusonii), either independently or concatenated (1 769 508 nt, 88 883 informative sites). The same analysis was also performed on the chromosomal backbone using the E. coli/Shigella multiple genome alignment (2 672 618 nt, 115 435 informative sites) that, in addition, integrates non-coding sequences and pseudogenes.
Using the concatenated genes of the core genome and a maximum likelihood approach, regardless of the method used to estimate a model (see Materials and Methods) we obtained a robust phylogeny with very high bootstrap values (Figure 4). When each of the 1878 individual gene phylogenies is compared to the concatenated gene phylogeny using various tree topology comparison tests (see Materials and Methods), about 25% are not significantly different from the concatenated gene tree. (It is worth noting that these tests are very stringent, as tree topologies differing by a single strain position can be significantly different.) Similarly, when the “consensus strength” of a node is defined as the percentage of genes that supports the bipartition at a specific node using CONSENSE, it can be shown that nodal consensus strength varies greatly, from 11% to 90% (Figure 4). However, in both approaches (tree topology comparison tests and consensus strength), the low values are largely due to an absence of phylogenetic signal differentiating the strains rather than to conflicting phylogenies, as 55% of genes have fewer than 40 informative sites (data not shown). All the classical groups described by multi-locus enzyme electrophoresis [13] and retrieved later on by genetic markers [14]–[18] are recovered as monophyletic apart from group D. The monophyly of group D in previous MLST studies never appeared to be very robust [16],[17],[46] and was presumably due to long-branch attraction. One D strain (IAI39) is closely related to the group B2 strains and belongs to the ECOR 35, 40, 41 subgroup [16],[46], whereas the other (UMN026), which belongs to the ECOR 46, 47, 49, 50 subgroup [16],[46], has emerged later. Our analysis retrieves the previously reported polyphyly of Shigella [6],[7]. Identical data were observed when using the multiple genome alignment (Figure S3), thus confirming the robustness of the phylogeny. A controversy has emerged about the more basal group within the E. coli species phylogeny, which some authors maintain is group B2 [16], [47]–[49] whereas others remain unconvinced [17],[46]. Our large data set using the closely related E. fergusonii as an outgroup, and thus avoiding the long-branch attraction artefact caused by the inclusion of Salmonella in some previous works, clearly shows that the first split in the E. coli/Shigella phylogenetic history leads on one hand to the strains of group B2 and a subgroup within group D, and on the other hand, to the remaining strains of the species. Groups A and B1, as well as the S1, S3 and SS Shigella groups, have emerged more recently (Figure 4).
Since lateral transfer is extensive in E. coli, we investigated how well gene repertoire relatedness fades with increasing evolutionary distance. We defined gene repertoire relatedness between two genomes as the fraction of shared orthologs in the smallest genome [50], and obtained the evolutionary distance from the phylogenetic tree in Figure 4. We found a negative association between the relatedness of gene repertoires and phylogenetic distance (Figure 5, R2 = 0.26, p<0.001). For very closely related genomes the association is quite clear (Spearman's ρ = −0.70, p<0.001, for the 12% closest comparisons corresponding to 2 of the 6 histogram bins of Figure S4). However, the more distant comparisons show much weaker association between relatedness and divergence time (Spearman's ρ = −0.30, p<0.001). Therefore, the number of shared orthologs is a poor phylogenetic marker and only among the most closely related genomes is there a high degree of similarity according to the repertoire of non-core genes of the pan-genome. This rapid saturation of phylogenetic signal in terms of gene repertoire relatedness might seem surprising in light of the ∼2000 genes shared among all genomes. Yet, if most gene deletions correspond to recent insertions, as we suggested previously, then the saturation of the phylogenetic signal results from the very small number of ancient acquisitions that are maintained among distant genomes. This effect is further enhanced by the frequent re-acquisition of some gene families such as phage and IS-associated genes. As a result, variance in gene repertoire relatedness increases quickly with phylogenetic distance to such an extent that some distantly related genomes actually exhibit greater gene repertoire relatedness than do more closely related ones.
The finding of a strong, reliable phylogenetic tree for the strains allows the inference of gene repertoire dynamics along the history of the species (Figures 6 and 7, Figure S5). We inferred the presence/absence of genes by maximum likelihood using the reference phylogeny at each ancestral node, including the inferred ancestor of all E. coli. We then quantified the flux of incoming and outgoing genes between consecutive nodes of the tree, i.e., at every branch, and inferred the associated change in genome length. There is a difference of almost one thousand genes between the gene repertoire we can infer reliably in the ancestor (4043 genes) and the expected one given the inferred genome length (∼5000). This is because most incoming genes are quickly lost. Anciently acquired volatile genes with no lasting adaptive value have been purged, if not re-acquired later on, whereas recently acquired ones may still persist in populations. Indeed, the gap between expected and inferred gene numbers increases linearly with the distance from the node to the tips of the tree, i.e., with the ancientness of the node (Pearson r = 0.75, p<0.001, Figure S6). Confirming this interpretation, a comparison of genomes separated by a lapse of time equivalent to the distance between the extant genomes and the ancestor, e.g., strains APEC O1 and 55989, shows a number of distinct genes close to the 1000-gene difference observed at the inferred ancestral genome. When accounting for E. coli's speciation process from the other Escherichia spp. it should thus be borne in mind that genes involved in speciation may have disappeared altogether from extant lineages.
To analyse in detail the gains and losses of genes we considered that genes were present at an ancestral node if the probability of presence was higher than 50%, and otherwise were absent. (Variations around this value had little effect of the overall results.) Genes were then classified in 4 mutually exclusive categories: core genome, clade-unspecific (i.e., also present in some genomes not descending from the focal node), clade-specific and present in all descendents from the focal node, or clade-specific but present in only some of the descendents (Figure 5). Most non-core genes are clade-unspecific, especially in nodes close to the root. This is best understood by revisiting Figure 2, which shows that most non-core genes are present in very few genomes. As a result, few genes in the internal nodes are clade-specific and present in all genomes of the clade. The last common ancestor is an exception because it contains many genes present in some E. coli genomes but absent in E. fergusonii. Elsewhere, very few genes are clade-specific, consistent with the idea that most transferred genes quickly disappear from the populations.
Very recent acquisitions are highly enriched in phage-related genes, except in the branches leading to Shigella where transposable elements dominate (Figure 7). Few terminal branches show significant amounts of acquisition of known function genes. The exceptions, UMN026 and IAI39, correspond to the largest terminal branches, which include very ancient and very recent acquisitions. This pattern is suggestive of rare acquisition of genes of known function followed by lower probability of loss for these genes. Stated otherwise, the acquisition of known-function genes is rare, but these genes have a higher probability of being adaptive and, thus, are less likely to be lost. At the opposite extreme, transposable elements and prophage-related genes have high probabilities of being acquired, but since they often have deleterious consequences, they are quickly purged from the populations. As a result, gains inferred in ancestral nodes, i.e., those for which we can still infer an acquisition regarding extant genomes, are enriched in adaptive genes and impoverished in transposable and phage elements.
The pan-genome includes the ancestral genome, which in turn includes the core genome. As one goes from the smallest to the largest gene set one expects to find more accessory and fewer essential functions. Indeed, functions encountered more frequently in the smaller sets include biosynthesis of amino acids, nucleotides, co-factors and proteins, and, to a lesser extent, metabolism of DNA, fatty acids, and phospholipids, transcription and protein fate (Table S5). On the other hand, regulators, cell envelope, biological processes and mobile elements are over-represented in the larger sets. Interestingly, the inferred ancestor of all E. coli lacks none of the 23 high-confidence essential genes that are missing in the core genome. It thus provides a better representation of the housekeeping and essential functions of the E. coli cell than does the core genome.
Gene acquisition and loss have important roles in transitions between commensalism and pathogenicity [51],[52]. Epistatic interactions between virulence determinants and the genetic background may also be important [22]. Indeed, the strains with the highest pathogenicity and classified as biosafety level 3 (S. dysenteriae serotype 1 and enterohemorragic E. coli O157:H7) (Table 1) are closely related (Figure 4). This high degree of pathogenicity is due to toxins that could require a specific genetic background to achieve appropriate expression.
To understand the link between virulence and genetic background, we first looked for functional genes categorically present (i.e., ubiquitous in the clade but absent elsewhere) or absent (i.e., absent in the clade but ubiquitously present elsewhere) within three main phylogenetic groups: A, B1 and B2 (with group D being unsuitable for the analysis as it is paraphyletic) (Table 4 and Table S6). Since only one group A strain was available (E. coli K-12 MG1655), we added to this analysis the genome of strain HS (http://msc.jcvi.org/e_coli_and_shigella/escherichia_coli_hs/index.shtml), a group A human commensal strain. Few genes (5 to 81 per phylogenetic group, depending on the group) were found to be specific to and ubiquitous within the particular phylogenetic group, in agreement with the high gene flow observed in the species. However, the numbe of specific genes was higher within group B2 than within other phylogenetic groups, despite the greater number of studied B2 genomes and the greater time of divergence of this phylogenetic group (two factors that should decrease the number of shared genes) (Table 4). This could indicate that these genes stably gained or lost, contribute to the fitness of the group B2 strains. Indeed, only one of these genes corresponds to a transposase and none to phages, whereas 75% have an assigned function. This is significantly higher (Chi square test, p<0.001) than the proportion of genes with assigned functions in the B2 pan-genome (4097 of 8439, 48.5%). Furthermore, the distribution of the genes with assigned functions among different functional categories (‘Product type’ annotations, Table 4) is significantly different for the specific genes as compared with the pan-genome (Chi square test, p = 0.049). The study of Pearson residuals shows that the enzymes and transporters and carriers categories contribute significantly to this difference. Integrative analysis of the documented functions of the specific genes shows a large part of them to be involved in metabolism (Table 5). These observations represent a hallmark of selection and suggest an important role for metabolism in the niche adaptation of group B2 strains that needs to be further substantiated by experimental analyses.
We then examined whether the presence of specific genes could be related to a specific phenotype. No gene was specific either to commensal strains or to pathogenic strains in general. However, in extraintestinal pathogenic strains (ExPEC pathotype) 16 genes were specifically present and 1 was specifically absent (Table 4). Most of these genes have an assigned function corresponding mainly to 2 clusters: (i) the pap operon, a well-known adhesin determinant involved in the pathogenesis of urinary tract infection [53], and (ii) two genes coding for an aldo-keto reductase activity (one of these genes shares 95% identity with akr5f1 gene from Klebsiella spp [54]) and a divergent lysR family regulatory gene (Table S6). In addition, when considering intrinsic extraintestinal virulence potential as assessed using a mouse model of septicaemia that avoids host variability [32], no gene specific to the virulent phenotype was identified. All these data indicate that extraintestinal virulence is a multigenic process resulting from numerous gene combinations and multiple redundancies. Furthermore, the fact that no gene specific to extraintestinal infection could be identified reinforces the hypothesis that extraintestinal virulence is a coincidental by-product of commensalism [42]. This suggests that the development of vaccines specific for extraintestinal infections will be extremely difficult. Any gene target likely will also be present in some commensal strains; therefore, such vaccines will presumably lead to potentially undesirable modification of the resident microbiota. Twenty and 4 genes were specifically present and absent, respectively, in intestinal pathogenic strains (with Shigella excluded from the analysis). All except 2 of these genes are of phage and IS origin or of unknown function.
We also took the unique opportunity to do a comparative genomic analysis of the recently reported B2 human commensal clone (represented by strain ED1a, as sequenced in this work), which is avirulent in the mouse lethality model [55]. Thirty-one genes were specifically present and 9 were specifically absent in the B2 strains that were virulent in the mouse lethality model (B2 mouse killer strains) (Table 4 and Table S6). Interestingly, among the 9 absent genes, 8 belong to the mhp operon. The catabolic pathway of phenylpropionate and its derivatives is split in E. coli into two operons, the mhpR mhpABCDFET and the hcaR hcaEFCBD operons. The hca operon is specifically absent in all the group B2 strains (Table 5). Strain ED1a is thus an exception, as it possesses the mhp, but not the hca operon. This may suggest some sort of involvement of aromatic compounds in the virulence of B2 strains.
A similar comparative genomic analysis involving the Shigella strains identified 38 genes (30 from the virulence plasmid [56], as expected) to be specifically present, but also 32 genes to be specifically absent (Table 4). Excluding the plasmid genes, 70% have an assigned function, which is significantly greater (Chi square test, p<0.001) than for the genes of the Shigella pan-genome (3832 of 9351, 41%). Here again, the distribution of the genes with assigned functions among different categories (Table 4) is significantly different from the Shigella pan-genome (Chi square test, p = 0.027), with a disproportionate emphasis on the transporters and carriers category, and more generally on metabolism-related functions (Table 6). The specificity of this pattern of gene loss suggests a footprint of selection through an antagonistic pleiotropy mechanism of adaptation [57] during the very peculiar Shigella intracellular life style. Such a life style also leads to the reduced effective population size of Shigella, and to less efficient selection [49]. Thus, it has been argued frequently that gene loss in Shigella is the result of independent mutation accumulation. It is likely that most gene loss in Shigella is indeed the result of less efficient selection, but our data suggest that inactivation of these 32 genes, or a fraction of them, is positively selected.
We further substantiated the role of polyamine metabolism and transport in Shigella virulence by identifying the absence of (i) speG involved in spermidine biosynthesis and (ii) the cad genes involved in cadaverine biosynthesis [52]. It has been shown that the presence of cadaverine prevents the escape of S. flexneri from the phagolysosome [58]. The absence of spermidine acetylation by SpeG could preclude export of acetyl-spermidine. Another negative phenotype of Shigella, not often discussed in relation to pathogenicity, is their lactose-negative character, arrived at by convergent evolution [7]. We found that within the lactose operon region, the only gene always inactivated is lacY, the permease coding gene. As the role of pH is essential for colonisation of a novel niche, the lactose permease, a proton-driven transporter, may act against adaptation of the bacteria to the acidic phagolysosome. One might speculate that a beta-galactoside present in the phagolysosome could be transported out with import of protons, leading to a proton influx that would rapidly kill the bacteria. Gene decay would thus have protected Shigella against this host protective mechanism.
Bacterial chromosomes are highly organised with respect to their interaction with cellular processes such as replication, segregation and transcription [59]. To understand how the massive flux of genes we have documented can be compatible with chromosome organisation we inferred the number of insertion and deletion events at each branch of the species tree (see Materials and Methods, Figure 7 and Figure S5). The average acquired fragment contains 4.3 genes, whereas the losses average only 3 genes (Wilcoxon test, p<0.001). These values are nearly half the previously published ones [60], most likely because our analysis includes many more closely related strains and uses the inference of ancestral states, leading to a more accurate estimation of multiple contiguous insertions and deletions. The total number of genes gained and lost is expected to be roughly similar, since enterobacterial genomes have relatively similar sizes. Therefore, gains correspond to larger fragments and losses to more frequent events. The size of the fragments of gains or losses varies widely. More than half of inferred losses and gains involve a single gene. Only 5% of losses and 8% of gains correspond to events including more than 10 genes, but these include around half of the genes involved in gains and losses (54% and 40%, respectively). These values are similar for internal branches, small external branches and long external branches (Kruskal-Wallis test, p>0.05), suggesting that our inference is unbiased with respect to successive events taking place at the same locations in long branches or by selection-purging older events in internal branches. Variation in gene repertoires has been described as being scattered on the chromosome of E. coli and balanced between the two replichores [61]. For the numerous small insertions and deletions this distribution results naturally from random insertion/deletion of genetic material. Such small indels are expected to have little impact on the large-scale organisation of the genome.
What about the very large insertions/deletions? The 554 such events that involve more than 10 genes over-represent insertions over deletions (Fisher exact test, p<0.001), as expected given that insertions are typically larger. These events involve an average of 29 genes each, with a maximum of 157 genes for a single event. Unsurprisingly, known pathogenicity islands and prophages are included in these large events. The insertion of very large DNA segments, even if it takes place in intergenic regions, will have important consequences for the organisation of genomes. Therefore, we investigated where such insertions took place. We used the ancestral order of the core genome and computed, for each genome, the number of non-core genes between consecutive core genes. (The rare positions corresponding to synteny breakpoints in a genome were ignored for that genome.) This analysis revealed that in most genomes gene acquisition and loss takes place at precisely the same locations across genomes, i.e., between the same two contiguous core genome genes (Figure 8, Figure S7). Thus, the E. coli genome contains striking integration hotspots.
An example of an insertion hotspot at pheV tRNA gene in 12 E. coli strains is represented in Figure 9. This example shows that very different genetic information occurs at the same hotspot in different genomes. Interestingly, it also shows a patchy structure, with the information segmented into modules that can be found independently in other locations of other genomes. The presence/absence of specific modules is uncorrelated with either the phylogenetic group or the pathotype. For example, module 14 (immunoglobulin-binding genes, which encode a surface-exposed protein that binds immunoglobulins in a nonimmune manner) is present in strains 55989 (group B1, EAEC), APEC O1 and S88 (group B2, ExPEC); module 19 (N-acetylneuraminic acid degradation) is present in strains UMN026 (group D, ExPEC) and CFT073 (group B2, ExPEC) only; and module 2 (N-acetylneuraminic acid synthesis), with the pattern [1-2-3-4-5] is absent in strains UMN026, CFT073, ED1a (group B2, commensal) and 536 (group B2, ExPEC). Actually, the organization of the modules is identical in APEC O1 and S88, and very similar in UMN026 and CFT073. Such a modular structure of the hotpots suggests either multiple integrations or frequent recombination between integrative elements.
While 51% of all intergenic regions between pairs of contiguous core genes show no single insertion or deletion in any of the 21 genomes, we found 133 such locations with an average of more than 5 non-core protein-coding genes per genome. These locations accumulate 71% of all non-core pan-genome genes. Nearly two thirds of the hotspots (62%) lack prophages in all genomes. Genes in hotspots have an average of 4 orthologs in the other genomes. Yet, this average is somewhat misleading since some genes have many orthologs and the majority has practically none. Therefore, hotspots correspond to regions of abundant and parallel insertions and deletions of genetic material. While the existence of large insertions and deletions in E. coli has been abundantly described [62],[63], our data shows that these events take place systematically at the same regions in different genomes.
The genomes of E. coli harbour many prophages and genomic (e.g., pathogenicity) islands, which typically integrate in the chromosomes by site-specific recombination in a tRNA gene through the action of phage-like integrases [64]. We assessed how frequently such elements are associated with hotspots. We found that 83% of the hotspots showed no tRNA gene at the edge of the element, within a 3-gene window, in any of the genomes. When tRNA genes were indeed found, they tended to be present in practically all genomes. Since each E. coli genome has close to 100 tRNA genes, the occurrence of tRNA genes in the neighbourhood of 17% of hotspots can partly be due to chance. We therefore searched the hotspots for homologs of a set of 8067 integrases obtained from Swissprot by using Blastx to include potentially pseudogenised integrases. Using our standard criteria for homology (see Materials and Methods) we found that more than half of the hotspots have no integrase homolog in any genome, whereas fewer than 6% have integrases in the majority of the genomes. Decreasing the similarity criterion for a homolog to 40% identity increases the number of putative integrases, but half of the hotspots still have at most two distant homologs of integrases, and these are present in the majority of genomes in only 17% of the hotspots. This seriously challenges the widely held view that E. coli integration hotspots are mostly determined by the distribution of tRNA genes and that such integrations systematically take place by phage-like integrase elements.
What else could create such hotspots? It would be predicted that selection for preserved integrity of composite regulatory elements, genes, operons, supra-operonic structures, nucleoid folding-domains and macrodomains should reduce the number of locations where large insertions can occur without causing significant loss of fitness [59]. For example, ∼90% of the genomes consist of genes and half of the remaining 10% represents intergenic regions within operons. Selection should thus effectively forbid most insertion points in the genome. However, once a permissive region has acquired a large element, and since most transferred DNA has no adaptive value, subsequent integration in the region becomes more likely because the region offers a larger target for neutral insertion. The insertion of a large element in a permissive region will then result in a founder effect that amplifies the likelihood of the permissive region becoming a hotspot.
Some regions may be more prone to recombination because of their sequence/motif composition, e.g., the presence of motifs recognised by integrases or the machinery of homologous recombination. We tested if the regions flanking the hotspots showed higher frequencies of chi sequences, but found no significant effect. DNA structure may also play a role, e.g., because chromosome folding leaves some regions more exposed than others for recombination with incoming DNA [65]. The 133 hotspots contain 61% of all synteny breakpoints, which is much more than expected given the number of these locations (Chi square test, p<0.0001), but close to the expected value if one considers that rearrangements cannot disrupt core genes and that the hotspots are very large (Chi square test, p>0.05). This shows that insertion/deletion hotspots are also rearrangement hotspots, even though we initially removed rearranged positions to identify the insertion/deletion hotspots (thus being conservative). It also suggests that rearrangements occur in these regions because they are permissive to change not because they are intrinsically recombinogenic, since the frequency with which they rearrange simply reflects their larger size. However, even if hotspots are not intrinsically recombinogenic they can still be caused by the brokering effect of homologous recombination. Indeed, incoming DNA once integrated in one genome can propagate within the population by lateral transfer via classical homologous recombination involving the homologous flanking regions. Given the observed rates of recombination in the species, this mechanism could quickly lead to the horizontal spread of highly adaptive newly acquired genes. We describe some evidence for this in the next section.
For any given sequence alignment, the likelihood of the overall gene tree topology, i.e., the phylogenetic congruence, reflects the extent to which the phylogenetic signal of the sequences was altered by recombination. While the concatenate of genes provides a strong phylogenetic signal, the individual genes' histories can be very diverse as a result of recombination. Furthermore, these histories may depend on the genes' positioning in the chromosome. Notably, if homologous recombination helps in disseminating recent acquisitions, as we propose, the core genome around these hotspots should show signs of recombination as indicated by phylogenetic incongruence. We therefore made an analysis in 5 kbp sliding windows along the multiple genome alignment to identify the most phylogenetically incongruent regions (see Material and Methods). This method identified two large regions of very strong incongruence, one centred around rfb (Figure S8), the operon involved in O antigen synthesis, and the other around the leuX tRNA gene, and including fimA, which is under diversifying selection and is involved in the adhesion of bacteria to host cells [66]. Both loci were previously identified as hotspots of phylogenetic incongruence [67],[68]; the present analysis reveals how much they affect the chromosome.
Recombination at the rfb locus significantly affects congruence within a striking 150 kbp surrounding region, i.e., from positions 1988 kbp to 2138 kbp (100% of windows tested had scores lower than 1.96 standard deviation away from the average, with an average of −4.84 and peaks at −10.19). The fim locus includes an incongruence region close to 200 kbp in length (from positions 4421 kbp to 4618 kbp, average −2.54 standard deviation and 73% with lower than −1.96 standard deviation and peaks at −6.65). Interestingly those two regions are centered on integration hotspots and encompass 11 of the 133 hotspots of integration. The genes present in such loci arose most likely by lateral transfer since they are highly dissimilar between strains. For example, genes at the rfb locus genes can exhibit less than 50% similarity, while the leuX locus encompasses a highly variable assortment of non-homologous inserts in all the genomes sequenced. Hence at least for those two major loci we find a striking link between hotspots of integration and hotspots of homologous recombination. In the case of the rfb locus, it is worth noting that the incongruence signal we observe might be a composite signal, due not only to rfb but also to neighbouring loci. Within the above defined rfb region of incongruence, a flagella locus (fli operon) associated with two hotspots of integration is also under diversifying selection. Moreover, the high pathogenicity island (HPI) is integrated within that high recombination region in many isolates and corresponds also to a hotspot of integration. It has been suggested that after a recent and unique integration event, the HPI has propagated within the species by homologous recombination [69]. The propagation or diversification of these loci, located to the left of rfb, through homologous recombination might generate the asymmetrical pattern of phylogenetic incongruence we observe around the rfb locus (extended incongruence on the left side of the rfb locus) (Figure S8).
We found 23 other regions with weaker signatures of incongruence (i.e., with a 5 kbp sequence incongruence score more than 2 standard deviations from the average), each spanning less than 20 kbp. It is important to note that most of these incongruent regions include genes involved in diversification of genetic information and often pathogenicity. The vast majority of these include 3 groups of common genes. First are regions with the porin-encoding genes ompA and ompC, the flagella-encoding genes, the rfa locus coding for the core lipopolysaccaride and genes coding for several membrane proteins such as LolCDE, CcmABCDE, ABC transporter, AroP APC transporter, LplT-aas, FadK, YeaY, EamB, YhgE and YicG membrane proteins. These loci are probably involved in diversifying selection since they code for antigenic proteins exposed at the cell surface. Second are two regions encompassing mismatch repair genes (mutS and mutH) that have been shown to be under selection for cycles of inactivation and reacquisition through recombination [70]. Third is a region associated with the integration of a locus that can provide resistance to phages through clustered regulatory interspaced short palindromic repeats (CRISPRs) [71].
All available methods estimate the effective, not the intrinsic, recombination rate. Effective recombination results from the intrinsic recombination rate and the ensuing selection on recombinants. Most of the phylogenetic incongruence hotspots we found contain genes under diversifying selection, for instance to escape immune pressure or to acquire resistance to phage. Hence, it is very likely that differences in the intensity of selection might be responsible for the differences observed in the size of the regions affected by a phylogenetic incongruence hotspot. A recombinant carrying a new allele at a locus under strong diversifying selection will be selected and thus will increase rapidly in frequency in populations. Hence, the recombinant will invade the local population before any further recombination occurs at the locus [72]. In that case, sampling the genome after the action of natural selection allows identification of the original recombining fragment. In contrast, if selection is moderate, the recombining fragment that brought the interesting allele into the genome will be covered by many further recombination events before it reaches high frequency. In this case, only fragments around the selected allele will retain the trace of the recombination event. As a consequence, when selection is intense, one expects to identify long recombinant fragments in some strains, as we did at the rfb or leuX loci. Our observations suggest that the intensity of diversifying selection acting on the rfb and leuX-fimH loci are under very strong selective pressure compared with the diversifying selection acting on the core LPS, the flagella or some of the porins. The fact that most hotspots of integration (117 among 133) do not result in hotspots of phylogenetic incongruence suggests that they carry neutral or deleterious genes. Conversely, it also suggests that some horizontally acquired genes can be highly beneficial (e.g., 11 hotspots of phylogenetic incongruence around the rfb or leuX-fimH locus) or moderately beneficial (e.g., 4 hotspots of integration associated with hotspots of phylogenetic incongruence) and that this results in different selection footprints in the neighbouring core genome.
The existence of integration and phylogenetic incongruence hotspots brings to the fore the conflict between genome dynamics and organisation. We therefore analysed the variation in recombination along the backbone sequence (as estimated by a population genetic-based approach), using a sliding window of 3 kbp on the multiple genome alignment and a step size of 500 bp. This analysis revealed a large region around the terminus of replication with a particularly low ratio between gene conversion and mutation rates (Cgc/theta) (Figure 10). The region between 1 Mb and 2 Mb shows lower gene conversion rates, since there is 20% lower chance for a base to be involved in a gene conversion event (Cgc×Lgc, unilateral t test: p = 1e-21). This region also shows 10% lower levels of polymorphism (theta of Watterson, p = 1e-7), i.e., variations within the E. coli species, and 2% lower G+C content (Figure 10). A+T richness at the terminus region has been suggested to result from higher mutation rates [73]. Based on comparative genomics with Salmonella, it was also shown that divergence, i.e., the genetic distance between species, slightly increased closer to the terminus [74],[75], further supporting the hypothesis of a higher local mutation rate. Using our newly sequenced outgroup genome E. fergusonii, which unlike Salmonella does not shows saturation of synonymous substitutions, we found that the terminus domain has synonymous and non-synonymous substitution rates twice as high as the rest of the chromosome. While decreased G+C content and increased divergence could reflect a higher mutation rate at the terminus, such an interpretation is contradicted by the observed lower polymorphism.
Theoretical population genetic studies have shown that the fluctuation of recombination frequency along chromosomes affects the level of polymorphism and the efficiency of selection [76]. When there are numerous deleterious mutations and low recombination rates, a fraction of the population bearing deleterious alleles is doomed to disappear in the long term without contributing to the gene pool of the future population. The relevance of this phenomenon, referred to as background selection, requires the existence of deleterious mutations of moderate effects, i.e., mutations that can persist for some time in the population before selection wipes them out. At the population level, this result in an excess of rare alleles, which can be estimated by Tajima's D statistics. We found that overall a gene's average Tajima's D was slightly negative (indicating an excess of rare alleles). However, the Tajima's D of synonymous mutations was null, while that of non-synonymous mutations was much more negative (Figure S9). This suggests that most non-synonymous mutations are deleterious since, in contrast to synonymous mutations, they do not increase in frequency within the population reflecting the purging effect of natural selection. Therefore the conditions for the action of background selection are met. Furthermore, under background selection, a reduced recombination rate results in a decreased polymorphism (such as we observed around the terminus), an increased fraction of rare alleles and a decreased efficiency of selection [76]. The terminus region shows a lower Tajima's D than the rest of the chromosome (Student bilateral test, p<0.00001). It also shows a reduced ratio of non-synonymous to synonymous polymorphism (Student bilateral test, p<0.002). This suggests that more non-synonymous mutations, presumably slightly deleterious, persist around the terminus. When applying the same approach to the ratio of non-synonymous to synonymous divergence, we found more non-synonymous mutations fixed around the terminus (Student bilateral test, p<0.05). All these observations are in agreement with a reduced efficiency of selection at this region, compatible with the effects of background selection in low recombination regions.
The observed co-occurrence of lower GC% and lower recombination rate at the terminus could also indicate a reduced action of recombination to purge deleterious mutations in that region. Most mutations tend to be from GC to AT and, as our analysis of Tajima's D revealed, most non-synonymous mutations are presumably deleterious. Consequently, if a segment of DNA lacking deleterious mutations replaces a fragment that contains many of them, presumably GC-towards-AT, the resulting recombinant will be selected for and, hence, will increase the GC content. Therefore, in regions of low recombination rate, a greater number of deleterious GC-towards-AT mutations will accumulate. This is in agreement with recent analyses showing an association between G+C enrichment and purifying selection of non-synonymous substitutions [77].
Alternatively, recombination could have a direct mutagenic effect. The biased gene conversion hypothesis, which enjoys growing popularity to explain the G+C heterogeneity in mammalian genomes, states that mismatches in recombination heteroduplexes are repaired in favour of G and C [78]. If in E. coli, as in humans and elephants, biased gene conversion results in G+C enrichment, then lower conversion rates at the terminus should result in the observed lower G+C content. Biased gene conversion results in the biased segregation of nucleotides and, therefore, in a gap between the composition of genomes and their mutation patterns. We had previously found that such a gap was common in bacterial genomes [79]. The re-assessment of those data showed that in all 6 E. coli genomes considered in our previous work the G+C content was higher than expected given the observed mutational patterns. This suggests that mutations towards G and C are more likely to attain fixation, in agreement with the hypothesis of biased gene conversion in E. coli.
Both hypotheses are compatible with the pattern observed, but attribute different meaning to reduced GC% at the terminus. In the biased gene conversion hypothesis, lower GC% is just a result of the mutational bias induced directly by recombination, while in the second one a lower GC% reflects the lower efficiency of recombination to purge slightly deleterious mutations and is therefore a signature of maladaptation.
Why should conversion rates be lower at the terminus? This could be explained by the patterns of genome organisation. Firstly, in exponentially growing E. coli cells the regions near the origin of replication are present in many more copies than the regions near the terminus [80]. Therefore, they provide more abundant targets for gene conversion with foreign DNA. Because of gene dosage effects the origin of replication is also enriched in highly expressed genes, which are under stronger purifying selection. This might lead to lower observed rates of mutation or to higher rates of recombination, if recombination's role is to maintain housekeeping functions [81]. Secondly, the low recombination / high A+T content region near the terminus coincides with the boundaries of the Ter macrodomain of chromosome folding in E. coli [82]. Four macrodomains (Ori, Ter and two flanking Ter named Right and Left: Figure 10) have been described [82]. These macrodomains are compacted structures that act as intra-chromosomal recombination insulators. Tight compaction of the Ter domain might lead to lower conversion rates with incoming DNA. The link between the frequency of gene conversion, biased sequence composition, chromosome compaction and selection highlights the intimate association between genome dynamics and chromosome organisation.
New high-throughput sequencing technologies will soon allow the sequencing of hundreds of strains of the same species, but not to completion and closure. The genomes of Escherichia that we sequenced, the previously sequenced ones, plus others and our re-annotation efforts, will provide a solid basis for the next phase of E. coli genomics in which population genetics and experimental evolution will have important roles. We also hope to have contributed to narrowing the gap between population genetic and phylogenetic approaches in studying genome evolution by showing that they both can be used to untangle the effects of gene dynamics on adaptation and genome organisation. Within a bacterial species, the core genome evolves mostly through mutation and recombination, whereas the rest of the genome is also subject to horizontal gene transfer. While this fits with qualitative observations in other species [83]–[85], in E. coli the rates of lateral transfer are particularly high and lead to very short gene residence times. Furthermore, once introduced by lateral transfer, genes can spread by homologous recombination at the flanking regions. Despite this very high gene flow, genes co-exist in organised genomes. The conflict between genome dynamics and organisation may have resulted in the striking integration hotspots, which confine regions of high instability. It may also have resulted in regionalised gene conversion.
Chromosomal plasticity certainly accelerates the adaptation of E. coli to varied environments. First, it allows many parallel and specific evolutionary pathways of gain and loss of genes leading to convergent phenotypes. Second, it allows multiple gene combinations that, with epistatic interactions, will result in phenotypic diversification. As a result of these complex evolutionary patterns, most often there is no simple association between the presence of a gene and a given phenotype. For example, our genomic analysis of the extraintestinal virulence phenotype suggests that it will be very difficult to develop a vaccine against extraintestinal infections without affecting also resident intestinal microbiota because there is no single determinant of the former. The vast diversity among E. coli genomes suggests that the key to understanding the emergence of such phenotypes resides in ampler sampling of natural isolates combined with a systematic analysis of the data at a physiological level.
Six E. coli strains as well as the type strain (ATCC 35469T) of E. fergusonii, the closest E. coli-related species [31], were selected for complete genome sequencing (Table 1). Among the E. coli strains, 2 were commensal: IAI1 (serogroup O8) was isolated from the faeces of a young healthy military conscript in the 1980s in France [23] and ED1a (serogroup O81) was isolated in the 2000s from the faeces of a healthy man in France and belongs to a human-specific widespread commensal clone that is increasing in frequency [55]. Four E. coli strains were pathogenic. Enteroaggregative E. coli strain 55989 was originally isolated from the diarrheagenic stools of an HIV-positive adult suffering from persistent watery diarrhea in Central African Republic [86]. The enteroaggragative pathotype is recognized as an emerging cause of diarrhoea in children and adults worldwide [87]. Among the three extraintestinal pathogenic strains, IAI 39 (serotype O7:K1) was isolated from the urine of a patient with pyelonephritis in the 1980s in France [23]. UMN026 (serotype O17:K52:H18) was isolated from a woman with uncomplicated acute cystitis in 1999 in the USA (Minnesota) and is a representative of a recently emerged E. coli clonal group (“clonal group A”) that is now widely disseminated and a cause of drug-resistant urinary tract and other extraintestinal infections [88]. S88 (serotype O45:K1:H7) was isolated in 1999 from the cerebro-spinal fluid of a new born with late-onset neonatal meningitis in France and represents what is now considered a highly virulent emerging clone in France [89]. These strains were distributed in 3 of the 4 main E. coli phylogenetic groups: IAI1 and 55989 belong to group B1, UMN026 and IAI392 belong to each of the two major subgroups within group D, and ED1a and S88 belong to subgroups VIII and IX, respectively, within group B2 [42]. Few data are available on E. fergusonii strains. They have been isolated from humans and warm blood animals, sometimes in pathogenic (intestinal and extraintestinal) conditions [90]–[92]. The main characteristics of the 14 strains (8 E. coli sensu strictu and 6 Shigella) with freely available genomes at the time of the study are presented in Table 1. These genomes were used for comparison purpose.
Three DNA libraries were constructed to determine, for each strain, the complete genome sequence. Two of the libraries were obtained after mechanical shearing of the genomic DNA and cloning of the resulting 3 kbp and 10 kbp inserts into plasmids pcDNA2.1 (Invitrogen) and pCNS (pSU18 derived), respectively. DNA fragments of about 30 kbp generated after partial digestion using HindIII and/or Sau3A were introduced into pBeloBac11. Vector DNAs were purified and end-sequenced using dye-terminator chemistries on ABI3730 sequencers to provide an average of 12-fold coverage for each genome. A pre-assembly was made without repeat sequences, as previously described [93] using Phred/Phrap/Consed software package (www.phrap.com). The finishing step was achieved by primer walking, transposition and PCR.
Once the consensus sequence of a first complete (single contig) assembly was available for one of the new genomes, gene prediction was conducted using the AMIGene software [94]. The predicted coding sequences (CDSs) were assigned a unique identifier prefixed with “ECED1_” for E. coli ED1a, “EC55989_”, for E. coli 55989, “ECIAI1_” for E. coli IAI1, “ECIAI39_” for E. coli IAI39, “ECS88_” for E. coli S88, “ECUMN_” for E. coli UMN026, and “EFER_” for E. fergusonii ATCC. These identifiers start with ‘p’ if the corresponding CDSs are encoded on plasmids. The sets of predicted genes were submitted to automatic functional annotation, as previously described [95]. Apart from the plasmid-encoded genes, the final functional assignation was based on the transfer of the recently updated E. coli K-12 MG1655 annotations [96] between strong orthologs i.e., 85% identity over at least 80% of the length of the smallest protein (Table S2A). Sequence data for comparative analyses were obtained from the NCBI database (RefSeq section, http://www.ncbi.nlm.nih.gov/RefSeq). Putative orthologs and synteny groups (i.e., conservation of the chromosomal co-localisation between pairs of orthologous genes from different genomes) were computed between each newly sequenced genomes and all the other complete genomes, as previously described [95]. All these data (syntactic and functional annotations, results of comparative analysis) are stored in a relational database, called ColiScope. Manual validation of the automatic annotation by multiple users in different locations was performed using the MaGe (Magnifying Genomes, http://www.genoscope.cns.fr) web-based interface. For each newly sequenced genome, only ‘specific’ regions, i.e., those containing genes not orthologous to ones in E. coli K-12 MG1655 or to expert annotated genes in another genome of the ColiScope project, were manually annotated (Table S2A). In total, 9776 genes were annotated by our group.
This expert work was also used to re-annotate the other public and Shigella genomes. This allowed the creation of a set of consistent expert annotations for the 20 genomes. First, we integrated these genomes into the ColiScope database by using MICheck, a method that enables rapid verification of sets of annotated genes and frameshifts in previously published bacterial genomes [97]. Some inaccurate or missed gene annotations were defined for these genomes (see Table S2B and Table S3 for the list of newly predicted genes in the 14 analyzed genomes). Second, we automatically transferred the functional annotation of E. coli K-12 MG1655 genes, or genes annotated in the context of this project to the genes in the other genomes that showed very strong sequence similarity (85% identity over at least 80% of the length of the smallest protein). The remaining genes, i.e., those without orthologs in E. coli K-12 MG1655 or one of the new Escherichia genomes, retained the original functional annotations (column ‘Specific’ genes in Table S2B).
The new E. coli and E. fergusonii nucleotide sequences and annotations data have been deposited in the EMBL database (http://www.ebi.ac.uk/embl; see accession numbers list below). In addition, the ColiScope database, which includes all data for the set of Escherichia and Shigella strains sequenced to date, is publicly available via the MaGe interface at https://www.genoscope.cns.fr/agc/mage.
A preliminary set of orthologs was defined by identifying unique pairwise reciprocal best hits, with at least 80% similarity (∼85% identity) in amino acid sequence and less than 20% difference in protein length. The analysis of orthology was made for every pair of E. coli/Shigella genomes. The core genome, consisting of genes ubiquitously found among all strains of the species, was defined as the intersection of pairwise lists.
For every pair of genomes this list of persistent orthologs was then supplemented, with attention to conservation of gene order. Because (i) few rearrangements are observed at these short evolutionary distances, and (ii) horizontal gene transfer is frequent, genes outside conserved blocks of synteny are likely to be xenologs or paralogs. Hence, we combined the homology analysis (protein sequence similarity ≥80%, ≤20% difference in protein length) with the classification of these genes as either syntenic or nonsyntenic, for positional orthology determination. The analysis was made for every pair of E. coli/Shigella genomes. The definitive list of orthologs of the pan-genome was then defined as the union of pairwise lists.
A syntenic block was defined as a set of consecutive pairs of genes in the core genome. Conserved order gene blocks are obtained by comparison of the localisation of best bi-directional hit pairs in the core genome, adopting a window size of one gap.
These lists were also used to perform gene accumulation curves using R, which describe the number of new genes and genes in common, with the addition of new comparative genomes (Figure 1). The procedure was repeated 1000 times by randomly modifying genome insertion order to obtain median and quartiles.
In the same bacterial species, homologs (paralogs, orthologs, xenologs) were defined by identifying reciprocal blastp, with ≥80% similarity in amino acid sequence and ≤20% difference in protein length. Among different proteobacterial species, orthologs were defined by identifying unique pairwise reciprocal best hits, with ≥40% similarity in amino acid sequence and ≤20% difference in protein length. The analysis of orthology was made with 99 proteobacterial genomes.
Whole genome alignments of the 20 E. coli/Shigella study strains were performed using the Aligner algorithm of the MAUVE program, version 2.0.0 [98], with the following parameters: –island-size = 20 –backbone-size = 20 –max-backbone-gap = 20 –seed-size = 19 –gapped-aligner = clustal –max-gapped-aligner-length = 10000 –min-recursive-gap-length = 5000 –weight = 5000. The MAUVE output file was further treated so as to assign each part of the alignment to either one of two categories, ‘backbone’ or ‘variable segment’ (previously named ‘loops’), as described [99]. Briefly, regions not belonging to a “match”, as defined by MAUVE and less than 10 kbp long were aligned using ClustalW and the alignment was automatically inspected. The region was considered as a backbone segment if all pairwise comparisons gave more than 76% identity, with never more than 20 consecutive gaps. In all other cases, the entire region was considered as a variable segment.
To produce the DNA alignment file from the above mentioned procedure, the coordinates of all backbone segments on each genome were extracted and aligned with MAFFT, version 6.24 [100], using a home made Perl script. Segments were first aligned with the ‘–globalpair option’, which is suitable for a suite of globally alignable sequences. When problems occurred (especially for long backbone segments), MAFFT alignments were computed using the ‘–auto’ option which automatically selects an appropriate alignment algorithm according to data size.
Statistical analysis along the chromosome (scans). Along the chromosomal multiple genome alignment we studied the variation of descriptive statistics, such as the GC% and estimates of the mutation and recombination rates. We estimated each statistics, F, on a sliding window of constant size along the concatenated alignment. We then estimated the average value of the statistic μ and its standard deviation with the median and the inter quartile distance (normalised by a factor of 1.38) as these estimates are less affected by the existence of extremes values. We then calculated the standardised cumulative sum along the genome . When the cumulative sum is decreasing in a region, it means that this region harbours a lower than average value of the statistics. Hence for each statistics we can identify the boundaries of regions having atypical values.
To reconstruct the phylogeny of the strains, we used two data sets: the genes common to all the E. coli/Shigella and E. fergusonii strains (Escherichia core genome) and the genome backbone, defined as above. We also used several methods for each dataset.
(i) The reference phylogenetic tree of the Escherichia core genome genes was reconstructed from the concatenated alignments of 1878 genes of the core genome of the E. coli/Shigella and E. fergusonii strains. We used Tree-puzzle 5.2 [101] to compute the distance matrix between all strains using maximum likelihood under the HKY+gamma (with 8 categories)+I model. The tree was then built from the distance matrix using BioNJ [102]. We made 1000 bootstrap experiments on the concatenated sequences to assess the robustness of the topology. (ii) We also inferred a tree for each of the 1878 genes in the core genome, using maximum likelihood with PHYML 2.4.4 with a GTR+gamma+I model for each gene [103]. For each tree we extracted the relevant parameters of the model and made a weighted average to obtain a global average model. We used the lengths of the genes as weights of the average. The global model thus obtained was used to infer a tree based on the concatenation of the genes using Tree-puzzle 5.2. The tree was then built from the matrix of distances using the BioNJ algorithm. To check that the branch lengths obtained with this method were correct we computed them by maximum likelihood by imposing the tree topology (baseML from package PAML 4 implementation [104]). The differences found were extremely small. To assess the robustness of the tree we bootstrapped 1000 times the concatenated sequences, each time launching Tree-puzzle with the same previously inferred global model. (iii) We performed comparisons among phylogenetic trees. To test if the phylogenetic tree of each gene (as inferred by maximum likelihood using the PHYML 2.4.4: GTR+gamma+I model) is significantly different from the global tree as reconstructed from the concatenation of genes of the Escherichia core genome, we performed several tests for comparing tree topologies using likelihood. These included a SH test [105], two types of Kishino and Hasegawa test (KH test) (i.e., the original two-sided KH test as described in [106] and the one-sided KH test [107] using pairwise SH tests), and the expected likelihood weights (ELW) [108]. For the simulations, we used these tests as well as the Robinson and Foulds test [45]. All tests used a 5% significance criterion. (iv) We also built a consensus tree (extended majority rule as implemented in CONSENSE) using PHYLIP 3.66 package [109] from the set of trees inferred in (ii).
Using MAUVE's global alignment we also extracted a backbone concatenate which we input into Tree-puzzle with the HKY+gamma (with 8 categories)+I model to obtain a matrix of distances. BioNJ was then used to reconstruct the unrooted tree from the distance matrix.
Using the chromosomal multiple genome alignment, we studied the likelihood of the species tree for any 5 kbp window of conserved sequence along the genome. Since the likelihood, as estimated with PHYML [103] under the HKY model, depends on both the length of the sequence studied and the fraction of informative polymorphic sites, we computed the regression between the number of sites and the likelihood for sequences of same size, then estimated a score as the deviation from that prediction. Hence, a phylogenetic score of 0 reflects a region for which the likelihood of the species tree equals the average across all the genome. A negative score reflects a lower than average likelihood, i.e., the phylogeny is affected more than average by recombination.
We simulated 2 million, 3 kbp sequences under a neutral coalescent framework with pure gene conversion using the MS software [110]. All simulations had different values of the per-base rate of mutation (theta), the per-base gene conversion rate (Cgc) and the average tract length (Lgc) (assuming a geometrical distribution). For each of these simulations, statistics of linkage disequilibrium specific to the gene conversion signature were calculated as described elsewhere [39]. Basically long distance and short distance linkage disequilibrium are measured for pairs and triplets of sites. Since we had previously estimated fairly small gene conversion tract lengths [42], we used window sizes of 1 kbp, 0.2 kbp and 0.1 kbp, instead of the larger default values.
Using ABCest software [41], an Approximate Bayesian Computation method, we estimated these parameters for all the genes of the genome and all the 3 kbp sliding windows along the genome alignment with a step of 500 bp. To assess the reliability of the method we tested it on 1500 new simulations. The Pearson correlation between the observed and estimated ratio Cgc/theta was very high (0.897, 0.885 for the log transformed values) and 92% of simulations provided a 95% confidence interval around the estimated value encompassing the true value. Tract length, Lgc, provided quite large 95% confidence intervals so even if 92% of simulations encompassed the real value in this interval, the Pearson correlation between observed and estimated value was lower: 0.585 (0.676 for the log transformed values). Hence, this approach provides adequate estimates of the parameters and once the 2 million simulations have been performed, it allows a rapid (several seconds) estimation of the parameters for each dataset.
To study how gene conversion affected the phylogenetic reconstruction process, we modified the MS software [110] to allow 25 kbp of sequences evolve in a pure gene conversion model, but maintaining 1 nucleotide without any conversion so that its history reflects the history of the chromosomal backbone. We then compared with various methods (see Phylogenetic analyses section) the topology of the phylogenetic tree as reconstructed with PHYML [103] from the 25 kbp, as evolved along MS-derived local topologies under the HKY model with Seq-Gen [111], with the true history of the non-recombining last nucleotide, as directly extracted from MS.
We used the function “ACE” (package “APE” in R [112]) to estimate ancestral character states for continuous (genome size) and discrete (presence or absence of genes) characters on all branches of tree involving these taxa. For continuous characters we used a Brownian motion model in which characters evolve following a random walk. This model was fitted by least squares [113]. We estimated ancestral discrete characters by maximum likelihood [114]. For this we built a matrix wherein the number of rows corresponds to the number of characters (i.e., 18 822 positional ortholog genes corresponding to the pan-genome) and the number of columns corresponds to the number of genomes (i.e., 1 E. fergusoni and 20 E. coli strains). The model has two character states (0 = absence of the gene, 1 = presence of the gene). Since genome sizes are relatively constant among the closely related genera Escherichia, Salmonella and Yersinia, we assumed a probability of insertion equal to the probability of deletion, i.e., we assumed that genomes are close to equilibrium in terms of genome size. Variations in size are thus seen as stochastic fluctuations associated with the insertion of certain large elements such as phages.
We used the reference phylogenetic tree and the phyletic pattern indicating the presence/absence of each gene (of the pan-genome) to infer the probability of presence of each gene in each internal node of the tree. For each such node a gene was considered as present if it had a probability of presence ≥0.5. The numbers of genes lost and gained, respectively, were then determined in the following way: if the gene was absent (vs. present) in a given node but present (vs. absent) in its ancestor, it was considered as gained (vs. lost) along the branch leading to the given node. Ancestral gene order was determined on all branches of tree using the parsimony criterion. Considering the internal node gene order, the numbers of acquisition and loss events was defined for sets of consecutive pairs of genes (by allowing gaps of 1 gene).
The number of events in each branch of the species tree was computed by reconstructing the relative order of the core genes in the ancestral genome by parsimony. We then combined in a single event the contiguous gains or losses of genes in the same branch, allowing gaps of 1 gene.
A mouse model of systemic infection was used to assess the intrinsic extraintestinal virulence of the available strains [23]. For each strain, 10 outbred female Swiss OF1 mice (3–4 weeks old, 14–16 gm) were challenged subcutaneously in the abdomen with a standardized bacterial inoculum (0.2 ml of Ringer solution with 109 cfu/ml of log-phase bacteria). Mortality was assessed over 7 days post-challenge. In this model system, lethality is a rather clear-cut parameter and, based on the number of mice killed, almost all strains were classified as non-killer (<2 of 10 mice killed) or killer (>8 mice killed) [32].
|
10.1371/journal.pgen.1004573 | Genome Wide Association Studies Using a New Nonparametric Model Reveal the Genetic Architecture of 17 Agronomic Traits in an Enlarged Maize Association Panel | Association mapping is a powerful approach for dissecting the genetic architecture of complex quantitative traits using high-density SNP markers in maize. Here, we expanded our association panel size from 368 to 513 inbred lines with 0.5 million high quality SNPs using a two-step data-imputation method which combines identity by descent (IBD) based projection and k-nearest neighbor (KNN) algorithm. Genome-wide association studies (GWAS) were carried out for 17 agronomic traits with a panel of 513 inbred lines applying both mixed linear model (MLM) and a new method, the Anderson-Darling (A-D) test. Ten loci for five traits were identified using the MLM method at the Bonferroni-corrected threshold −log10 (P) >5.74 (α = 1). Many loci ranging from one to 34 loci (107 loci for plant height) were identified for 17 traits using the A-D test at the Bonferroni-corrected threshold −log10 (P) >7.05 (α = 0.05) using 556809 SNPs. Many known loci and new candidate loci were only observed by the A-D test, a few of which were also detected in independent linkage analysis. This study indicates that combining IBD based projection and KNN algorithm is an efficient imputation method for inferring large missing genotype segments. In addition, we showed that the A-D test is a useful complement for GWAS analysis of complex quantitative traits. Especially for traits with abnormal phenotype distribution, controlled by moderate effect loci or rare variations, the A-D test balances false positives and statistical power. The candidate SNPs and associated genes also provide a rich resource for maize genetics and breeding.
| Genotype imputation has been used widely in the analysis of genome-wide association studies (GWAS) to boost power and fine-map associations. We developed a two-step data imputation method to meet the challenge of large proportion missing genotypes. GWAS have uncovered an extensive genetic architecture of complex quantitative traits using high-density SNP markers in maize in the past few years. Here, GWAS were carried out for 17 agronomic traits with a panel of 513 inbred lines applying both mixed linear model and a new method, the Anderson-Darling (A-D) test. We intend to show that the A-D test is a complement to current GWAS methods, especially for complex quantitative traits controlled by moderate effect loci or rare variations and with abnormal phenotype distribution. In addition, the traits associated QTL identified here provide a rich resource for maize genetics and breeding.
| Maize (Zea mays L.) is one of the most important food, feed and industrial crops globally. Grown extensively under different climate conditions across the world, maize shows an astonishing amount of phenotypic diversity [1]. Identifying the underlying natural allelic variations for the phenotypic diversity will have immense practical implications in maize molecular breeding for improving nutritional quality, yield potential, and stress tolerance.
With the rapid development of next generation sequencing and high-density marker genotyping techniques, there emerges tremendous interest in using association mapping to identify genes responsible for quantitative variation of complex traits [2]. The use of GWAS has been well demonstrated in model plants such as Arabidopsis [3] and rice [4]. In maize, we examined the genetic architecture of maize oil biosynthesis in 368 diverse maize inbred lines with over 1.06 million SNPs obtained from RNA sequencing and DNA array using the GWAS strategy [5]. Despite the great potential that GWAS has to pinpoint genetic polymorphisms underlying agriculturally important traits, false discoveries are a major concern and can be partially attributed to spurious associations caused by population structure and unequal relatedness among individuals in a given panel [6]. A number of statistical approaches have been proposed, among which the mixed linear model (MLM) is one of the popular methods that can eliminate the excess of low p values for most traits [6], [7]. However, Zhao et al. [8] performed GWAS using a NAÏVE model in each sub-population and MLM with inferred population structure as a fixed effect in the whole mapping panel of rice, and their results suggested that MLM may lead to false negatives by overcompensating for population structure and relatedness. To improve the MLM, some strategies to best utilize marker data have been proposed [9], [10]. The more we know about the genetics of a trait, the greater our power is to detect the rest of the genetic contribution. The problem is, of course, that we usually do not know what the causal loci are, and methods that try to identify them are prone to over-fitting [11]. Beló et al. [12] adopted the Kolmogorov–Smirnov (KS) test for association analysis in each subpopulation of the mapping panel and an allelic variant of fad2 associated with increased oleic acid level was successfully identified based on modest density markers. However, detailed instructions for the algorithm were not published. Most current GWAS methods lack the power to detect rare alleles and this has limited the application of GWAS, since rare alleles are common in maize diversity collections [1], [5]. Parametric tests of association are sensitive to SNPs with minor allele frequencies, which can artificially increase association scores. Balancing samples across population subdivisions can homogenize allele frequencies, elevating rates of globally rare variants that are common in certain subdivisions [5].
In this study, 513 diverse maize inbred lines [13], representing tropical/subtropical and temperate germplasm, were genotyped by MaizeSNP50 BeadChip containing 56,110 SNPs [14]. RNA sequencing (RNA-seq) was performed on 368 of these 513 lines and 556,809 high quality SNPs with a minor allelic frequency greater than 0.05 were obtained [5], [15], [16]. Seventeen agronomic traits were systematically phenotyped for the 513 lines under multiple environments and seasons (see Materials and Methods). The objectives of this research were (1) to explore an efficient imputation method to infer missing genotypes for the 145 inbreds that were only genotyped by SNP-chip (low density), not by RNA-seq (high density); (2) to develop a powerful statistical method for GWAS to identify robust QTL for complex agronomic traits in maize; and (3) to methodically analyze the underlying genetic architectures of the 17 agronomic traits in the diverse maize association mapping panel.
A brief description of each trait, its acronym, and evaluation methodology was summarized in Supplementary Table S1. All of the 17 traits in the 513 maize inbred lines were in accordance with a normal distribution (Figure S1A, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14, S15, S16, S17A). But the phenotype of each trait showed distinct differences among four subgroups (Figure S1B, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14, S15, S16, S17B). Analysis of Variance (ANOVA) showed that population structure explained 39.5% of phenotypic variation (PVE) for tassel main axis length, which was the highest among the 17 agronomic traits included in best linear unbiased prediction (Table S2), indicating vulnerability of this particular trait to the population structure and variable sensitivity of different traits to population structure. Furthermore, heritability (h2) was highest (0.683) for tassel main axis length (TMAL), while the lowest heritability (0.386) was observed for kernel number per row (KNPR) among the traits. Pair-wise Pearson's correlation coefficients of the 17 traits revealed that phenotypes within a category were more correlated. The values ranged from 0.001 between kernel width and plant height to 0.95 between days to anthesis and heading (Figure S18). These results indicated that all the tested lines possessed significant genetic variability and can be used for further genetic analyses.
The whole panel, 513 maize inbred lines, was genotyped using the MaizeSNP50 BeadChip containing 56,110 SNPs (Illumina). RNA sequencing was performed on immature seeds for 368 out of the 513 maize inbreds using 90-bp paired-end Illumina sequencing, resulting in 2,445.9 Gb of raw sequencing data. 556,809 high quality SNPs obtained by combining the two genotyping platforms (RNA-seq and SNP array) [5], [16] were used in the study. For the additional 145 maize lines, the genotype calls of unique loci from the integrated SNP data were projected based on regions of IBD to physical maps constructed using 56110 SNPs, and then high-density markers with more than 0.5 million SNPs were obtained for all the lines. Out of 56,110 SNPs from MaizeSNP50 data set, 49728 SNPs overlapped with the integrated SNPs data based on their physical positions (B73 RefGen_v2). The 49,728 common SNPs were regarded as core or frame markers for projection based on IBD regions. In order to evaluate the performance of IBD [17] based projection, training and validation datasets were established for chromosome 1, which had 7818 core markers from Illumina Maize SNP50 and 88581 SNPs from the integrated data set. The genotypes for one maize line with RNA-seq data in IBD regions were assigned to the matched target line without RNA-seq data for each SNP. The projection accuracy was calculated by comparing inferred genotypes of 368 lines with their real genotype obtained from RNA-seq. In addition, KNN algorithm [4] which infers a large number of missing genotypes generated from low-coverage genome sequencing was used to impute the missing genotypes of the unique loci from RNA-seq SNP data based on 49728 frame markers. Single method analysis, either IBD based projection or KNN algorithm, cannot achieve both optimal accuracy and coverage (see Materials and Methods). However, the combination of IBD based projection and KNN seemed effective to infer a large number of missing genotypes. In order to optimize the set of imputation parameters, a simulation was performed on chromosome 1 in 368 lines (Figure S19). The simulation result on chromosome 1 in 368 lines indicated that the missing rate was reduced from 91.6% (1–7,818/88,581) to 12.8%, with an accuracy rate 96.6% (Table S3). The optimal parameter combination (IBD: SNPs number≥150 in 5 Mb window size; KNN: w = 20, k = 6, p = −7, r = 1) was then used to impute the missing SNPs for the remaining 145 inbred lines, resulting in an 85.5% filling rate. Therefore, our approach combining SNP-chip data and RNA-seq SNP data with an effective projection procedure permits the quick construction of a high-density physical map and integration of SNPs from RNA-seq data set onto the whole population. This approach is also applicable to other genomes and genotyping data from different platforms for a variety of downstream analyses.
The 368 maize inbreds with 556,809 SNPs genotyped by RNA-seq and Maize SNP50 array were defined as Data set 1. In addition, Data set 1 and 145 maize inbred lines with joint IBD-based projection and KNN imputed genotypes were defined as Data set 2 together. The 513 maize inbreds with Maize SNP50 array genotyped were defined as Data set 3. To evaluate the reliability of imputed genotypes for 145 inbred lines, GWAS was performed using MLM, with both Data sets 1, 2 and 3 focusing on kernel oil concentration, which has been thoroughly analyzed in our previous study [5].
For GWAS performed using MLM with both Data sets 1 to 3, a total of 26, 32 and 8 significant loci were identified in Data sets 1 to 3, respectively, at the Bonferroni-corrected threshold (−log P>5.74, α = 1,) (Figure 1). Almost all strong signals identified in data sets 1 and 3 were also identified in data set 2 (Table S4). More interestingly, we identified six additional significantly associated loci in dataset 2 (−log P>5.74, α = 1), including the phosphoinositide 3-kinases gene (PI3Ks) and the phosphatidylinositol transfer protein, which is known to be involved in the oil concentration trait [18] (Figure 1, Table 1, Table S4). This suggests that GWAS carried out using the imputed genotypes with a larger population (n = 513) increased the statistical power compared to the analyses of RNA-seq genotyped SNPs with the smaller population size (n = 368) or low density as DNA array SNPs with the same population size (n = 513).
GWAS for 17 agronomic traits using MLM was conducted with Data set 2 and the results are summarized in Figures S1C, D, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14, S15, S16, S17C, D. A total of 19 significant SNPs from 10 loci were identified for five traits (ear leaf width, ear length, kernel width, plant height and tassel main axis length) (Table 2). No significant SNP was found to be associated with the other 12 tested traits at the Bonferroni-corrected threshold (−log P>5.74, α = 1). If we set the Bonferrroni-corrected threshold as −log P>7.05 (α = 0.05), no SNPs were significant for all the 17 traits. It may be too strict to use 0.05/n as the cutoff since not all the markers are independent. One thousand permutation tests were conducted for three typical traits with different level of population structure (kernel width, ear height and days to heading) (Table S2). The results showed the cutoff value at α = 0.05 is quite similar (Table S5) with 0.05/n.
The A-D test [19] is a nonparametric statistical method and a modification of the KS test [12], [20] that gives more weight to the tails of the distribution than the KS test. Since the identified loci were much less numerous than expected using the MLM method, the data set was reanalyzed using the A-D test. The same three traits (kernel width, ear height and days to heading) were used to perform 1000 time permutation tests to determine the cutoff values. The results showed the cut off value at α = 0.05 varied around the Bonferrroni-corrected threshold as 0.05/n (Table S5). To simplify the procedures, we used the uniform cutoff (−log P>7.05, α = 0.05) for further analysis. Flowering time is an important and well-studied trait, and many QTL or candidate genes have been identified [21], [22]. Recently, several studies have confirmed that ZmCCT is the gene underlying the major QTL affecting flowering time on chromosome 10 [22], [23]. Taking flowering time as an example, it provides a good opportunity to test whether A-D test is a feasible GWAS method for agronomic traits or not. Using the A-D test, we identified 30 loci associated with days to heading in Yunnan 2010. Around 20% of 30 loci were located within a QTL support interval reported in NAM population [22]. If the significant loci are randomly distributed in the genome, the probability by chance is equal to the ratio between the whole-length of QTL interval and the whole genome length (12%), which represents an almost twofold enrichment compared with the 12% expected by chance. A strong association (−log10 (P) = 7.59) was identified in 1.7 Kb upstream of ZmCCT (Figure 2A). Four other loci seem to be strong candidates including: one homologous gene (CIB1) [24] shown to be involved in the regulation of flowering time in Arabidopsis, two homologues containing CCT domain that was demonstrated as key photoperiod regulatory gene in plants [25], and one locus previously shown to affect flowering time in maize (Id1) [26] (Figure 2A). Using the MLM method, we were only able to identify the marginally significant association for ZmCCT (−log10 (P) = 5.64) and there were no strong signals in other genome regions (Figure 2B). Therefore, A-D test could be a more appropriate GWAS method for agronomic traits and we performed GWAS using the A-D test in each subpopulation of Data set 2 without controlling of population structure for all tested traits. The total number of unique SNPs significantly associated with the 17 traits was 678, of which 310 represented unique loci (Figure S1E, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14, S15, S16, S17E). The numbers of significant SNPs associated with different traits ranged from 1 (Cob Weight) to 35 (Tassel branch number and Days to silking). For plant height, a total of 107 loci were identified at the Bonferroni-corrected threshold (−log P>7.05, α = 0.05) (Table S6, S7). About 10% of the loci were detected to affect two or more different traits that were consistent with the observed correlations among the measured traits (Figure S18). There were 101, 71 and 171 loci detected in three subpopulations: SS (subpop-1), NSS (subpop-2) and TST (subpop-3), respectively. A reasonable number of spurious associations should be existed in the detected loci since the population structure is not properly addressed in each subpopulation. Genomic control [27]–[30] is a popular method to control the population stratification and cryptic relatedness that was applied to adjust A-D test statistic in each subpopulation in present study. In total, 19 loci were significantly associated with 13 traits at the Bonferroni-corrected threshold (−log P>5.74, α = 1) (Table S7).
To further examine the nature of statistically significant associations, we examined the phenotype distributions of individuals carrying each allele. Interestingly, some associations that differed in the width of the phenotype distribution but which had nearly identical trait means were found to be highly significant by the A-D test but not significant by MLM. Figure 3 illustrated significant loci for ear height with nearly identical trait means (Figure 3A) and significant loci for ear leaf width with an obvious shift of the means (Figure 3B). In total, 14.6% of significant loci identified by A-D test do not have an obvious shift of the mean between the two alleles (t-test, p>0.05) (Table S7). In this case the differences between distributions are real, but the corresponding genetic markers would not be useful in breeding if the objective is to change the phenotypic means.
Causal allele frequencies and trait distributions are the main factors that affect association mapping efficiency [1], [31]. GWAS data were simulated by adding phenotypic effects to real genotypic data considering the population structure and epistasis from MaizeSNP50 BeadChip [13] under three scenarios: a normal distribution model, an abnormal distribution model caused by uncertain effectors like phenotyping errors and an abnormal distribution model caused by a larger effect QTN with rare alleles (Methods). We compared our noticed A-D test with three other mapping methods: Kruskal-Wallis (K-W) test and linear model (LM) which does not correct for population structure; MLM which corrects for population structure and kinship. Statistic power of the four methods were compared under the same level Type I error. For each method, QTNs were considered to be detected if their P value were below the threshold determined by 1,000 times permutation.
The results for the three simulation schemes are shown in Figure 4 and can be summarized as follows: First, MLM has more (Figure 4A, C, I) or similar (Figure 4G) power among the four methods for major QTNs in schemes 1 and 3. Second, regardless of the allele frequency, nonparametric methods usually have greater power than LM and MLM for moderate QTNs (Figure 4D–F, J–I). Third, A-D test is more powerful than K-W test in terms of QTNs with rare alleles (Figure 4J–I). Fourth, nonparametric methods are much more powerful than parametric methods in scheme 2 (Figure 4B, E, H, K) that the phenotype has an abnormal distribution model caused by uncertain effectors.
We compared our mapping results for 17 agronomic traits with QTL identified using different linkage segregation populations and with previously reported known genes. Loci identified by the A-D test which overlap with previously identified genes and loci mapped in biparental populations are summarized in Table S7. Considering the large confidence interval of previously reported QTL, 3 independent RIL populations genotyped with high-density SNP markers were used to conduct QTL analysis for 3 traits (kernel width, ear length and kernel number per row) of the 17 traits tested. For the compared traits, 9 loci (20% of the detected loci) identified by the A-D test were within the QTL confidence interval. One example is a major kernel width QTL which was mapped in chromosome 7 with BK/Yu8701 RILs and explains 18.7% of phenotypic variation (Figure 5B). Within the QTL interval, significant SNPs- kernel width association was detected (Figure 5A). Six candidate genes: GRMZM2G354539, GRMZM2G052893, GRMZM2G052817, GRMZM2G354525, GRMZM2G052610, and GRMZM2G052509 are found in the associated interval. An expression quantitative trait locus (eQTL) was detected for one (GRMZM2G052509, −log10 (P) = 10.16) of the six annotated genes, which can therefore be regarded as a candidate gene for further study. The second example of overlap included the SNP chr2.s_1972207(C/G) with −log10 (P) = 9.06and SNP chr2.s_1972176(C/G) with −log10 (P) = 8.49 which were both significantly associated with ear length, and a QTL affiliated with ear length identified in B73/By804 RIL near the associated peak (Figure 5D–F). SNP chr2.s_1972207 (C/G) and SNP chr2.s_1972176 (C/G) were the only two of the 36 SNPs within the gene GRMZM2G061877, which encodes a DHHC zinc finger domain containing protein, and both of them are in the CDS region. SNP chr2.s_1972176 (C/G) makes no difference to the translated protein sequence, while SNP chr2.s_1972207 (C/G) results in a change of Isoleucine to Methionine. Several zinc finger proteins that play important roles in maize inflorescence development, for instance transcription factors RA1, RA2 and RA3 in ramosa pathways [32], have been identified. In rice, a zinc finger transcription factor DST directly regulates OsCKX2 expression in the reproductive meristem leading to OsCKX2 regulated CK accumulation in the shoot apical meristem (SAM) and, therefore, controls the number of the reproductive organs; the dst mutant leads to lower plant height and longer rice panicle length [33]. These zinc finger genes are functioning as transcription factors. Since the DHHC protein domain product of GRMZM2G068177, which was strongly suggested as a candidate gene for the regulation of ear length, acts as an enzyme, this may suggest a novel function of zinc finger proteins in monocot reproductive organ development. However, further work is needed to test this hypothesis. The third example, one QTL located on chromosome 1 using K22/Dan340 RIL population which explains 11.8% of phenotypic variation for kernel number per row, also overlapped with significant association signals (Figure 5G–I) Four candidate genes: GRMZM2G088524, GRMZM2G022822, GRMZM2G108180 and GRMZM2G052666, located in a 200 kb window around the significant signals, were predicted.
The genome-wide imputation of genotypes has attracted much attention given its broad applicability in the GWAS era. There are a number of methods for imputing missing genotypes, but many factors influence the accuracy of imputed genotypes [34], [35]. In this study, we proposed a two-step method combining IBD based projection and KNN algorithm to infer missing genotypes, resulting in 96.6% accuracy and 85.5% genome coverage in the tested samples. Considering that the missing genotypes consist of over 91.6% of our raw data set, this level of accuracy is acceptable. Compared with other methods [4], [17], [34], [35], the two-step method has its advantages. Imputation based only on IBD regions ensures high accuracy but a relatively low coverage rate. KNN algorithm has been proven to be a good strategy for sequencing data [4], however, it alone does not represent the true similarity of the inbred lines due to low density of frame markers and rapid LD decay in maize in our study. Therefore, we first used IBD based imputation to increase marker density, and then the KNN algorithm was used to infer the missing genotype, leading to high coverage rate and imputation accuracy. Imputation error is often caused by ignoring recombination and mutation within IBD regions. In addition, if an inbred line with low density markers share a region with two or more inbred lines with high density markers and the missing genotypes are inferred on the basis of only one of these lines, there is a high risk of error since the accuracy of the projection depends on the identity between the projected and chosen lines. Reanalysis of GWAS for kernel oil concentration revealed consistent results and a higher detection power. Six more associated loci were identified, most likely due to the increase in sample size. The implication is that mapping resolutions are enhanced by extracting moderately more information from the genome and expanding sample size.
The detection of loci controlling complex traits using GWAS has flourished and numerous statistical approaches for GWAS analysis in plants have recently been described [14], [36]–[40]. Linear statistical models like ANOVA, general linear model (GLM), and MLM establish significance cutoff by relying on the assumption that target traits have normal distribution. However, sometimes phenotype distribution in the moderate plant population is not normal in the tails that may be due to the population size, field experiment such as phenotyping errors, or genetic effects [31]. Based on our simulated data, nonparametric methods including A-D and KW tests usually have greater power than LM and MLM for abnormal phenotypes, rare alleles and moderate QTNs. It also implies that A-D and KW tests should perform well to detect the shifts of distribution as well as changes in the shape of distributions [31]. A-D test possesses advantages than K-W test in the detection of QTNs with rare alleles. However, MLM performs better than A-D and KW test for the major QTNs especially those with common alleles. However, we need to keep in mind that population structure of the studied samples is the key confounder for GWAS. In the measured 17 agronomic traits of present study, we observed the phenotypic variation explained by population structure ranged between 0.9% and 32.3% (Table S2). In the A-D test, we didn't account the confounding by population structure in the subpopulation that may lead to false-positive findings. Genomic control is a good alternative for controlling the statistic inflations [27]–[30], different inflation factors were observed in different traits and different subpopulations in present study (Table S7). We detected 19 loci significantly associated with 13 traits at the Bonferroni-corrected threshold (−log P>5.74, α = 1) (Table S7) using genomic-control () to adjust our real phenotype test statistic from A-D test. However, we also need to be careful that the adjusted −Log P might be over corrected, since A-D test has already controlled part of the population structure and genomic control method is affected significantly by the true association signals , even for the agronomic traits may involve a larger number of loci with small effects [21], [36], [37], [41]. And the influence of epistatic genetic effect to the genomic control is still not explored [27]–[30]. Another thing need to be noticed is testing within subpopulations (A-D test) and across the whole panel with controlling the population structure (MLM) are different. Testing within subpopulations changes allele frequencies of background alleles and therefore possibly changes the epistatic interactions that are mapped in an additive manner within subpopulations but were not mapped across populations.
In general, A-D test could be a good complement to current popular GWAS methods. As each method owning its own advantages, the preliminary understanding of the traits studied is needed for choosing GWAS methods or trying different GWAS methods would be helpful especially for those studies only few or none significant signals were identified by using only one method. In this study, we performed GWAS using both MLM and A-D test for 17 agronomic traits. In total, 18 overlapped regions were detected by the two approaches (Table S7). The A-D test also showed high concordance with previous studies in identifying a higher number of QTL related to agronomic traits.
Our noticed nonparametric statistical approach is robust with respect to non-normality, similarly to the KS test [12]. The KS test tends to be more sensitive around the median value and less sensitive at the extreme ends of the distribution. Thus, the KS test is not always appropriate for calculating the significance of data sets which differs at the tails of the probability distribution, while the median remains unchanged [31]. The A-D test improves upon the KS test because it has more sensitivity towards the tails of the pooled sample. More importantly, the performance of the A-D test for small samples is quite good, as demonstrated by numerous Monte Carlo simulations [19]. This means that, for complex traits, the A-D test can make a good use of SNPs that have minor allele frequency and keep detection ability to the relatively small effect loci. At the same time, it is important to recognize that there are always limitations to what can be achieved using statistics. It seems that A-D test does not work well for all traits. Interestingly, we identified 14.6% associations by A-D test that differed in the width of the phenotype distribution but which had nearly identical trait means (Figure 3A). In these cases the differences between distributions are real, but the corresponding genetic markers would not be useful in breeding if the objective is to change the phenotypic means. Instead, the associations appear to represent allelic differences in the apparent trait stability. Therefore, to confirm candidate loci, it is necessary to check both frequency distribution and normality of the distribution curves (Figure 3). Several studies in humans have confirmed that using multiple methods for statistical inference critically enables the interpretation of results and engenders stronger candidates for experimental follow-up [42].
We identified some genes affecting important agronomic traits in maize that are very good candidates for future detailed analysis, for allele mining to identify functional variation, and for marker development. As whole genome sequences become available for many crop species including maize, as well as for multiple genotypes of the same species through resequencing, along with cost-effective high-throughput genotyping systems and the next generation of sequencing technologies, GWAS becomes practical and its use in plant breeding will allow the manipulation of many traits at the whole-genome level. Association mapping using a set of global diverse breeding germplasm and high-throughput SNP markers, as shown in this study, provides high-resolution dissection of the genetic architecture of complex traits. This knowledge in turn will be useful not only for designing marker-assisted selection strategies but also for optimizing conventional breeding systems.
A total of 513 maize lines with tropical, subtropical and temperate backgrounds representing the global maize diversity were employed for genome-wide association mapping in this study. All maize inbred lines have been well described in previous studies [13], [15] and the 513 maize lines were classified into four subgroups based on population structure Q matrix: Stiff stalk (SS) with 112 lines, Non-stiff stalk (NSS) with 116 lines, Tropical-subtropical (TST) with 258 lines, and an admixed group with 27 lines (detailed information also can be downloaded at (www. maizego.org/resource). A Randomized complete block design with one to two replications was used for field trials in five environments, including Ya'an (30°N, 103°E), Sanya (18°N, 109°E), Yunnan (25°N, 102°E) in 2009, Guangxi (23°N, 110°E) and Yunnan (25°N, 102°E) in 2010. A row length of 3 m was used for each line including 11 plants plot−1 with 25 cm plant to plant and 60 cm row to row distance. Five randomly selected plants were used for phenotypic data acquisition in each line and the mean data in each replication was used for phenotypic analysis. A total of 17 economically important traits were phenotyped (Table S1). These traits were divided into three categories: morphological attributes (plant height, ear height, ear leaf width and length, tassel main axis length, tassel branch number, and leaf number above ear), yield related traits (ear length and diameter, cob diameter, kernel number per row, 100-grain weight, cob weight and kernel width), and maturity traits (days to heading, anthesis, and silking). Best linear unbiased predictions (BLUP) for each line across five environments were calculated using the MIXED procedure in SAS (Release 9.1.3; SAS Institute, Cary, NC), and employed for evaluating trait variation in the association panel.
Imputation methods have not been developed to deal specifically with low density of SNP marker data. Of the available imputation models, identity by descent (IBD) based projection [17] and the k-nearest neighbor algorithm (KNN) [4] seemed to effectively infer a large number of missing genotypes. To assess the performance of IBD based projection, preliminary tests for chromosome 1 in 368 maize lines were conducted. We removed genotype data without frame SNPs and then compared the observed genotypes with those generated by projection. The number of IBD regions with consecutive SNPs for 368 lines varied from 1 to 285 on chromosome 1, and projection accuracy, defined as the percentage of correctly projected genotypes ranged from 74.8% to 99.1%, with an average of 92.6% (Table S3). The mean error ratio pooled over in 32,015 IBD regions on chromosome 1 for the 368 maize lines was also calculated (Figure S19A, B), and gradually declined with the increasing number of identical SNP and size in IBD regions. Preliminary testing suggested that IBD regions with 150 consecutive SNPs and a size of 5 Mb or more were highly conserved in maize and error rate for projection was well controlled within 5% (Figure S19A, B). The number of qualified IBD segments ranged from 0 to 20 for 368 lines and coverage rate, defined as the percentage of projected genotypes, accounted for 61.99% of genomic regions on chromosome 1, with projection accuracy increased from 92.59% to 96.62% on average (Table S3). Therefore, IBD based projection for regions with 150 consecutive SNPs and a size of 5 Mb were applied for integration of SNPs from RNA-seq data set onto the 145 additional maize inbred lines. Alternatively, the K-Nearest Neighbor (KNN) algorithm was also used to enrich the physical map of each line constructed by 56,110 SNPs from MaizeSNP50 chip by inferring the missing genotypes of the unique loci from RNA-seq SNP data. In the preliminary test, this method was efficient and the imputation accuracy and coverage rate for 368 lines were 97.48% and 75.35%, respectively (Table S3). The IBD based projection and KNN imputation revealed high inferred accuracy; however, the coverage rates were relatively low, with an average of 62% and 75%, respectively. In order to increase the coverage rate and keep high imputation accuracy, IBD based projection and KNN algorithm were combined to infer missing genotypes. The IBD method can provide more frame SNPs for the KNN algorithm, and simultaneously the KNN algorithm compensates for the weakness of the IBD method in coverage rate. About 38% of the genotypes were missing after prediction of IBD regions with 150 consecutive SNPs and 5 Mb size, and then the KNN algorithm was used to impute the missing data, resulting in 95.8% of accuracy for the missing data. The joint IBD based projection and KNN imputation of the genotypes of 368 lines increased coverage rate from 62% to 87.2%, with a total accuracy of 95.9% in the preliminary test. The projection accuracy was also affected by heterozygosity of each line, which increased from 95.88% to 96.60% after excluding 44 lines with more than 10% heterozygosity. The joint IBD based projection and KNN imputation that performed well in the preliminary test was used for the integration of SNPs from the high density SNPs data set onto 145 maize lines genotyped by 56110 SNPs. For 145 maize lines, 54.18% and 32.28% of loci across 10 chromosomes were inferred through IBD based projection and subsequent KNN imputation, respectively. As a result, 85.46% of loci for the whole maize genome were filled. The average density for the whole panel increased from 20 SNPs to more than 200 SNPs per Mb.
The linkage analyses of ear length, kernel number per row, and kernel width were performed in three recombination inbred line (RIL) populations, BY804/B73 (197 individuals), K22/Dan340 (197 individuals), and BK/Yu8701 (165 individuals). All the RIL lines and their parents were genotyped using Maize SNP50 assays (Illumina) containing 56,110 SNPs [14]. The phenotype of BK/Yu8701 in Henan 2011 and BLUP value from 5 environments of BY804/B73 and K22/Dan340 were used. QTL mapping using the composite interval mapping method [43] was performed in the package QTL cartographer version 2.5 [44].
ANOVA, correlation, and repeatability analyses for 17 agronomic traits were conducted using SAS software (Release 9.1.3; SAS Institute, Cary, NC). Heritability analysis and association analysis for the 17 agronomic traits in Data set 2 were conducted by MLM using TASSEL [45] software package. The observed p values from marker-trait associations were used to display Q-Q plots and Manhattan plots, using R. Permutation tests were used to determine the cutoff for GWAS. Considering the computation time, we only choose three typical traits with different population structure effects (kernel width, ear height and day of flowering time) as examples. The results showed that the cutoff values are similar with the Bonferroni correction. To simplify the procedures, we use the uniform Bonferroni-corrected thresholds at α = 1 and α = 0.05 as the cutoffs. When performing n tests, if the significance level for the entire series of tests is α, then each of the tests should have a probability of P = α/n. When the numbers of markers was 556809 SNPs, at α = 1 and α = 0.05, the Bonferroni-corrected thresholds for the p values were 1.796×10−6 and 8.95×10−8, with corresponding −log p values of 5.74 and 7.05, respectively. Regression estimator () of Genomic Control inflation factor was used [28]. Percentage of PVE by associated SNPs was calculated by ANOVA. Informative SNPs and candidate genes at the identified loci for the corresponding traits were from public maize genome data set B73 RefGen_v2.
To compare the power and FDR of A-D test, Kruskal-Wallis (K-W test) test, linear model (LM) and mixed linear model (MLM), three schemes with different phenotype distribution were simulated by considering the QTN effects and allele frequency.
Scheme 1 was used to simulate a normal distribution phenotype with the contribution of population structure, additive genetic effect, epistatic genetic effect and residual effect [6]. The population structure and epistasis explained 10% of the total phenotypic variation, respectively. The additive effect was the sum of all additive effects for 20 causal QTNs. For approaching the real genetic architecture, we set 20% major QTNs explaining 30% of the sum of all assigned genetic effect and 80% moderate QTNs explaining 70% of the sum of all assigned genetic effect. Half of major and moderate QTNs were rare alleles (MAF = 0.05–0.1) and half were common alleles (MAF = 0.25–0.45). Larger genetic effects were assigned to the rare alleles QTNs to ensure them could explain the same proportion of phenotypic variation as common alleles QTNs. The ratio of assigned genetic effects between rare alleles QTNs (at MAF = 0.075) and common alleles QTNs (at MAF = 0.35) was calculated based on . The genetic effect was assigned to all SNPs, one at a time [6]. The proportion of the additive effect was defined by narrow-sense heritability which is the proportion of additive variance over the total variance, and was examined. The residual effect followed a normal distribution and had a variance to satisfy the contributions from additive and epistatic effects at the designated level [6].
Scheme 2 was used to simulate an abnormal distribution phenotype with a long tail on one side. On the basis of scheme 1, 10% of lines were randomly selected and added an extra residual effect (1 to 6 fold standard deviation of the phenotype). All the others were same.
Scheme 3 was designed to simulate an abnormal distribution phenotype caused by a larger effect background rare QTN. The additive effect was still the sum of all additive effects for 20 causal QTNs. 1 background QTN, 3 major QTNs and 16 moderate QTNs explaining 25%, 20%, 60% of the sum of all assigned genetic effect respectively. The population structure effect, epistatic effect and residual effect were consistent with scheme 1.
Simulations of the phenotypes were repeated 500 times in all schemes. All simulated phenotypes had been analyzed with the four methods presented in the main text. 1,000 permutations had been done separately for the four methods to obtain the threshold at different type I error risk.
The Anderson-Darling two-sample procedure assumes that the two samples have a continuous distribution function and we are interested in testing the null hypothesis that the two phenotype samples divided by two alleles of one SNP have the same distribution, without specifying the nature of population:
The test procedure is as follows:
1. Calculate :
The computational formula for not adjusted for ties is,and the corresponding adjusted for ties is,where:
, indicates the two phenotype distribution function
k = 2; i = 1, 2
= data number in the ith sample; j = 1,2,…,
N = total number of two samples' individuals;
= data in the i sample and j observation within that sample
L = the number of unique data, where it will be less than n with tied data
z(j) = distinct values of all combined data ordered in ascendant way denoted z(1),z(2),…,z(L)
= number of values in the pooled sample equal to z(j)
= number of values in the combined samples less than z(j) plus one half of the number of values in the combined samples equal to z(j)
= number of values in the ith sample which are small than z(j) plus one half the number of values in this sample which are equal to z(j)
2. Calculate :
Under , the variance of is,with:where:
3. Calculate :
4. Refer to the upper α percentiles of the distribution table below, reject at significance level if exceeds the given point .If is outside the range of the table. Plotting the log-odds of versus , a strong linear pattern indicates that simple linear extrapolation should give good approximate p values.where:
URL. One R package (ADGWAS) for GWAS by Anderson-Darling test can be downloaded here: http://www.maizego.org/Resources.html
|
10.1371/journal.ppat.1002439 | HIV-1 Capsid-Cyclophilin Interactions Determine Nuclear Import Pathway, Integration Targeting and Replication Efficiency | Lentiviruses such as HIV-1 traverse nuclear pore complexes (NPC) and infect terminally differentiated non-dividing cells, but how they do this is unclear. The cytoplasmic NPC protein Nup358/RanBP2 was identified as an HIV-1 co-factor in previous studies. Here we report that HIV-1 capsid (CA) binds directly to the cyclophilin domain of Nup358/RanBP2. Fusion of the Nup358/RanBP2 cyclophilin (Cyp) domain to the tripartite motif of TRIM5 created a novel inhibitor of HIV-1 replication, consistent with an interaction in vivo. In contrast to CypA binding to HIV-1 CA, Nup358 binding is insensitive to inhibition with cyclosporine, allowing contributions from CypA and Nup358 to be distinguished. Inhibition of CypA reduced dependence on Nup358 and the nuclear basket protein Nup153, suggesting that CypA regulates the choice of the nuclear import machinery that is engaged by the virus. HIV-1 cyclophilin-binding mutants CA G89V and P90A favored integration in genomic regions with a higher density of transcription units and associated features than wild type virus. Integration preference of wild type virus in the presence of cyclosporine was similarly altered to regions of higher transcription density. In contrast, HIV-1 CA alterations in another patch on the capsid surface that render the virus less sensitive to Nup358 or TRN-SR2 depletion (CA N74D, N57A) resulted in integration in genomic regions sparse in transcription units. Both groups of CA mutants are impaired in replication in HeLa cells and human monocyte derived macrophages. Our findings link HIV-1 engagement of cyclophilins with both integration targeting and replication efficiency and provide insight into the conservation of viral cyclophilin recruitment.
| During infection HIV-1 enters the nucleus by crossing the nuclear membrane and incorporating itself into the host DNA by a process called integration. Here we show that the viral capsid protein gets tethered to a cyclophilin protein called Nup358, a component of the nuclear membrane gateways that allow transport between the cytoplasm and the nucleus. Altering the capsid protein so that it cannot use Nup358 prevents viral replication in macrophages, a natural target cell type for HIV-1. Intriguingly, these viral mutants are not less infectious in certain immortalised cell lines suggesting that in these cells nuclear entry is regulated differently. In this case similar to wild type virus, the mutant viruses integrate into host chromosomes but they integrate into different regions suggesting that the pathway into the nucleus dictates where the virus ends up in the host chromatin. We also show that another cyclophilin, the cytoplasmic protein cyclophilin A, influences the engagement of Nup358 as well as other proteins involved in HIV-1 nuclear entry. We hypothesise that HIV-1 has evolved to use cyclophilins so that it can access a particular pathway into the nucleus because alternative pathways lead to defects in integration targeting and viral replication in human macrophages.
| The ability to infect terminally differentiated cells of the monocyte-macrophage lineage is a conserved property of lentiviruses, including HIV-1 [1]. This process requires pre-integration complexes (PICs) to traverse the nuclear pore, though the molecular mechanism remains unclear. The HIV-1 proteins matrix, Vpr and integrase, as well as a DNA triplex at the central polypurine tract, have been proposed to contribute, but contrary evidence has been presented for each [2]–[5]. Gammaretroviruses such as murine leukemia virus (MLV) are dependent on cell division for infectivity and infect non-dividing cells inefficiently [6]. Characterization of HIV-1/MLV chimeric viruses has suggested a role for the HIV-1 capsid (CA) in nuclear entry [7]. Furthermore, certain HIV-1 CA mutants are selectively defective in arrested cells but not in actively dividing cells again implicating a role for CA in HIV-1 nuclear entry [8]–[10].
The nuclear pore complex (NPC), through which HIV replication intermediates must pass, consists of multiple copies of at least 30 different nuclear pore proteins (Nups). Nup358 is a large 358 kDa protein that constitutes the cytoplasmic filaments and has a C-terminal cyclophilin (Cyp) domain. It was first named Nup358 [11] but has also been called RanBP2 [12]. We use its original name Nup358 throughout this study. Several roles have been proposed for Nup358 involving cell cycle control, nuclear export, and transportin/importin dependent nuclear import (reviewed in [13]). In addition, Nup358 is a co-factor for HIV-1 replication, supporting nuclear entry of viral PICs and influencing target site preference for integration [14]–[17]. It has been unknown how the virus engages Nup358 and influences PIC traffic across the nuclear pore.
Here we demonstrate that HIV-1 CA binds directly to the Nup358 Cyp domain (Nup358Cyp) with an affinity within three fold of its binding of the monomeric cytoplasmic cyclophilin, CypA, which is known to be important during HIV-1 infection. We also demonstrate that CypA is important for directing HIV-1 into a nuclear entry pathway involving Nup358 and subsequent engagement of the nuclear basket protein Nup153, ensuring integration into preferred genomic loci. We report that altering CA interactions with Nup358 or CypA results in alterations in integration targeting preference, and reduced replication in macrophages. Our study provides the first evidence for direct interaction between HIV-1 CA and the NPC and suggests possible models for links between nuclear import, integration site selection and effective replication in primary human cells.
Several studies have shown that depletion of Nup358 reduces HIV-1 infectivity. We sought to define the HIV-1 determinant that confers its sensitivity to Nup358 depletion by studying infections with VSV-G pseudotyped viral vectors encoding GFP. Stable Nup358 depletion by transduction of HeLa cells with MLV or HIV-1 based shRNA expression vectors reduced HIV-1 GFP vector infectivity by 6- to 8-fold confirming Nup358's role as an HIV-1 cofactor [14], [17] (Figure 1A, B and Figure S1). We validated effective shRNA targeting by western blotting, using a Nup358 specific antibody (Figure 1B), as well as by co-transfecting the shRNA expression vector and a plasmid encoding GFP-tagged Nup358 into 293T cells (Figure S2). Studies on the role of Nup358 in HIV-1 replication have used M-group HIV-1 isolates [14]–[17]. In order to confirm the importance of Nup358 as a cofactor for other HIV-1 isolates we also tested the O-group HIV-1 virus MVP5180 [18], [19] as a distantly related HIV-1 and found that this too was sensitive to Nup358 depletion, suggesting that Nup358 use is a conserved feature of HIV-1 biology (Figure 1A and Figure S1C). We next tested whether the even more distantly related simian immunodeficiency virus from macaques (SIVmac) was sensitive to Nup358 depletion. In contrast to HIV-1, infectivity of SIVmac, was not reduced by Nup358 RNAi, suggesting species-specificity of Nup358 use (Figure 1A and Figure S1C). We next sought to identify the viral determinant for Nup358 RNAi sensitivity. Given that the HIV-1 capsid protein (CA) has been implicated in HIV-1 nuclear import [7], [8], [10], we tested whether the different sensitivities to Nup358 depletion between HIV-1 and SIVmac could be accounted by their different CA proteins. We exchanged CA coding regions between HIV-1 and SIVmac and analyzed infectivities of chimeric viruses on Nup358 depleted cells. Replacement of SIVmac CA with CA from HIV-1 [20] rendered the chimeric SIVmac sensitive to Nup358 depletion, while replacement of HIV-1 CA with SIVmac CA [21] rendered HIV-1 largely insensitive to Nup358 depletion (Figure 1C). For comparison, we examined the sensitivity of these viruses to transportin 3 (TRN-SR2) depletion, and confirmed that TRN-SR2 specific shRNA reduced infectivity of both HIV-1 (∼8 to 10-fold) and SIVmac (∼20-fold) (Figure 1A and Figure S1). MLV GFP vector infectivity was not affected by depletion of these proteins as reported previously, consistent with MLV's inability to traverse the nuclear pore and infect non-dividing cells (Figure 1A and Figure S1).
If Nup358 and TRN-SR2 facilitate nuclear entry of wild type HIV-1, then their depletion should inhibit HIV-1 infection at the level of nuclear import. We confirmed that 2-LTR circle products of HIV-1 were modestly reduced in abundance in the Nup358 or TRN-SR2 depleted cells, whereas late reverse transcript production was unaffected (Figure 1D) [22]. However, we observed that the ten-fold reduction in infectivity was greater than the two to four-fold reduction in 2-LTR circles, possibly explained by an integration defect increasing the amount of 2-LTR circles. To measure integration we infected Nup358 or TRN-SR2 depleted cells with HIV-1 GFP vector, grew the cells for 2 weeks and measured the number of integrated proviruses by Taqman qPCR. We observed that the reduction of integrated proviruses in Nup358 depleted cells (5-fold) was similar to the reduction of 2-LTR circles (4-fold) (Figure 1D and Figure S3). In contrast, the reduction of proviruses in TRN-SR2 depleted cells was significantly greater (50-fold), than the reduction in 2-LTR circles (2 to 3-fold). This observations may suggest that Nup358 depletion blocks HIV-1 at a step prior to nuclear import but after reverse transcription, whereas TRN-SR2 depletion imposes two blocks, one at the stage of nuclear import (reduction of 2-LTR circles) and a second at integration. However, we suggest caution in interpretation of 2-LTR circle assay as a measure of nuclear entry given that 2-LTR circles are non productive for infection and their formation may have different co-factor requirements. Importantly, replication of wild type NL4.3GFP-IRES was also impaired in Nup358 or TRN-SR2 depleted HeLa cells expressing CD4 (Figure 1E). Equivalent CD4 expression in these cells was confirmed by flow cytometry using fluorescent CD4 specific antibody (Figure S4). These data suggest that Nup358 and TRN-SR2 contribute to optimal viral nuclear entry, integration and eventually replication.
Nup358 contains a cyclophilin domain (Nup358Cyp) at its extreme carboxyl-terminus. The HIV-1 N-terminal CA domain (CANTD) resides on the surface of the virion core and recruits CypA to viral cores [23], [24]. The CA-dependent sensitivity of HIV-1 to Nup358 depletion led us to hypothesize that HIV-1 CANTD might also interact with Nup358Cyp in a similar manner to its interaction with CypA. To test this, we purified recombinant CypA and Nup358Cyp and measured binding to recombinant CANTD, using isothermal titration calorimetry (ITC) [25]. We found that the HIV-1 CANTD bound Nup358Cyp with a Kd of 16 µM, in a similar range to its Kd of 7 µM for CypA (Figure 1F) [25]. Surprisingly, the CypA inhibitor cyclosporine (Cs) did not prevent Nup358Cyp binding to HIV-1 CANTD whereas it did inhibit CypA binding. Whilst capsid interaction with both Nup358Cyp and CypA was entropically favourable, interaction with Nup358Cyp was more strongly entropically favourable than CypA. This does not markedly alter the affinity with respect to CypA as CANTD interaction with CypA is more enthalpically favourable than with Nup358Cyp. The different thermodynamic signatures between CypA and Nup358Cyp suggest that the two proteins do not form identical interactions. The entropic component of any interaction is the sum of changes in protein and solvent dynamics. Given that the ligand, CANTD, is the same in each experiment whilst Nup358Cyp and CypA comprise a single globular fold it is likely that the entropically favourable nature of both interactions is a consequence of releasing ordered water molecules upon complex formation. The larger entropic change associated with Nup358Cyp interaction may indicate a greater release of ordered water upon complexation, suggesting that the interface is larger than in CypA:capsid.
SIVmac CANTD did not bind Nup358Cyp (Figure 1F), and bound CypA with a very low affinity (∼800 µM) (Figure 1F), which becomes important below. The inability of Nup358Cyp to bind to SIVmac CANTD correlates with the insensitivity of SIVmac to Nup358 depletion in HeLa cells (Figure 1A).
To probe Nup358Cyp binding further we designed an HIV-1 inhibitor based on the simian restriction factor TRIMCyp. Owl monkey TRIMCyp blocks HIV-1 by binding incoming capsids via its CypA domain [26]. We replaced the TRIMCyp cyclophilin domain with human CypA or Nup358Cyp to make TRIMCypA and TRIMNup358. We found that both TRIMNup358 and TRIMCypA blocked HIV-1 infectivity (Figure 1G) whereas SIVmac infection was not restricted, as expected from the lack of binding (Figure S5). Importantly, Cs treatment only rescued infectivity from TRIMCypA but not from TRIMNup358, corroborating Cs sensitivity measured by ITC (Figure 1F). We confirmed similar expression levels of chimeric proteins by western blot (Figure 1G). These data are consistent with HIV-1, but not SIVmac, CANTD efficiently binding Nup358Cyp in the context of TRIMNup358 in the cytoplasm of infected cells. This suggests that HIV-1 PICs containing CA or possibly entire capsid cores can interact directly with a component of the NPC, providing insight into how HIV-1 contacts the nuclear pore during the process of nuclear entry.
Host proteins that interact with pathogens are often under positive selective pressure [27]. A higher rate of non-synonymous nucleotide substitutions (dN) than synonymous substitutions (dS) at a particular codon in an interspecies comparison provides evidence for such positive selection. We aligned Nup358Cyp DNA sequences from 12 different species (Figure S6) and performed an analysis of codon-specific selective pressures using the program Random Effect Likelihood (REL) implemented on the online version of the HyPhy package [28]. Despite overall strong negative selection across the Cyp domain, we found Nup358Cyp codon 61 to be positively selected at a statistically significant level (Bayes factor >50; Figure 2A). Indeed, residue 61 is extremely conserved as methionine across the whole vertebrate Cyp family except in Nup358Cyp. (Figure 2B, C). In the case of Nup358Cyp lower vertebrates encode the ancestral methionine, whereas higher vertebrates encode valine, leucine or isoleucine at this position. Fixation of the positively selected site appears to have occurred after the divergence of fish and tetrapods, since Nup358Cyp sequences from fish (e.g. Danio rerio) retain methionine at position 61 (Figure 2B, C). This observation suggests that Nup358 has been under selective pressure to evolve and that this has led to variation in the sequence at this position. The cyclophilin domain of Nup358 has been proposed to possess prolyl cis-trans isomerase activity, similarly to CypA [29], [30]. Assuming that the Nup358Cyp active site is homologous to that of CypA then according to the CypA structure (PDB:1FGL) residue 61 is located directly at the bottom of the active site suggesting that it might impact on substrate specificity (Figure 2D). To examine this further we made a TRIMNup358 mutant in which the valine at Nup358Cyp position 61 was changed to the ancestral residue methionine. TRIMNup358 V61M was no longer able to restrict HIV-1, suggesting that binding to HIV-1 CA is influenced by this residue (Figure 2E). Our results suggest that during evolution selective pressure, possibly from ancient pathogenic viruses, has driven the change of Nup358Cyp position 61, altering substrate specificity in higher vertebrates. In turn HIV-1 is adapted to use this modified protein for nuclear entry in humans.
If binding of HIV-1 CA to Nup358 is important for HIV-1 infectivity, then amino acid substitutions in CA that affect interaction with Nup358 should influence infectivity. Indeed, we found that whilst wild type HIV-1 infectivity is sensitive to both Nup358 as well as TRN-SR2 depletion, certain HIV-1 CA mutants were not, suggesting an inability to utilize these cofactors. We infected HeLa cells stably expressing Nup358 or TRN-SR2 shRNA with GFP-encoding VSV-G pseudotyped HIV-1 vectors bearing wild type or mutant CA. We found that the cyclophilin-binding mutants G89V and P90A were insensitive to Nup358 depletion but remained sensitive to TRN-SR2 depletion (Figure 3A). We hypothesized that an inability to bind Nup358Cyp or CypA might underlie infectivity defects of other HIV-1 CA mutants. The HIV-1 CA mutant N57A is more severely defective in arrested cells than dividing cells (Figure 3A) [10], suggesting that this residue may have a role in nuclear entry. Indeed, ITC demonstrates that N57A is impaired in binding Nup358Cyp (Kd 55 µM) but not CypA (Kd 7 µM) (Figure 3B). As, N57A is less sensitive to both Nup358 and TRN-SR2 depletion (Figure 3A), we hypothesize that its infectivity defect is caused by an inability to engage these proteins. We found that N57A was still restricted by TRIMNup358 (Figure S5), suggesting that increased avidity through Nup358Cyp dimerization in the context of TRIMNup358 may overcome the reduced affinity to monomeric Nup358Cyp. Importantly, N57A's insensitivity to Nup358 depletion suggests that it does not engage Nup358 during nuclear entry.
Finally, we studied the HIV-1 CA mutant N74D, which is reported to be less sensitive to Nup358 or TRN-SR2 depletion (Figure 3A) [17], [31]. Like N57A, N74D bound monomeric Nup358Cyp in ITC experiments with significantly lower affinity than wild type (Kd 95 µM) (Figure 3B) and like N57A, N74D was also restricted by TRIMNup358 (Figure S5A). As for N57A, it seems contradictory that HIV-1 CA mutants that are less sensitive to Nup358 depletion are restricted by TRIMNup358. We assume that binding characteristics of TRIMNup358, and Nup358 itself, to CA are different particularly given that TRIM5 is reported to form cytoplasmic dimers [32] and higher-order multimers [33], Thus a forced dimerization of Nup358Cyp by fusing it to TRIM5α, could increase binding of Nup358Cyp to CA by increasing avidity, thereby allowing restriction. In addition, it is possible that the decreased binding of N57A as well as N74D to TRIMNup358 is disguised by an increased sensitivity to restriction by this TRIM5 chimera. HIV-1 CA N57 is located at the base of helix 3 and N74 in helix 4 (Figure 4C, D), suggesting that amino acid residues outside the Cyp-binding loop can impact on Nup358Cyp binding. We propose that Nup358 and TRN-SR2 define an import pathway used by wild type HIV-1 and that CA amino acid substitutions direct the virus to use Nup358 independent (G89V, or P90A), or Nup358/TRN-SR2 independent (N74D, or N57A) import pathways.
To test whether HIV-1 dependence on Nup358 is increased in non-dividing cells, we arrested HeLa cells with aphidicolin and measured infectivity of the CA mutant viruses. We found that only N57A and MLV infectivities were inhibited by aphidicolin treatment (Figure 3A and Table S1) whereas mutants G89V, P90A and N74D were not affected. This suggests that G89V, P90A and N74D use Nup358/TRN-SR2 independent routes into the nucleus even in the absence of cell division. This hypothesis is supported by the observation that neither the wild type virus nor these mutants become additionally sensitive to Nup358 or TRN-SR2 RNAi in aphidicolin-arrested cells (Figure 3A and Table S1). On the other hand N57A is slightly increased in its sensitivity to aphidicolin particularly after Nup358 depletion. We conclude that these co-factors are required for HIV-1 infection of dividing and non-dividing cells.
HIV integration is favored in chromosomal regions rich in genes and associated features such as CpG islands, DNAaseI hypersensitive sites, and high G/C content. We have shown that Nup358 or TRN-SR2 depletion reduces HIV-1 integration frequency near these features [16]. To test this for the CA mutants studied above, we sequenced 19,546 unique integration sites from HIV-1 and its mutants by 454/Roche pyrosequencing and compared their chromosomal distributions as described [16], [34]–[37].
Although the HIV-1 CA mutants retained the preference for integration within transcription units, their integration site distributions diverged from wild type HIV-1. The patterns clustered into two groups that map to two distinct areas on the CA surface (Figure 4). HIV-1 CA mutants N57A and N74D integrated into regions of chromatin associated with a significantly lower density of transcription units and associated features. For wild type HIV-1 this density was 15 transcription units/MB, whereas for CA mutants N57A or N74D the density was reduced to what is expected for random integration (7–9 transcription units/MB) (Figure 4A and Figure S7). In contrast, the two Cyp-binding mutants, G89V and P90A, exhibited an opposite phenotype, with favored integration into regions of increased density of transcription units (∼20 transcription units/MB). The chimeric HIV-1 containing SIVmac CA, showed a further increased preference for regions dense in transcription units (25 transcription units/MB) (Figure 4A and Figure S7A). These latter three viruses similarly showed increased frequency of integration in areas rich in active genes, CpG islands, DNase sites, and high in GC content, features correlating with high gene density. Hierarchical clustering of the CA mutants based on these data separated the viruses into two groups: N57A and N74D, and the Cyp-binding mutants G89V, P90A and chimeric HIV-1(SIVCA) (Figure 4B). Wild type HIV-1, which has an intermediate targeting phenotype, was an outlier within this second group. Thus amino acid substitutions in CA can alter integration targeting preference, resulting in either of two phenotypes. Because G89V and P90A influence targeting in the same direction, we infer that disruption of normal CypA interactions, and possibly Nup358 interactions, result in increased frequency of integration in regions with high densities of transcription units. The N74D and N57A substitutions are less sensitive to depletion of both Nup358 and TRN-SR2, and N74D gains sensitivity to depletion of other nuclear pore proteins [17]. We thus infer that this pathway leads to favored integration in regions with lower densities of transcription units.
We were surprised that HIV-1 CA mutants P90A and N74D, which are less sensitive to depletion of TRN-SR2 and/or Nup358 (Figure 3A), and have different integration site preferences in unmodified cells (Figure 4), were as infectious as wild type virus in single round assays. This is true when the virus is pseudotyped with the VSV-G envelope (Figure 3A) or the natural HIV-1 gp160 envelope (Figure S9). To test whether these CA substitutions affect HIV-1 replication we compared replication of wild type HIV-1 NL4.3 (Ba-L Env) with CA mutants P90A and N74D in spreading infection in HeLa TZM-bl cells [38]. Interestingly, we found that replication of both HIV-1 CA mutants was impaired in these cells compared to wild type virus, suggesting that cofactors used for nuclear entry and/or integration site selection are important for optimal replication (Figure 4E). We also found that HIV-1 NL4.3 (Ba-L Env) bearing CA alterations N74D or P90A replicated poorly in primary human MDM from four independent donors, whereas wild type virus replicated efficiently (Figure 4F and Figure S7B). These data demonstrate that HIV-1 CA mutants P90A and N74D do not support optimal replication. One possible explanation is that this is due to differences in their integration site targeting as compared to the wild type virus, though other models are possible. Whether the defect in replication is due to a defect in viral gene expression remains unclear. However, it is clear that the mutant viruses that are unable to effectively utilize Nup358 or TRN-SR2 display a replication defect in a cell line and in primary human macrophages.
The observation that HIV-1 CA mutants P90A and G89V, as well as chimeric HIV-1(SIVCA) integrate into genome regions with higher densities of transcription units and associated features raised the possibility that integration targeting might be influenced by CypA binding to CA. Since cyclosporine (Cs) selectively inhibits CypA but not Nup358Cyp binding (Figure 1F, G), we investigated whether Cs could retarget integration by HIV-1. In fact, Cs treatment retargeted viral integration preferences in a way that phenocopied the CA G89V/P90A substitutions shifting integration preferences into regions of higher gene density (Figure 5A, B). Thus preventing CypA-CA interactions with Cs has the same effect on integration targeting as amino acid substitutions in HIV-1 CA that block CypA binding, supporting the idea that integration targeting is truly affected by cyclophilin-CA interactions.
Reduction of Nup358 by RNAi led to integration into low gene density/activity regions [16] but preventing CypA binding by CA amino acid substitutions (G89V/P90A) or Cs treatment shifted virus integration preferences into high gene density/activity regions. This suggested to us that Nup358 and CypA have different, possibly opposing effects on HIV-1. Alternately, Cs treatment may somehow change the availability of Nup358 in the cell. To investigate this further we tested whether CypA inhibition in Nup358 depleted cells influences HIV-1 infectivity. Remarkably, Cs treatment specifically rescued HIV-1 infectivity reduced by Nup358 depletion to the level observed in control cells (Figure 6A and Figure S8). We note that the small inhibitory effect of Cs on HIV-1 infectivity is preserved and infectivity is rescued to the level of infectivity on control cells treated with Cs. Thus Cs inhibits HIV-1 GFP infectivity by 2–3 fold but concomitantly rescues infectivity from the effects of Nup358 depletion. Transient CypA depletion using shRNA expression had a similar effect as CypA inhibition with Cs, also rescuing infectivity reduced by Nup358 depletion (Figure 6B). As expected, the CypA insensitive mutants G89V or P90A did not respond significantly to Cs treatment or CypA depletion by RNAi respectively (Figure 6A, B). The infectivity of the HIV-1 CA mutant N74D was slightly reduced by Cs consistent with its reduced sensitivity to Nup358 depletion and supporting the notion that it is still able to recruit CypA as confirmed by ITC (Figure 3B, 6A and Figure S8). Cs also partially rescued HIV-1 infectivity in cells with strong TRN-SR2 depletion (Figure 6A and Figure S8), suggesting that TRN-SR2 participates in the Nup358 dependent import pathway into which the virus is directed by CypA.
We were also able to show that the distantly related HIV-1 O-group virus MVP5180 was also specifically rescued upon Cs treatment/CypA depletion in Nup358 or TRN-SR2 depleted cells but was unaffected in control cells (Figure 6A, B). This suggests that MVP5180 functionally interacts with CypA in a similar way to NL4.3 and this is concordant with the very similar co-crystal structures of M-group HIV-1 CANTD with CypA and O-group HIV-1 CANTD with CypA (PDB ID: 1M9D) [39]. Together these observations made using both NL4.3 and MVP5180 suggest that CypA acts upstream of Nup358 and that Nup358 is not required for HIV-1 infectivity in the absence of CypA activity. In other words we propose that CypA activity directs the virus to engage Nup358.
If CypA activity directs HIV-1 to interact with cytoplasmic Nup358 to traverse the NPC then reduced HIV-1 infectivity through depletion of nuclear pore proteins that act downstream of Nup358 should also be rescued by CypA inhibition. To test this we analyzed infectivity of HIV-1 NL4.3 and its CA mutants in HeLa cells depleted for Nup153 (Figure 6C). Nup153 is a NPC component in the nuclear basket and has been highlighted in genome wide siRNA screens as co-factor for HIV-1 [14], [40], [41]. We found that HIV-1 NL4.3 infectivity was strongly reduced in Nup153 depleted cells by ∼10-fold, whereas MLV infection was not affected (Figure 6C). However, the Cyp non-binding mutant HIV-1 CA G89V as well as mutants N74D and N57A, which are less dependent on Nup358/TRN-SR2 were only moderately affected (∼3-fold). When cells were treated with Cs during infection, infectivity reduced by Nup153 depletion was specifically rescued for wild type virus, whereas the HIV-1 CA mutants remained unaffected (Figure 6C). The O-group HIV-1 MVP5180 was affected by Nup153 depletion similarly to NL4.3 and CypA inhibition rescued its infectivity similarly to what we observed in Nup358 depleted cells. These observations support the notion that inhibition of CypA recruitment leads HIV-1 to use different cellular cofactors, and perhaps a different pathway, for nuclear entry.
Importantly, Cs treatment prevented spreading infection of wild type HIV-1 in human MDM [42](Figure 6D). Thus, productive infection in a biologically relevant cell type is dependent on the conserved use of cyclophilins. Our results suggest that inhibition of CypA may not only prevent the use of Nup358 but rather may direct HIV-1 into a Nup358/Nup153 independent nuclear entry pathway that may not be available or functional in MDM. In one possible model, for some cell lines such as HeLa cells, the use of alternate pathways and retargeting of integration preferences may not lead to large infectivity defects particularly when measuring infectivity using VSV-G pseudotyped HIV-1 vectors with GFP driven from a heterologous promoter. In replication assays using full length HIV-1 and primary targets of HIV-1 infection such as macrophages, Cs treatment (Figure 6D) or CA residue changes (Figure 4F) have their strongest inhibitory effects. Whilst our observations can be explained by various models, they support the notion that the cofactors that the virus has evolved to use, and has conserved the use of, such as Nup358 and TRN-SR2, may be most important in the primary cells in which the virus naturally replicates. A model is presented in cartoon form (Figure 7).
Here we have presented data suggesting that HIV-1 uses a pathway that includes the cytoplasmic cyclophilin CypA and the nuclear pore associated cyclophilin Nup358 to access the nucleus and target preferred regions of the genome for integration. We find that a determinant for the use of this pathway is CA and we demonstrate the first direct interaction of HIV-1 CA with a component of the nuclear pore complex. Disrupting engagement of CypA/Nup358 by mutating CA or inhibiting CypA with Cs appears to cause HIV-1 to use a Nup358/Nup153 independent pathway. The role of CypA in this process remains obscure but our data suggests that it directs HIV-1 to utilize a nuclear entry pathway involving Nup358 and Nup153. Indeed, roles for CypA in nuclear transport of cellular factors have been proposed before [43]–[45]. Our data illustrate that this HIV-1 nuclear import pathway is directly linked to integration site preference, which provides a candidate explanation for reduced replication in human MDM. Intriguingly, HIV-1 CA substitutions can influence the regions of the genome that the virus targets for integration. We have distinguished between the regions targeted for integration using criteria related to the density of transcription units, including GC content, DNAaseI hypersensitivity and gene expression. Infections with HIV-1 variants containing substitutions in CA that prevent CypA binding (e.g. G89V), and inhibiting CypA binding with Cs, both lead to increased frequency of integration in regions with higher densities of transcription units (Figure 4 and 5), supporting the consistency of our observations. Furthermore, mutations that render HIV-1 less sensitive to both Nup358 and TRN-SR2 depletion (CA N57A and N74D) both shift integration preferences to regions with lower densities of transcription units (Figure 4). Thus mutations that prevent HIV-1 utilizing Nup358 and TRN-SR2 have the same effect as depletion of these proteins, as described in our previous study [16]. These observations suggest that nuclear entry pathways may lead to different areas of chromatin and provide probes to investigate this possibility.
Several reports have suggested a role for HIV-1 CA in nuclear entry [7], [8], [10], [17], [40], [46]. Using ITC experiments we demonstrate here that HIV-1 binds to the nuclear pore through interactions between CA and the C-terminal Cyp domain of Nup358. This is the first direct evidence for an interaction between CA and the NPC and suggests that CA-containing PICs or whole capsid cores dock at the NPC prior to nuclear entry as previously inferred from microscopy studies [47]. Recruitment of cores through Nup358 may assist appropriate uncoating and interaction of PICs with the nuclear transport machinery including TRN-SR2 and Nup153. Remarkably, Nup358Cyp shows evidence for positive selection and a positively selected residue affects restriction by TRIMNup358 suggesting that this residue impacts on HIV-1 CA binding. This is the first case of an HIV-1 co-factor displaying signs of positive selection. We speculate that ancient pathogens, possibly viruses, may have provided the necessary selective pressure for the change of residue 61 from methionine, which has been conserved in the entire cyclophilin family, to valine, isoleucine or leucine in Nup358Cyp. It will be interesting to examine whether other viruses that encounter the nucleus during their life cycle use Nup358 and whether this position influences their recruitment. Indeed, Nup358 has been suggested to be involved in HSV-1 capsid attachment to the nucleus, however the viral determinants for this process remain obscure [48].
We also demonstrate that HIV-1 CA sequence influences the sites in which HIV-1 integrates. Although there are many possible explanations for that, we hypothesize that this occurs through selection of the cofactors for nuclear import or the nuclear import pathway. In the future, it will be interesting to investigate whether TRN-SR2 functions to enhance cytoplasmic availability of HIV-1 co-factors required for nuclear import or integration site selection. Interaction with such co-factors may be disturbed by CA mutations leading to impaired nuclear import or integration.
Surprisingly SIVmac, a primate lentivirus from rhesus macaques that was derived experimentally from SIV from sooty mangabeys [49] does not appear to utilize Nup358 during infection. SIVmac is however sensitive to TRN-SR2 depletion suggesting that it uses a related but somewhat different set of co-factors to enter the nucleus as compared to HIV-1. SIVmac is known to integrate into genes in a similar way to HIV-1 but subtle differences between HIV-1 and SIVmac integration targeting may exist. The significance of these observations remains unclear and characterization of the pathways used by a variety of lentiviruses to enter the nucleus and target favored sites will undoubtedly be informative.
Whilst our data don't rule out partial cytoplasmic uncoating we envisage the HIV-1 CA acting as a protective cage around the reverse transcription complex, shielding the viral macromolecules from pattern recognition by innate immune mediators present in the cytoplasm. Antagonistic Nup358 and CypA activities could be explained by a model in which CypA stabilizes or protects the core [50], whilst Nup358 regulates uncoating at the nuclear pore [47]. In this regard Nup358 binding to the conical viral core could have different effects from monomeric CypA, as eight Nup358 proteins are attached to the NPC [51]. Multiple simultaneous Nup358-CA interactions might destabilize the HIV-1 core and liberated PICs could then interact with TRN-SR2 and the nuclear located Nup153 ensuring transport through the NPC to appropriate sites [52]. This model provides a rationale for conservation of Cyp binding and explains how CA might influence TRN-SR2 or Nup153 usage without direct interaction [22], [40], [46], [52]. This model may also explain how the use of TRN-SR2 does not correlate with the ability of various integrase proteins to bind TRN-SR2 protein in vitro [46]. If lentiviruses regulate uncoating through interactions between CA and other host factors then their integrase proteins may be exposed to different karyopherins during this process. In this way the CA sequence and structure might be a stronger influence on the choice of karyopherins than integrase despite integrase being the ultimate target for karyopherin interaction. The Nup358 cyclophilin domain has been suggested to act as a chaperone by mediating prolyl cis-trans isomerization of cellular proteins [29], [30]. It will be interesting to investigate whether Nup358 is enzymatically active on the HIV-1 capsid core and whether this causes uncoating at the nuclear pore.
Currently available cyclosporins do not antagonize Nup358Cyp binding to HIV-1 CA but the fact that cyclophilins can be pharmacologically inhibited suggests the possibility of specifically inhibiting HIV-1 CA-Nup358Cyp interaction and possibly HIV-1 replication. Overall, our data demonstrate that rather than being lost during cytoplasmic uncoating, HIV-1 CA binds to the nuclear pore component Nup358 and directs the virus into a pathway that regulates its traffic between the cytoplasm and chromatin, playing a key role in the integration site targeting required for optimal continuation of the viral replication cycle.
VSV-G pseudotyped vectors derived from HIV-1, SIVmac and MLV-B have been described as has their preparation by 293T transfection [53]. The HIV-1/SIVmac chimeric vectors have been described [20], [21] as has the HIV-1 vector encoding MVP5180 Gag [19]. HIV-1 NL4.3GFP-IRES has been described [54]. Viral doses were measured by reverse transcriptase (RT) enzyme linked immunosorbant assay (Roche). Viral vector infection assays using VSV-G pseudotyped viruses encoding GFP were analyzed by enumerating the number of green cells 48 hours post infection by flow cytometry. Viral vector infectivity experiments were performed in a 24-well plate format as described [53]. To measure late RT products, 2-LTR circles and integrated provirus, control or shRNA expressing cells were infected with VSV-G pseudotyped HIV-1 GFP encoding vector and then grown for indicated times. Total DNA was purified from 2 samples at each time point (QiaAmp, Qiagen) and 600 ng were subjected to Taqman quantitative PCR using late RT [55], 2-LTR circle [56] or GFP [53] primers and probe to detect provirus as described. Infectivity was measured in parallel samples by flow cytometry 48 hours post infection.
MDM were prepared from fresh blood from healthy volunteers as described [57]. Cells were infected with 400 pg RT/well in 24-well plates and subsequently fixed and stained using a CA specific antibody (CA183) and a secondary antibody linked to beta galactosidase as described [57]. For measuring the effect of CypA inhibition on HIV-1 replication the assay was performed in the presence of 5 µM DMSO or Cs throughout the whole time course. TZM-bl infection assay was performed with 50 pg RT/20000 cells in 24-well plate dishes and RLU were measured at indicated time points.
Methods for integration site sequencing and heat map and dendogram analysis have been described [16].
All RNA interference experiments were performed by expressing short hairpin RNA from either MLV vector pSIREN RetroQ (Clontech) (for Nup358, TRN-SR2 and Nup153) or pSUPER (Oligoengine) (for CypA) or if indicated from the HIV-1 vector pCSRQ, which was derived by subcloning the shRNA expression cassette from pSIREN RetroQ into pCSGW. The CypA shRNA target sequence has been described [58]. The Nup358 shRNA target sequence that was used throughout the study was 5-GCGAAGTGATGATATGTTT-3. Nup153 shRNA target sequence was 5-CAATTCGTCTCAAGCATTA-3. Both sequences were selected as 1 of the 4 target sequences from the Dharmacon siRNA smartpool for Nup358 or Nup153, respectively. Both shRNAs had only minor toxic effects on the cells, unlike shRNAs derived from the other three target sequences of each smart pool (Figure S1A, and data not shown). Additional Nup358 shRNA target sequences used in the experiment shown in Figure S1A were shRNA2 5-CAAACCACGTTATTACTAA-3, shRNA3 5-CAGAACAACTTGCTATTAG-3 and shRNA4 5-GAAGGAATGTTCATCAGGA-3. Specificity for each target sequence was confirmed by BLAT (UCSC genome browser). For Nup358, we confirmed effective targeting by co-transfecting the shRNA expression vector with a plasmid encoding GFP-tagged Nup358 (Figure S2), as well as by western blotting using a Nup358 specific antibody (Figure 1B). TRN-SR2 target sequence and control have been described and were also validated by co-transfecting a plasmid encoding for TRN-SR2-IRES-eGFP with the expression vectors encoding shRNA or control (SC) (Figure S2) [22]. The observations made for shRNA expressing HeLa cells were similar between populations of puromycin selected cells and clonal cells but the phenotype of cell clones was more stable, thus we used single cell clones for all experiments (Figure S1A, B and data not shown). Nup358, TRN-SR2, Nup153, CypA and beta-Actin were detected by western blot using a Nup358 antibody kindly given by Frauke Melchior, mouse TRN-SR2 antibody ab54353 (Abcam), mouse Nup153 antibody ab24700 (Abcam), rabbit CypA antibody SA296 (Biomol) and mouse beta-Actin antibody ab6276 (Abcam) and appropriate horseradish peroxidase linked secondary antibodies. TRIMCypA and TRIMNup358 were detected using anti-HA antibody 3F10 (Roche). Cyclosporine (Sandoz) and aphidicolin (Sigma) were diluted in DMSO and used at 5–8 µM and 2 µg/ml, respectively.
Isothermal titration calorimetry was performed as described [25].
Codon-specific selection analysis was performed using the Random Effect Likelihood (REL) algorithm as described [27] using the alignment in Figure S6.
|
10.1371/journal.pcbi.1000391 | How to Get the Most out of Your Curation Effort | Large-scale annotation efforts typically involve several experts who may disagree with each other. We propose an approach for modeling disagreements among experts that allows providing each annotation with a confidence value (i.e., the posterior probability that it is correct). Our approach allows computing certainty-level for individual annotations, given annotator-specific parameters estimated from data. We developed two probabilistic models for performing this analysis, compared these models using computer simulation, and tested each model's actual performance, based on a large data set generated by human annotators specifically for this study. We show that even in the worst-case scenario, when all annotators disagree, our approach allows us to significantly increase the probability of choosing the correct annotation. Along with this publication we make publicly available a corpus of 10,000 sentences annotated according to several cardinal dimensions that we have introduced in earlier work. The 10,000 sentences were all 3-fold annotated by a group of eight experts, while a 1,000-sentence subset was further 5-fold annotated by five new experts. While the presented data represent a specialized curation task, our modeling approach is general; most data annotation studies could benefit from our methodology.
| Data annotation (manual data curation) tasks are at the very heart of modern biology. Experts performing curation obviously differ in their efficiency, attitude, and precision, but directly measuring their performance is not easy. We propose an experimental design schema and associated mathematical models with which to estimate annotator-specific correctness in large multi-annotator efforts. With these, we can compute confidence in every annotation, facilitating the effective use of all annotated data, even when annotations are conflicting. Our approach retains all annotations with computed confidence values, and provides more comprehensive training data for machine learning algorithms than approaches where only perfect-agreement annotations are used. We provide results of independent testing that demonstrate that our methodology works. We believe these models can be applied to and improve upon a wide variety of annotation tasks that involve multiple annotators.
| Virtually every large-scale biological project today, ranging from creation of sequence repositories, collections of three-dimensional structures, annotated experiments, controlled vocabularies and ontologies, or providing evidence from the literature in organism-specific genome databases, utilizes manual curation.
A typical curation task in biology and medicine involves a group of experts assigning discrete codes to a datum, an experimental observation, or a text fragment. For example, curators of the PubMed database assign topics to each article that is registered in the database. These topics are encoded in a hierarchical MESH terminology [1] to ensure that curators have a consistent way to define an article's content. Other curation examples include annotation of function of genes and proteins, description of genetic variation in genomes, and cataloguing human phenotypes. A standard approach to assessing quality of curation involves computation of inter-annotator agreement [2], such as a kappa-measure [3].
Manual curation is tedious, difficult, and expensive. It typically requires annotation by multiple people with variable attitudes, productivity, stamina, experience, tendency to err, and personal bias. Despite its difficulties and the imprecision in outcome, curation is critical. Existing curation approaches can be improved and enhanced with careful experimental design and appropriate modeling. This study aims to address the following questions:
In this study we propose a holistic approach to quantify our certainty in individual annotations for a group of several annotators, which allows to retain the complete dataset as a basis for training and testing machine learning methods.
Specifically, we suggest an internally consistent way to design annotation experiments and analyze curation data. We created two alternative probabilistic models for such analysis, tested these models with computer simulations, and then applied them to the analysis of a newly annotated corpus of roughly 10,000 sentences. Each sentence in this corpus was annotated by three experts. To test the utility of our computational predictions, we randomly sampled a subset of 1,000 sentences (out of the original 10,000) to reannotate by five new experts. Using these two rounds of annotation, we evaluated the models' predictions by comparing the three-experts-per-sentence results against the “gold standard” eight-experts-per-sentence analysis.
First, to generate the corpus, our homemade scripts extracted 10,000 full sentences randomly from diverse scientific texts, making sure that all sentences are distinct and that section-specific and topic-specific constraints are met. Specifically, we randomly selected 1,000 sentences from the PubMed database, which at the time of our analysis stored 8,039,972 article abstracts (note that not every PubMed entry comes with an abstract). We also sampled 9,000 sentences from the GeneWays corpus (368,331 full-text research articles from 100 high-impact biomedical journals). We put the following constraints on these 9,000 sentences: 2,100 sentences were sampled from articles related to WNT pathways, apoptosis, or schizophrenia research (700 sentences per topic, with random sampling within each pool of topic-specific articles). The remaining 6,900 sentences were sampled with restriction on article section: 20% of the sentences came from abstracts, 10% from introductions, 20% from methods, 25% from results, and 25% from article discussion sections. We did not process sentences in any way before the annotation. Because the current study is not concerned with automatic annotation of sentence fragments per se, we do not elaborate on machine-learning features that we described in our earlier study [4].
Second, we randomly reordered the 10,000 sentences and partitioned them into eight equal-size sets. We arranged eight annotators recruited for the first cycle of analysis into eight 3-annotator groups, assigning to each group a unique sentence set. This way each annotator analyzed three sets of sentences, and utilizing the loop-design of the analysis (see Figure 1A) we were able to computationally compare annotators' performances with each other. This concluded the first cycle of annotation. The part of this corpus on which all three annotators perfectly agreed, as well as the part on which at least two out of the three agreed, were used for training, testing and analyzing supervised machine learning methods for automatic annotation assignment, in a recent study, reported elsewhere [4].
As the models for annotation reliability introduced here are based on the above corpus, to reliably validate the models, we performed a second cycle of annotation. To do this, we recruited five additional annotators, sampled a subset of 1,000 random sentences out of the original 10,000, and asked the new annotators to annotate the 1,000-sentence subset. The result of the second cycle of annotation was a 1,000-sentence set that was annotated by five annotators per sentence in the second cycle and by three annotators per sentence in the first cycle.
The whole annotated corpus is publicly available along with this manuscript (see Dataset S1).
When defining guidelines for our present annotation effort [5] we aimed at distinguishing among several types of scientific statements, varying across multiple dimensions. Specifically, we tried to distinguish commonplace knowledge from original conclusions, high certainty statements from uncertain ones, experimentally supported evidence from speculations, and scientific statements from methodological or meta-statements. The goal of this effort was to generate a manually annotated corpus that can be further used to train computers to automatically perform well-defined annotation tasks at a large scale.
In the long run, we hoped to learn to automatically highlight portions of research articles that fit a particular search goal. Such a goal can be, for example, to identify all original conclusions supported by experiments. Another plausible goal (out of many imaginable) is to find the scientific statements made with high certainty, with or without experimental support. A tool of this kind would be a useful addition to the armamentarium of a biomedical text-miner.
We asked experts to annotate sentences along the following six dimensions (with two of them, polarity and certainty, combined), described in great detail in an earlier article [5] :
In addition, each annotation of a dimension is allowed to have code Error, indicating erroneously extracted or jumbled sentence.
As the focus of this work is the construction of models for annotation correctness, we next describe these models.
Despite the apparent complexity of the generative process under Model A, in its simplest form the model requires only one parameter per annotator for any number of allowed annotation values. In contrast, for Model B, given n permissible annotation values, there are n−1 independent values of γ's and one independent value of λ(i)x|x for each annotator. As a result, for the number of fragments in a sentence that allows 9 values, Model B requires optimization of a likelihood function depending on 16 free parameters (584 for the full model), whereas the likelihood for Model A depends only on 8 (11 for the full model).
It is well known that if we estimate parameters using numerical function optimization over a fixed-sized dataset, it is much easier and quicker to obtain the maximum-likelihood estimates when the number of model parameters is small. As the number of parameters increases, the data is typically insufficient to uniquely determine the parameter values, and parameter estimates may widely vary. As our experiments with simulated data illustrate (see below), the number of local optima grew exponentially with the number of permitted annotation values for Model B. While Model A also had multiple optima, their number was smaller, and only one optimum occurred within the parameter area where all annotators performed with correctness >0.5.
The multimodal shape of our likelihood functions is a direct consequence of the inability to directly observe or determine the correctness of annotation values. Multimodal likelihood surfaces are a common feature of models involving latent variables (e.g., see [9] ), suggesting that multiple explanations are possible for the same data and each corresponds to a mode on the likelihood surface. Moreover, the larger the number of parameters, the larger the number of possible configurations explaining the same dataset.
One additional advantage of Model A is, when we annotate the same fragment of text along multiple dimensions, Model A can easily be altered to allow for non-independence among distinct types of annotations. (See Text S1 for details.)
To test our methodology, before applying it to real annotator data, we conducted a study in which data were simulated from one of the models and then parameters were estimated under both (see Figure 3). When we obtained simulated data from Model A, the parameters estimated for Model A clustered nicely along a perfect diagonal (given that both true values and the initial optimization values of the correctness parameters were >0.5) (yellow circles in Figure 3A). The parameters for Model B produced a much greater scatter of likelihood values, with better likelihood estimates closer to the expected values (blue circles in Figure 3A). This result (along with additional repeated-estimation analysis) indicates that poorer likelihood values for Model B correspond to convergence to the numerous local optima on the B-model likelihood surface. In one example (detailed in the Text S1), we made 300 estimates under Model B for the same simulated dataset (3 allowed annotation values). The estimation search ended in the same local optimum only 2 times out of 300; 298 sets of estimates were all distinct from each other.
When simulation was performed by generating data under Model B, the parameters estimated for Model A tended to be more widely scattered than when the data was simulated under Model A (yellow circles in Figure 3B). Nevertheless, Model A estimates still tend to follow the diagonal of the plot. As expected, when estimation for Model B was initialized at the true parameter values, the resulting estimates grouped tightly around the perfect-estimate diagonal (black circles in Figure 3B). However, when estimation under Model B was initialized at a random point in the parameter space, (blue circles in Figure 3B), estimation scatter became extreme due to convergence on local optima.
The simulations indicate that we can indeed obtain reasonable estimates of the annotator correctness parameters, and this was practically easier with Model A. Estimating parameters for Model B was computationally more expensive, requiring the estimation of many more parameters, while frequently settling into local optima. As such, we use Model A and the simplified B-with-thetas to analyze our real annotator data.
Figure 4 shows estimates of key parameters for Model A using approximately 10,000 sentences, each annotated by three annotators. Figure 4A shows maximum likelihood estimates of the correctness parameters for the eight annotators (the first round of evaluation) and four dimensions of annotation.
Surprisingly, not only did the value of correctness vary significantly among annotators, but the same annotator's correctness fluctuated widely across the annotation tasks. The same annotator could perform extremely well at one annotation task and terribly at another (see Figures 4A and 5A for results of analysis under models A and B-with-thetas, respectively). We observed very similar absolute values of correctness and nearly identical patterns of annotator-specific correctness across dimensions under the two models. Thus, it is more likely that the features of our annotator correctness estimates reflect properties of annotator performance rather than being artifacts of model design.
Estimates of conditional probabilities of agreement patterns given correctness status (denoted by α's, see Figure 4B–E) and estimates of code frequencies (denoted by ω's, see Figure 4F–I) also tell an interesting story. As we have noted previously [4], frequencies of annotation values for each dimension were far from uniform: The probability was almost 0.75 that a sentence would be annotated as having a single fragment (Figure 4F). Similarly, there was a greater than 0.60 chance that a fragment would contain either no reference to experimental evidence at all (E0) or direct evidence (E3), but not a value in between (E1 and E2, see Figure 4G); a 0.55 chance that the sentence would be annotated as having scientific focus (Figure 4H); and a greater than 0.75 chance that the fragment would contain the most certain positive statement (Figure 4I). Distributions of the code-frequency values, ω's, were mirrored fairly closely by the annotation correctness distributions (γ-distributions), estimated for Model B-with-thetas, Figure 5 (B–E).
The direct consequence of the skewed distribution of annotation codes is that under Model A the probability of random convergence to incorrect annotation values was high. Consider the conditional probabilities of agreement patterns given correctness states for the number of fragments in the sentence (Figure 4B). When all three annotators had incorrect annotations (III), the most likely observed agreement pattern was a perfect consensus (aaa, Figure 4B). Other dimensions of annotation showed a similar trend (Figure 4C–E). Why are these observations important? Because, depending on the annotation task, relying on annotator consensus annotations can lead to accepting erroneous annotations, while a proper stochastic modeling can rectify the problem.
The online Text S1 provides all equations required to identify the annotation with the highest posterior probability for each annotated fragment of text.
While there are numerous approaches for comparison of models in terms of their goodness-of-fit to data (e.g. [10] ), we do not apply them in our comparison of models A and B, because comparison of the raw log-likelihood values makes application of more sophisticated approaches unnecessary. Indeed, when we apply both models to our real annotator data, the most complicated version of Model A (namely, A-with-alphas) has 11 parameters to resolve the number of sentence fragments while the simplest version of Model B has 16 parameters. The best log-likelihood values we achieved after performing hundreds of independent runs of our random-start likelihood-maximization processes with A-with-alphas and B-with-thetas were −19,215.544 and −22,897.744, respectively. It is clear, even without any more sophisticated model-selection approaches, that the simple Model A fits the data orders of magnitude better than the more complicated Model B. The simplified Model A (−19,289.269), as expected, does not fit the data as well as its parameter-enriched version, but still significantly better than Model B. Curiously, estimates of annotator-specific accuracies (θ-values) are virtually identical under both versions of Model A (data not shown).
That said, it is important to note that comparison of models is never absolute but is always relative to the data on which the models are being compared. In other words, despite our observations on the particular dataset, it is likely that there are datasets on which the performance of the models is reversed.
The critical question regarding a study like this is whether the suggested approach is actually useful. To compare model-based predictions with external evaluations, we selected a random subset of 1,000 sentences (out of the original 10,000) and recruited five additional independent annotators to provide 5-fold re-annotation of these 1,000 sentences.
The most obvious way to demonstrate the utility of our models would be to re-evaluate predictions for cases where the three original annotators provided three different annotations for the same fragment of text, and compare these annotations to those produced by the additional five independent annotators.
To understand details of the underlying computation, consider the specific task of annotating the number of fragments in a sentence. For example, the original three annotators had estimated accuracies under the simplified Model A (θ-values) of 0.91776, 0.91335, and 0.82234, and detected 2, 1, and 4 fragments in the sentence, respectively (3-way disagreement). The correctness values alone suggest that we should trust annotator 1 most and annotator 3 least. We can further quantify our trust by computing the posterior probabilities that each annotator is correct given this particular triplet of annotators and annotation values.
The posterior probabilities (again, under simplified Model A) that the correct number of sentence fragments is 2, 1, and 4 are 0.4223, 0.3989, and 0.1752, respectively. That is, we have more than twice as much confidence that annotator 1 is right than that annotator 3 is right. Furthermore, we have more than four times the confidence that the correct number of fragments is either 1 or 2 as opposed to 4. To check the validity of our prediction, we looked at five additional (independent) annotations for the same sentence: 2, 2, 2, 2, and 1. Combining the original three annotations with the five new ones, we obtained an 8-way annotator vote value: 2. In this case, clearly, the 8-way vote coincided with our maximum a posteriori probability (MAP) prediction.
Due to the nature of our annotation protocol, where annotations are assigned to fragments rather than to complete sentences, the validity of polarity, focus, and evidence annotations was confounded by the validity of sentence segmentation (see Table 1 for an example). When annotators disagreed on the number of fragments and, especially on fragment boundaries, our analysis had to deal with small spurious fragments. To clarify, consider the case in which three annotators annotate a hypothetical sentence consisting of just three one-letter words: “A B C.” The annotators are allowed to break the sentence into fragments and annotate fragments with one of two codes: 1 or 2. Suppose that annotator 1 broke the sentence into fragments “A” and “BC”, annotating the first fragment with code 1 and the second with code 2 – for brevity we write this as A1|BC2. Similarly, evaluator 2 produced annotation AB1|C2 – breaking the sentence also into a pair of fragments, but, unlike annotator 1, grouping A and B. The third evaluator did not break the sentence at all, assigning annotation 2 to the whole sentence: ABC2. In combining these annotations, in order to enable analysis of the results, we first find the minimal fragmentation that incorporates all breakpoints –in our case, A, B, and C. Then we re-write the original annotations by transferring codes from larger fragments to smaller ones: A1B1C2, A1B2C2, and A2B2C2, for annotators 1, 2, and 3, respectively. As such, we pooled all breakpoints from the annotators to determine the final fragmentation and each of the original annotated fragments propagates its annotation down to all the final fragments composing it. We recognize that in future studies the segmentation and annotation should be performed in two stages. The first stage should focus on annotating boundaries of the fragments and finding the maximum a posteriori boundary. The second stage should involve annotation of fragment codes given the MAP sentence fragmentation. Despite this additional noise, our analysis below demonstrates that MAP predictions were significantly enriched with correct answers.
The main difficulty with the practical application of Model B is that even after a large number of numerical optimization runs, starting with different initial values, we had no confidence that we had identified the global optimum of the posterior probability. Nevertheless, we used the set of parameter estimates marked with the best likelihood value and the highest prior probability observed in a set of about 100 independent runs. Our results show that even these imperfect parameter estimates provide surprisingly robust prediction results (see Table 2).
For evaluating the quality of our model-specific predictions we need to establish a baseline corresponding to a naïve random-predictor method. If we consider only three-way annotation disagreements, a naïve random-predictor method would work by sampling an annotation out of three choices with a uniform probability (1/3). Similarly, the probability that two annotations out of three include the correct answer (given that one of the three answers is correct) is 2/3. Examining Table 1, we can immediately see that the number of correct MAP predictions under both models was almost invariably greater than the randomly expected number (with the one exception of the two-best-predictions analysis of Polarity—Certainty annotations).
Both models appear to do their prediction job extremely well, with Model B-with-thetas performing marginally better. Despite the relatively small numbers of test cases for each type of annotation (31, 157, 108, and 87 three-way disagreements for distinct annotation types, see Table 2), we observe highly significant deviation from random prediction for each annotation type. The majority of our model-specific p-values, computed with Pearson's chi-squared test, are smaller than 10−3 and a few are smaller than 10−7 and 10−10 (see Table 2), indicating the extreme improbability that our prediction success is accidental.
While Model A fits the data better, Model A assumes that annotations are dependent only on the agreement pattern of judges and, given agreement pattern, are conditionally independent of their correctness. We suspect this independence assumption is violated to some extent and this explains Model B's slight advantage in predicting the eight judge results based on the three judge data.
In summary, our method picks the correct prediction (as determined by a larger panel of new additional independent experts) much more frequently than random, proving that our approach offers a practical aid to annotation tasks.
We performed our probabilistic analysis using programs written in MatLab (MathWorks); all corresponding scripts are available to anyone interested.
For our numerical analysis of posterior probability distributions, we used our own implementation of a simulated annealing algorithm [11], the MatLab implementation of the multidimensional simplex method, and common sense, see Dataset S2.
Our analysis above, demonstrates the advantages of careful experimental design, hopefully sufficiently so to convince the biological data curation community regarding the value of an experimental methodology in implementing and analyzing data curation results. It appears that a comparison of curator performance already justifies the effort, but the benefits go well beyond quality control. Our analysis offers the possibility of probabilistic data annotation, where alternative annotations are presented with appropriate degrees of certainty. This represents the plurality of opinions and disagreements among human experts in a much more organic way than does exclusive, deterministic (“crisp”) annotation. Our probabilistic, Bayesian approach to data annotation allows preservation and use of all annotation data, rather than the discarding conflicting parts. Furthermore, probabilistic machine learning methods, such as the maximum entropy and conditional random fields approaches, are well suited for imitating human curators and learning from such annotations.
As is further exemplified in the following section, the methodology described in this paper is directly applicable to a wide spectrum of annotation tasks, such as annotation of large fragments of text (articles, paragraphs, books), nucleotide sequences, phenotypes, three-dimensional models, and raw experiments. One could even use it to compare computational methods, for example, in the computational annotation of genomic regions, or in the detection of copy number variation using expression array data. In these applications, computation-generated predictions take the role of annotators (with unknown accuracies) annotating the same piece of data.
In the spirit of exploring mathematical symmetries [12], we notice that extrema in likelihood optimization under Model B form a permutation group that has n! group members for annotation with n admissible values. We can show (see Text S1) that every mode (solution) that belongs to the same permutation group has exactly the same height (the maximum likelihood value). We exploited this property in our implementation of the Expectation-Maximization algorithm, as explained in the Text S1. While each optimum has a corresponding permutation group of equivalent solutions yielding the same probability for the data, the likelihood surface is replete with local optima which are not equivalent and which we cannot currently count or characterize.
Both of the proposed models give rise to multiple solutions for the same data, although Model B is especially rich in alternative modes at the likelihood surface. At first we viewed this property disparagingly. Later, however, we realized a positive aspect of this multiplicity. It is true that a practical minded researcher looks for a unique solution to a mathematical problem. However, reality can often be explained in multiple ways. We can think of the multiple solutions to a set of equations as merely an invitation to consider alternative logically consistent ways to interpret data. This is not an unprecedented situation: the famous field equations formulated by Albert Einstein [13] allow for numerous solutions; each consistent solution, discovered by different thinkers during the last century, suggested a unique view of the physical world with profound and distinct philosophical implications.
How can the real-world data curation efforts, such as Arabidopsis thaliana annotation [14], Mouse Genome Database [15], UniProt and Swiss-Prot [16], GenBank [17] and numerous other repositories heavily used by bench biologists, benefit from our methodology?
It would be naïve on our part to expect that every curation team in the world will immediately switch to annotating each piece of data three times, using a loop design for multiple annotators (it would be nice, though). However, it is likely to benefit the curation teams to conduct small-scale annotation experiments, estimating error rates specific to the task at hand and to the group of annotator experts. Such estimates can be immediately used to assign confidence to data annotated by a single expert with a known correctness rate. Furthermore, estimates of annotator correctness are useful in conducting randomized quality control checks, where a randomly chosen datum is re-annotated by a group of three annotators with known performance metrics. We further illustrate the applicability of the method in the following example.
Consider a team of curators at the Jackson Laboratory in Bar Harbor, Maine, working on curating mouse phenotypes for mouse strains with genetic differences within corresponding genomes.
A genome of a given mouse strain can harbor a spectrum of variations relative to the genome of another mouse strain. Mouse phenotypes are arranged into a hierarchical terminology [15],[18] where each term is assigned to a unique code. While in some cases assignment of genetic variation to a phenotype is clear and unambiguous, in others the curators have to resolve some degree of ambiguity of assignment of rearrangement to a specific gene or genes (e.g., when multiple genes are affected) or of genetic variation to a phenotype (e.g., when pleiotropic variation is considered).
We can directly relate such an annotation task to our modeling framework. Suppose that eight curators (1, 2, …, 8) are arranged into eight groups of three experts each: (1, 2, 3), (2, 3, 4), (3, 4, 5), …(7, 8, 1). We ask curators within the same group to assign discrete phenotypic codes to the same subset of genetic variations. From the annotated data we can estimate model parameters for Models A and B as described in the paper, and estimate curator-specific error rates.
Such error-rates are immediately useful in order to:
The above example illustrates the applicability and potential utility of the models within the setting of a current and ongoing curation effort.
|
10.1371/journal.pcbi.1002201 | Changes in Dynamics upon Oligomerization Regulate Substrate Binding and Allostery in Amino Acid Kinase Family Members | Oligomerization is a functional requirement for many proteins. The interfacial interactions and the overall packing geometry of the individual monomers are viewed as important determinants of the thermodynamic stability and allosteric regulation of oligomers. The present study focuses on the role of the interfacial interactions and overall contact topology in the dynamic features acquired in the oligomeric state. To this aim, the collective dynamics of enzymes belonging to the amino acid kinase family both in dimeric and hexameric forms are examined by means of an elastic network model, and the softest collective motions (i.e., lowest frequency or global modes of motions) favored by the overall architecture are analyzed. Notably, the lowest-frequency modes accessible to the individual subunits in the absence of multimerization are conserved to a large extent in the oligomer, suggesting that the oligomer takes advantage of the intrinsic dynamics of the individual monomers. At the same time, oligomerization stiffens the interfacial regions of the monomers and confers new cooperative modes that exploit the rigid-body translational and rotational degrees of freedom of the intact monomers. The present study sheds light on the mechanism of cooperative inhibition of hexameric N-acetyl-L-glutamate kinase by arginine and on the allosteric regulation of UMP kinases. It also highlights the significance of the particular quaternary design in selectively determining the oligomer dynamics congruent with required ligand-binding and allosteric activities.
| Protein function requires a three-dimensional structure with specific dynamic features for catalytic and binding events, and, in many cases, the structure results from the assembly of more than one polypeptide chain (also called monomer or subunit) to form an oligomer or multimer. Proteins such as hemoglobin or chaperonin GroEL are oligomers formed by 2 and 14 subunits, respectively, whereas virus capsids are multimers composed of hundreds of monomers. In these cases, the architecture of the interface between the subunits and the overall assembly geometry are essential in determining the functional motions that these sophisticated structures are able to perform under physiological conditions. Here we present results from our computational study of the large-amplitude motions of dimeric and hexameric proteins that belong to the Amino Acid Kinase family. Our study reveals that the monomers in these oligomeric proteins are arranged in such a way that the oligomer inherits the intrinsic dynamic features of its components. The packing geometry additionally confers the ability to perform highly cooperative conformational changes that involve all monomers and enable the biological activity of the multimer. The study highlights the significance of the quaternary design in favoring the oligomer dynamics that enables ligand-binding and allosteric regulation functions.
| The biological function of proteins is usually enabled by their dynamics under native state conditions, which, in turn, is encoded by their 3-dimensional (3D) structure. Unraveling this functional code has been the aim of many experimental and theoretical studies [1]–[9]. In particular the slow conformational dynamics of proteins in the micro-to-milliseconds time scale has been pointed out to be consistent with the changes in structure or domain/subunit movements observed between the substrate-bound and -unbound forms of enzymes [4]–[7],[10], and potentially limit the catalytic turnover rates of enzymes [11]–[14]. The quaternary structure of oligomeric proteins adds another layer of complexity to this code as the assembly of the subunits entails additional constraints while possibly inducing new types of collective motions. The structural hierarchy in oligomers indeed gives rise to a wide diversity of dynamical events [15]. For instance, in allosteric proteins, such as the paradigmatic hemoglobin [16], [17], the coupling between the internal dynamics of the subunits and the intrinsic ability of pairs of dimers to undergo concerted reorientations with respect to each other underlies the cooperative response to ligand binding [18]–[20]. Analysing the slow conformational dynamics thus emerges as a crucial step towards understanding the structure-function code in oligomeric proteins.
Two classical models have been broadly used in the literature to interpret the conformational changes observed upon ligand binding: the Koshland-Némethy-Filmer (KNF) model [21] where the ligand ‘induces’ a conformational change in the allosteric protein, in line with the classical induced fit model, and the Monod-Wyman-Changeux (MWC) model [22] where the ligand selects from amongst those pre-existing conformers accessible by the intrinsic dynamics of the 3D structure. The former is usually a stepwise process, while the latter is all-or-none. The experimentally observed structural changes appear to result from a combination of intrinsic and induced effects: the intrinsic dynamics of the protein prior to substrate binding is essential to enabling cooperative changes in structure, while induced motions, usually more localized, help optimize and stabilize the bound conformers [4], [23].
Protein-protein interfaces are usually characterized by their size, shape complementarity and hydrophobicity [24], [25]. The dynamics at the interfacial residues are usually given little attention, although the functional significance of the structural changes triggered by complex formation or oligomerization is widely recognized. The interface between subunits often plays a key role in mediating the activity of each monomeric subunit [25]. Protein-protein interactions provide, not only thermodynamic stability to the folded state of the subunit in the complex (or assembly), but also a new spectrum of collective motions. Furthermore, the oligomeric arrangement provides an efficient means of communication that may modulate allosteric regulation [19]. The present study focuses on the following questions: (1) Is the intrinsic dynamics of the component subunit modified by the oligomerization process, and if so, in which ways? (2) What is the role of interfacial interactions and overall contact topology in the functional dynamics of the oligomer and, in particular, in signal transduction or allosteric communication?
The effect of multimerization on protein dynamics is investigated here in the context of the Amino Acid Kinase (AAK) family of enzymes. Members of this family have different degrees of oligomerization (Figure 1). Rubio and co-workers have significantly contributed to our current knowledge of this family of enzymes: they have resolved the X-ray structures of most family members [26]-[33] and suggested a shared mechanism of action on the basis of their sequence and folding similarities [28]. This mechanism was elucidated by our recent computational study of the softest modes of motion intrinsically accessible to different members of the AAK family of proteins [34].
The most exhaustively studied member of the AAK family is N-acetyl-L-glutamate kinase (NAGK) (Figure 1A). NAGK phosphorylates the amino acid N-acetyl-L-glutamate (NAG) in the bacterial route of arginine biosynthesis. In many organisms, NAG phosphorylation is the controlling step of the route, as NAGK is feedback inhibited by the end product arginine. Rubio and co-workers [30] characterized the structures of two hexameric NAGKs (from Thermotoga maritima (Figure 1B) and Pseudomonas aeruginosa) that are cooperatively inhibited by arginine [35]. In Escherichia coli, NAGK (EcNAGK) is homodimeric and arginine-insensitive (Figure 1A). Indeed, several studies have proven that the hexameric arrangement is a requirement for the cooperative inhibition by arginine [30], [36]. The distinctive feature of this biosynthetic route in bacteria is that it produces N-acetylated intermediates, in contrast to mammals that yield non-acetylated intermediates. This turns NAGK into a potential target for antibacterial drugs by selective inhibition. Another member of the AAK family is carbamate kinase (CK; Figure 1C). CK catalyses the formation of ATP from ADP and carbamoyl phosphate (CP; a precursor of arginine and pyrimidine bases), and undergoes a substantial change in its structure upon substrate binding [37]. A third member is the hexameric UMP kinase (UMPK) (Figure 1D). UMPK catalyzes the reaction ATP + UMP ADP + UDP to yield uridine diphosphate (UDP). It is involved in the multistep synthesis of UTP, being regulated by the allosteric activator GTP and inhibited by UTP itself. Its monomer fold is very similar to the rest of family members, but presents a strikingly different assembly of the subunits that has not been explained so far.
Notably, while the AAK family members do not exist in monomeric form, they share the same monomeric fold. This commonly shared monomeric fold is stabilized by oligomerization. The selection of a common monomeric fold in different oligomers suggests that that particular architecture possesses structure-encoded dynamic features that are exploited for enzymatic activity in oligomeric state. It is essential to analyze what the intrinsic dynamics of the monomeric units are, and to what extent, if any, they are maintained in the oligomeric state, or how they are coupled to, or complement, the dynamics of the biologically active (oligomeric) state. Calculations are thus performed for the monomeric fold alone as well as the monomer in the context of different oligomeric states, and the intact oligomers. As will be shown below, the oligomers do maintain some intrinsic dynamic features of the monomeric units, while the different assembly geometries of the monomers give rise to global motions uniquely defined for the particular oligomerization states. The method of analysis presented here is applicable to any protein that functions in different multimeric states. The effect of oligomerization on the dynamics of the component subunits can be experimentally examined provided that the protein exists in monomeric and different oligomeric states, which, in turn, may be controlled by environmental conditions [38] and few mutations at the protein surface [39]. However, such studies may be challenging in practice, and a computational examination emerges as an alternative promising tool.
The most collective movements of biomolecular systems, also called the global modes of motions, can be determined using Elastic Network Models (ENMs) in conjunction with Normal Mode Analysis (NMA) at very low computational cost. A wealth of studies have shown the robustness of the global modes predicted by the ENMs (e.g., by the anisotropic network model, ANM [40], [41]) and their close relevance to experimentally observed structural transitions related to ligand binding [4]-[6], [10], [18], [41]–[46], or to the essential modes extracted from converged molecular dynamics (MD) simulations [47]–[49]. The global modes are the low-frequency modes extracted from NMA, also referred to as slow modes. They correspond to large-amplitude motions taking place at long timescales (e.g. microseconds to milliseconds); and they are also called soft modes due to their lower energy cost associated with a given level of fluctuation away from the equilibrium state, compared to other modes. Given their robustness and efficiency, ENMs are uniquely suited for exploring the collective motions and allostery in oligomers. Previous such studies have highlighted the significance of multimeric arrangement in defining the collective dynamics [50]–[54].
The present study adds new evidences to the role played by multimerization in defining functional dynamics. First, we contrast the low-frequency modes favoured by the EcNAGK and PfCK monomers to those preferentially selected by the corresponding dimers. Secondly, the modes of the monomeric and dimeric components of hexameric TmNAGK are compared to those collectively accessible in the hexameric form. Third, a detailed analysis of the softest modes accessible to the EcUMPK dimeric form is presented to shed light onto the role played by different dimeric assemblies found in the AAK family in selecting the functional motions of the family members. Overall, the different designs of interfaces and assembly geometries observed among the members of the AAK family are shown to practically define the collective modes that are being exploited by the oligomers for achieving their particular activities, including substrate binding and allosteric regulation.
How does the intrinsic dynamics of the monomeric subunits affect the oligomerization process or vice versa? To what extent the intrinsic dynamics of the monomers prevail in the oligomers? Or to what extent they are perturbed by oligomerization? To analyse these issues, we have first compared the low-frequency ANM modes of the dimeric PfCK and EcNAGK with those of their respective monomers. The two enzymes exhibit close structural similarities (Figure 2). Their sequence identity is 24%, and their ATP-binding site and catalytic sites exhibit similar structural features. In fact, our previous comparative analysis of their collective dynamics showed that the slowest three ANM modes, which essentially modulate the opening/closure of the ATP-binding site, are commonly shared between these two enzymes; and they yield an overlap of 0.75 with the experimentally observed reconfiguration from open to closed state of NAGK [34].
The main structural difference between PfCK and EcNAGK, on the other hand, resides in their amino acid substrate binding site, and here we focus on the softest modes that control those sites. In EcNAGK, the β3–β4 hairpin serves as the lid of the NAG binding site and interlinks helices B and C, which are key components of the interface (Figure 1A); in PfCK (Figure 1C), a subdomain protruding away from the interface serves as the lid of the CP binding site. This subdomain (PS) is formed by the strand β5, helix αD and hairpin β6–β7. Both lids exhibit significant conformational changes closely linked to substrate binding, as shown by the crystallographic studies performed by Rubio and co-workers [27], [32]. Among the ANM modes that affect the substrate-binding sites, those simultaneously leading to closure/opening of the substrate-binding site in both subunits will be called symmetrical modes, and others, asymmetrical (Figure 2).
In EcNAGK, the symmetrical opening/closure of the substrate-binding sites is enabled by the 5th mode (red arrows in Figures 2B and 2D; see Video S1), whereas the corresponding asymmetrical motion takes place in the 4th (green arrows) mode (Video S2). Note that our previous work [34] showed that ANM modes 1–3 were instrumental in accommodating the structural changes at the ATP-binding site, but had practically no effect on the NAG-binding site. This nicely illustrates how the enzyme takes advantage of different types of motions accessible to its native structure for achieving different types of functional motions. In mode 5, the two β3–β4 hairpins (Figure 1A), the lids of the NAG-binding sites, undergo an almost rigid-body rotation about the dyadic (z-) axis of the molecule while the ATP binding domains undergo smaller but coupled anticorrelated rotations. On the other hand, the asymmetrical motion (mode 4) induces a translation along the y axis in both lids, along with the C-terminal part of the two helices B which are connected to the lids. No symmetric opening/closing of the lids is observed about the y-axis because these movements would be prohibited by steric clashes between the two B-helices (blue arrows in Figure 2D). Rotational motions about the z-axis, on the other hand, are favored by the overall architecture of the dimeric enzyme. Indeed, tight interfacial interaction between the two B-helices is considered to be a key element for the stability of the dimer [28]. The interfacial region thus coincides with the central hinge site that mediates the opening/closing of the two monomers. This example emphasizes the effect of inter-subunit surface and topology on the character of the movements allowed/prohibited, or selected, in the oligomer.
As to PfCK, the two substrate-binding subdomains are able to undergo both symmetric (1st and 3rd mode; see Video S3) and asymmetric (4th mode; see Video S4) motions because these two subdomains protrude away from the interface and their rotational rigid-body motions are not constrained by potential clashes between the adjacent B-helices. Indeed, the motion is parallel, rather than normal, to the plane defined by the two B-helices, and the two B-helices remain tightly packed and almost immobile in these modes. Notably, the global fluctuations of two PSs on PfCK dimer appear to modulate the access to the substrate-binding sites, suggesting a role in mediating substrate-binding.
The selection of particular modes by EcNAGK for achieving its specific functions (e.g., modes 1 and 3 enabling ATP-binding; and mode 5, substrate binding) [34] raises the following question: is the rotation of the hairpins an acquired mode of motion originating from the topology of the dimer interface and not accessible to the monomer? Or, is it an intrinsic dynamical ability of the monomer that is conserved and exploited in the dimer? To address this issue, we compared the modes obtained for the isolated monomer with those of the monomer in the dimer, using the subsystem/environment coupling method described in the Methods. The monomer is the subsystem, and the second monomer stands for the environment in this case. For the sake of clarity, herein the modes that include the coupling to the environment are indicated with a superscript, i.e., monomer(dimer) refers to the behaviour of the monomer within the dimer.
The results are presented in Figure 3 (and Supplementary Tables S1 and S2). Therein the overlaps between the eight lowest-frequency modes accessible to the monomer in the isolated state (y-axis) and within the dimer (x-axis) are displayed for EcNAGK (panel A) and PfCK (panel B), and Tables S1 and S2 lists the corresponding values. The orange-red entries along the diagonal in panel A demonstrate that the modes intrinsically accessible to the EcNAGK are closely maintained in the dimeric enzyme. Notably, both the order of the modes (i.e., their relative frequency and size, as defined by the respective eigenvalues), and their shapes are closely conserved.
The picture is different in the case of the PfCK dimer (panel B). While in EcNAGK all of the top-ranking seven modes are maintained with an overlap of 0.70 or above, in PfCK significantly fewer global modes favored by the isolated monomer are maintained, and with a weaker correlation and reordering of the modes. Thus, the PfCK monomer dynamics is strongly affected by dimerization. Examination of the individual modes showed that the monomer modes that induce high fluctuations at particular secondary structural elements such as the helix B and the β10–β11 hairpin (shown in cyan in Figures 2A and C) are practically absent in the dimer. As shown in Figure 2 these are key elements at the intersubunit interface, and dimerization imposes high constraints quenching their motion. The intersubunit surface of PfCK (2453 Å2) [27] is remarkably bigger than that of EcNAGK (1279 Å2) [28]. This higher surface area, and ensuing closer association of the two monomers, may be partly responsible for the larger perturbation of the intrinsic dynamics of the monomer upon dimerization in PfCK, compared to EcNAGK.
Figure 2 and videos S3 and S4 in the Supporting Information demonstrate that the global motions preferentially undergone by the two PSs in the PfCK dimer induce conformational changes near the substrate-binding site; and Figure 3 shows that the global dimer dynamics departs from that of the isolated monomers. So, dimerization promotes in this case collective motions that affect substrate recognition and/or binding. The PS has been proposed to have evolved, together with the intersubunit interface, to play a key role in the specificity of CK for its substrate carbamate, as opposed to more abundant analogues, i.e., acetate, bicarbonate or acetylphosphate [37]. This conjecture originally inferred from the examination of crystal structure alone is supported by our examination of PfCK dynamics. ANM global modes clearly indicate the ability of the PS to undergo movements toward the substrate-binding site, and the enhanced mobility at this particular region may indeed underlie the adaptability of CK to bind its substrate.
The next case we studied is the hexameric form of the NAGK enzyme from Thermotoga maritima (TmNAGK). The higher degree of multimerization of TmNAGK will permit us to contrast the dynamics of the whole enzyme with those of its dimeric and monomeric components.
On the basis of the X-ray crystallographic structure, the hexameric arrangement of TmNAGK is considered to be a trimer of EcNAGK-like dimers [30], herein called the AB dimer (see Figures 1B and 4A). The dimeric scaffolds are interlaced by a mobile N-terminal helix, not present in the dimeric EcNAGK, and organized with a ring shape. An alternative dimeric building block being considered is the one constituted by the two monomers that interlink two adjacent AB dimers, herein called the AF dimer (see Figure 4A). In the present study, we have compared the 20 lowest-frequency modes of the hexamer with those of the monomeric subunit and the two different dimeric building blocks.
The results are presented in the panels B–F of Figure 4. In each panel, the x-axis refers to the modes observed in the oligomer (hexamer or dimer), and the y-axis refers to those intrinsically accessible to the components (dimers or monomers) that make these oligomers, e.g., panel B compares the global modes of the AB dimer in the hexamer (x-axis) to those accessible to the AB dimer itself when examined in isolation (y-axis). The comparative examination of these maps discloses two distinctive patterns: panels C and E reveal the conservation of global modes, in general, between the entities that are being compared, while panels B, D and F reveal that about ½ of the modes accessible to the substructures when examined in isolation are not represented in the assemblies. This behavior is clearly seen, and quantified, by the dashed lines on the maps, which represent a linear fit by weighted least squares regression to the entries that exhibit a correlation of 0.5 of higher. The dashed line in the former groups lies along the diagonal (slope -1.04 and -1.01 in the respective panels C and E), whereas in the latter case, the slope varies as -1.81 (panel B), -1.72 (D) and -1.44 (F).
Let us first examine the 1st group more closely: panel C essentially tells us that the monomers participating in the AB dimer maintain in the dimer their intrinsic dynamics favored by their monomeric architecture. As to panel E, it simply reflects that AF dimer in the hexamer behaves practically in the same way as in the isolated AF dimer, indicating that multimerization does not alter the global dynamics favored by the AF dimeric structure. In other words, the TmNAGK hexamer exploits the intrinsic dynamics of the AF dimer; and likewise, the AB dimer takes advantage of the structure-encoded dynamics of its monomers. Notably, the top four modes are conserved in this case with a correlation of more than 0.95. This is in agreement with the high conservation of the monomer dynamics in the EcNAGK dimer, as pointed out in Figure 3A, given the structural and dynamical similarities [34] between the AB dimer and EcNAGK.
We now turn our attention to the 2nd group. Here we see the dimer AB in the hexamer which is unable to sample several modes that are accessible to the same dimer in isolation (panel B). Thus, the environment provided by the hexamer constrains the intrinsic dynamics of the AB dimer. Why is the AB dimer rigidified in the hexamer? We note that in the hexamer, these EcNAGK-like (AB) dimers make close, interlacing interactions with the adjacent dimer by swapping their N-terminal helices and also making contacts with the C-domain, i.e. the interactions of AB-type dimers with the adjacent dimer through the AF interface impose topological constraints that impair several modes in the hexamer (panel B). Likewise, the monomer in the hexameric environment is more restricted than the isolated monomer, such that many modes accessible to the isolated monomer cannot be effectuated in the hexamer (panel D). Given the different degree of conservation of the dynamics of the AB and AF dimers within the hexamer (panels B and E), we can add a complementary perspective to the structural view of TmNAGK as a trimer of EcNAGK-like dimers. The stronger conservation of the dynamics of the AF dimer supports a dynamical view of TmNAGK as a trimer of AF-like dimers.
Finally, it is worth pointing out that the surface area of the AF interface (1186 Å2) is slightly smaller than that of the AB interface (1381 Å2) [30]. This might suggest that the monomeric modes would be more severely constrained in the AB dimer, but this does not hold true as explained above. The small difference in the surface area is therefore not sufficient to explain the observed behavior. The major determinant of accessible global motions is not the surface area but the topology of the interfacial contacts, or the overall shape/architecture of the dimer. In the present case, the overall architecture of the hexamer selectively hinders a number of global modes accessible to the AB dimer, while those of the AF dimer are mostly preserved. It is widely accepted that the size of the interface is closely linked to the thermodynamic stability of the oligomer [25], [55]. The dynamics of the oligomer, on the other hand, is suggested by the present analysis to be predominantly controlled by the quaternary arrangement and contact topology of the subunits.
The results discussed above focus on the preservation or the obstruction of the global motions of the subunits upon oligomerization. Nevertheless, in many cases, oligomeric proteins are subject to cooperative processes that regulate the biological activity. This raises the question whether such cooperative processes are linked to new modes of motion unique to oligomeric arrangement.
TmNAGK is cooperatively inhibited by arginine in contrast to the dimeric EcNAGK and PfCK, which do not exhibit an allosteric regulation. The available X-ray crystallographic structure of TmNAGK represents the T state of the enzyme, which is bound to arginine. The apo form of the enzyme (R state) has not been structurally resolved, but the X-ray structure of the same enzyme from Pseudomonas aeruginosa (PaNAGK) serves as a suitable model for the R state on the basis of sequence and structural similarities [30]. Taking into account that the transition of TmNAGK between the R and T states is intimately linked to its allosteric regulation, those modes of motion that favor this conformational change will be the most functional. Therefore, the cumulative overlap of the lowest modes with the deformation vector between the R and T states has been calculated. Given that the T and R states correspond to proteins with different sequences, we have structurally aligned the two structures with DALI [56] and used the subsystem/environment coupling method (see Methods) to compute the ANM modes of TmNAGK, considering as subsystem those residues of TmNAGK structurally aligned to PaNAGK. Likewise, the deformation vector was calculated for the structurally aligned residues.
Strikingly, a single non-degenerate mode (6th) accessible to TmNAGK is found to describe 75% of the R↔T deformation (see Figure 5D showing the cumulative overlap). A deeper analysis of this mode can shed light on the structural origin of the functionality of this enzyme. The aim is to ascertain whether this mode arises from the intrinsic dynamics of the subunits or is acquired in the hexameric state. Mode 6 is an expansion/contraction of the ring, accompanied by cooperative rotational and twisting motions of each monomer (see Video S5). The axis of rotation goes through each AF interface (Figure 5A) and performs an almost rigid rotation of the EcNAGK-like dimers (Figure 5C). Residues close to these axes of rotation form minima in the mode fluctuations profile (Figure 5B) and belong to the AF interface. The axis involves a part of the N-terminal helix (6–20) of chains A and F, where the two helices interact tightly. Indeed, this interface stabilizes the hexameric arrangement and no NAGK dimer has been structurally characterized with an AF-like interface. The AF interface is unique to the hexameric arrangement.
As shown in Figure 4, the hexamer dynamics is affected by the intrinsic dynamics of the component subunits. Therefore, mode 6 could be associated with particular global modes accessible to the AB and/or AF dimers. We have examined the inter-residue distance variations maps induced by the low-frequency modes of the isolated AB and AF dimers to explore this possibility. AF dimer proves to be the major source of the rigid body movements of monomers observed in the hexamer (see Videos S6 and S7). The distance variation maps of the 1st and 4th modes of the AF dimer (Figure S1) illustrate that the internal motions within a given subunit are negligible, but the relative movements between the two subunits are significant. The AF interface, thus, emerges as a key mechanical region that confers to the two linked subunits suitable flexibility to undergo functional changes in their relative orientations. This dynamic feature of the AF interface, whose size is smaller than the AB interface, is in accord with Hubbard and co-workers [57], who stated that those interfaces that are not optimally packed may confer functional mobility to the oligomer. This inherent dynamical ability of the AF interface is therefore exploited in the hexameric arrangement to couple the rigid-body movements of the subunits, complementing their intrinsic internal dynamics.
The topology of the AF interface appears to be evolutionary selected to provide two essential features for the functionality of the enzyme: (1) flexibility to allow for the cooperative reorientations of the dimers, which is inextricably linked to allostery, and (2) thermodynamic stability of the whole hexamer. Taking into account the crucial role of the AF interface and with the aim of providing further insights into the allosteric regulation of this enzyme, we considered the maximum likelihood pathway (MLP) for each combination of pairs of residues (endpoints) belonging to the respective chains A and F, and evaluated the fractional occurrence of each residue in the ensemble of MLPs (see Methods). Figure 6A displays the percent occurrence of each residue, which also provides a measure of the relative allosteric potential of the residues. Peaks are observed at K17, E18, F19, Y20, K50 and Y51 (ribbon diagram color-coded from blue (peaks) to red (minima) in Figure 6B). The significance of this first set in allosteric communication could be anticipated due to their location at the tightest part of the AF interface and proximity to the arginine inhibitor (Figure 6B). However, our approach helps to identify other distal residues important for the communication, which behave as hubs. In particular, K196 and I162 channel most of the pathways to the AF interface via interactions with F19 (and the arginine inhibitor) and K50, respectively.
The communication across the AF interface can be summarized namely by two symmetric pathways distinguished by the MLP analysis: I162A → K50A→ Y51A→ K17F → E18F → F19F → K196F and its counterpart I162F →…→ K196A (colored yellow and green in Figure 6B). Aromatic residues tend to be favored at protein interfaces [25], and in this case, F19 and Y20 play a critical role. Not surprisingly, F19 is highly conserved among arginine-sensitive NAGKs [30] and, together with Y20 (violet in Figure 6B), it establishes an efficient communication pathway of the form F19(A/F)→ Y20(A/F)→ Y20(F/A)→ F19(F/A).
The structure of the monomeric subunit of EcNAGK is preserved among all family members, but the assembly geometry is less conserved. The arrangement of the monomeric subunits of NAGKs and CKs is strikingly similar, as shown above, but has significant differences with the assembly of UMP Kinases. Structurally, UMPKs are trimers of dimers in which the two helices that build the intersubunit surface of each dimer are parallel (Figure 7C and D), whereas in NAGK (and CK) these helices at the interface make an angle of ∼65° (Figure 7A and B). To our knowledge, a clear functional reason for this difference in monomer-monomer packing has not been reported so far. Although this difference has been argued to be necessary for hexameric assembly [58], there might be another functional reason since TmNAGK is an example of a hexameric assembly that selectively adapts the EcNAGK-like dimer packing (AB dimer). Here we compute the ANM modes of the UPMK dimer from Escherichia Coli (EcUMPK) in order to examine whether such a difference in packing geometry gives rise to significant changes in the global dynamics.
The first mode of motion of the isolated EcUMPK dimer entails a rotational rigid-body movement with respect to an axis across the αC helices (Figure 7, panels C and D, and Video S8). The anticorrelated motion of both subunits leads to an opening/closure movement of the whole dimer. This is in sharp contrast to the EcNAGK dimer dynamics, whose low-frequency modes do not exhibit rigid-body movements of the subunits. Does this dynamic feature of the EcUMPK dimer play a functional role?
Gilles and co-workers determined the X-ray crystal structure of EcUMPK complexed with GTP (PDB code 2VRY) [59], which is an allosteric activator, and characterized a functional conformational change. They argued that GTP induces a rearrangement of the quaternary structure that involves a rigid-body rotation of 11° that opens the UMPK dimer. Strikingly, the first ANM mode predicted for the UDP-bound dimer describes the structural transition between the UDP- and GTP-bound forms. The overlap is outstandingly high (0.78) (see Figure 8E for cumulative overlap). Moreover, it is worth pointing out that we have checked that this mode of motion is totally conserved in the hexamer (see Figure S2).
Why does the different assembly in the UMPK dimer give rise to a normal mode with a rigid-body character not present in EcNAGK? In UMPK the interface between the monomers is constituted mainly by two long parallel helices (αC) able to build a rotational axis that promotes an en bloc motion of both subunits. In contrast, the crossed orientation of the helices of NAGK (∼65°) and the presence of other intersubunit contacts (B-helices and β9–β10 hairpins) hinders a rigid-body rotation of the two subunits. This suggests that the unique dimeric assembly of UMPK gives rise to a particular soft mode not present in other AAK family members. This example further indicates that the design of the interfacial contact topology and oligomerization geometry is crucial in defining the functional mechanisms of oligomers.
In some cases, a single residue may significantly affect the contact topology at the interface and, thus, the allosteric regulation. This has been explored in the context of the UMPK analogue from Mycobacterium tuberculosis (MtUMPK), for which crystallographic and site-directed mutagenesis studies have been recently conducted [60]. The X-ray structure of MtUMPK bound to GTP shows striking similarities to EcUMPK structure. Notably, this similarity is extended to their global motions: the lowest frequency ANM modes of the two structures exhibit an overlap of 0.97. Given that the global modes of motion are fully determined by the overall shape of the protein, local perturbations are indeed unlikely to affect the low-frequency modes.
Site-directed mutagenesis studies, on the other hand, show the importance of some residues in both the activity and the cooperativity of the enzyme. Among them, P139 was pointed out to to be a key residue in the allosteric regulation of the enzyme. P139 is located close to the trimeric interface where three GTP molecules are bound. What is the dynamical role of this residue? The mean-square fluctuations profile obtained with the ANM shows that P139 occupies a position close to a local minimum (a rigid part of the protein) (Figure 8A). Such regions usually play a key mechanical role for mediating collective changes in structure, and mutations at such positions may potentially affect the allosteric dynamics of the protein.
We have analyzed the importance of P139 in mediating the allosteric communication among subunits A, D and F, which build one of the two trimeric interfaces where three GTP molecules are bound. We computed the communication pathways between GTP binding residues (starting from subunit A and ending at subunits D and F) and the percent contribution of each residue to MLPs, as done for TmNAGK. Figure 8 shows the trimeric interface color-coded according to the percent contribution in the same way as in Figure 6B. We note that the participation of P139 (in yellow) to these pathways is minimal (note the red color in the backbone), but the adjacent residues Y137 and L138 are important mediators of inter-subunit communication via interactions with Q132.
This analysis suggests that the importance of P139 lies in constraining the orientation of nearby residues Y137 and L138 involved in inter-subunit signal propagation. The fact that this residue is highly restricted position in the global mode profile emphasizes its role in constraining the neighboring residues in a precise orientation pre-disposed to enable inter-subunit communication. The experimentally tested mutants (P139A, P139W and P139H) all showed a diminished allosteric regulation, but to different extents [60]. Further simulations at atomic scale might help explain the relative sizes of the effects induced by these mutations, but this is beyond the scope of the present work. It might be interesting to experimentally test the effect of mutations at L18, Y137 and Q132, since these residues emerge here as key elements enabling inter-subunit communication and they are distinctly restricted in the collective dynamics (Figure 8A) despite the relatively low packing density at the interface.
To summarize, the present study reveals several dynamic features of oligomeric proteins by means of an ENM analysis of family members with different degrees of oligomerization. A common dynamic feature of the oligomers presented here is the conservation of the inherent dynamics of their monomeric or dimeric building blocks. The way these blocks are assembled in different oligomers confers different types of collective mechanisms unique to particular oligomerization geometries. Here are the main observations:
In summary, the oligomers in the examined AAK family appear to selectively exploit the inherent dynamic abilities of its components, on the one hand, and favor coupled movements of intact subunits, on the other, to effectively sample cooperative movements (soft modes) that enable motions required for substrate binding and efficient allosteric responses. The architecture of the interfaces and the assembly geometry play an essential role in defining the most easily accessible (or softest) modes of motion, which in turn, are shown to be relevant to the functional mechanisms of the different oligomers, being presumably optimized by evolutionary pressure.
The low-frequency modes described by the NMA of different ENM variants [40], [61]–[64] have proven to be robustly determined by the overall fold [7], [65], [66] and provide a consistent description of the conformational space most easily accessible to the protein [67]. Among them, we use here the most broadly used model, the anisotropic network model (ANM) [40], [41]. In the ANM, the network nodes are located at the Cα-atoms' positions, and pairs of nodes within close proximity (a cutoff distance of 15 Å, including bonded or non-bonded pairs of amino acids [41]) are connected by springs of uniform force constant γ. The interaction potential of the molecule is given by(1)
where M is the number of springs, and |Rij|-|Rij0| is the inter-residue distance with respect to the equilibrium (crystal) structure. The second derivatives of VANM with respect to residue displacements yield the 3Nx3N Hessian matrix H, the eigenvalue decomposition of which yields 3N-6 nonzero eigenvalues λk and eigenvectors uk corresponding to the frequencies (squared) and shapes of the normal modes of motion accessible to the examined structure. Numbering of modes in this work starts from the first mode with a nonzero eigenvalue.
The cross-correlation between the displacements of residues i and j, contributed by mode k scales as(2)
where the subscript ij designates the element of the matrix in square brackets. For i = j, equation (2) reduces to the square displacement of residue i in mode k. Clearly, lower-frequency modes (smaller λk) drive larger-amplitude motions.
Conformations sampled upon moving along mode k are generated using(3)
where R0 is the 3N-dimensional vector representing the initial coordinates of all residues and s is a parameter that rescales the amplitude of the deformation induced by mode k. The movies S1-S8 in the Supporting Information are generated using this equation with a series of different s values for selected modes of examined proteins.
The degree of overlap between a conformational change Δr observed by X-ray crystallography and the structural change predicted by the ANM to take place along mode k is quantified by (Δr · uk)/|Δr|. Here Δr is the 3N-dimensional difference vector between the α-carbon coordinates of two different forms resolved for the same protein under different conditions (e.g., substrate-bound and -unbound forms of enzymes, or inward-facing or outward-facing forms of transporters). The cumulative overlap CO(m) between Δr and the directions spanned by a subset of m modes is calculated as (4)
CO(m) sums up to unity for m = 3N-6, as the eigenvectors form a complete orthonormal set of basis vectors in the 3N-6 dimensional space of internal conformational changes (see Figures 5D and 7E)
The similarity between the conformational spaces described by two subsets of m and n modes, uk and vl, evaluated for two different systems can be quantified in terms of a double summation over squared overlaps as in Eq. 4, among all mxn pairs of modes (divided by m or n, depending on the reference set). The overlap O(uk,vl,) between the pairs of modes uk and vl calculated for different systems (e.g., Figure 3) is given by the inner product of the eigenvectors, i.e., (5)
Note that O(uk,vl,) is equal to the correlation cosine between the two N-dimensional vectors, since the eigenvectors are normalized.
The change in a given inter-residue distance |R0ij| induced by a given mode k, , is given by the projection of the deformation induced by the kth mode onto the normalized distance vector, scaled by the inverse frequency, (6)
Here (uk)i designates the ith super element (a 3D vector) of uk, and describes the relative displacement of the ith residue (x-, y-, and z-components) along the kth mode direction.
Inter-residue communication has been suggested to play a key role in allosteric regulation and enzymatic catalysis [68], [69], and has been the subject of many computational studies [48], [70]–[72]. Here we use a Markov model of network communication [73], [74] to identify communication pathways. The interactions between residue pairs connected in the ANM are defined by the affinity matrix A, whose elements are aij = Nij/(Ni Nj)½ where Nij is the number of atom-atom contacts between residues i and j based on a cutoff distance of 4 Å, and Ni is the number of heavy atoms belonging to residue i. The density of contacts at each node i is given by .The Markov transition matrix M = {mij}, where mij = aij/dj, determines the conditional probability of transmitting a signal from residue j to residue i in one time step [73]. We define –log(mij) as the corresponding ‘distance’. The maximum-likelihood paths (MLPs) for signal transfer between two end points are evaluated using the Dijkstra's algorithm [73]. In order to identify the residues that play a key role in establishing the communication between pairs of subunits, we considered the communication between all pairs of residues belonging to the two subunits of interest. In the application to the communication between the A and F subunits of TmNAGK (Figure 6), an ensemble of N2 = 2822 combinations of residue pairs (endpoints) have thus been considered (each chain consists of N = 282 residues). For each pair, we evaluated the MLP and thus determined the series of residues taking part in the MLP. To quantify the contribution of a given residue to intersubunit communication, we counted the occurrence of each residue in the complete ensemble of MLPs. Figure 6, panel A displays the resulting curve, peaks indicating the residues that make the largest contribution.
In many applications the dynamics of a part of the protein (subsystem, S) may be of interest in the context of its environment (E). The Hessian of the whole system is conveniently partitioned into four submatrices [75], [76]:(7)
where HSS is the Hessian submatrix for the subsystem, HEE is that of the environment and HSE (or HES) refers to the coupling between the subsystem and the environment. Inasmuch as the environment responds to the subsystem structural changes by minimizing the total energy, the effective Hessian for the subsystem coupled to the environment is (8)
This approach has been advantageously employed in determining potential allosteric sites [77] and locating transition states of chemical reactions [78]. It will be used below in conjunction with the ANM for assessing the effect of oligomerization on the dynamics of monomeric and/or dimeric components (subsystem).
We examined four enzymes belonging to the AAK family (Figure 1): EcNAGK (dimer), TmNAGK (hexamer), PfCK (dimer) and EcUMPK (hexamer). To this aim, we use the X-ray structures of EcNAGK in the open state (PDB code: 2WXB), the arginine-bound TmNAGK (PDB code: 2BTY), the ADP-bound PfCK (PDB code: 1E19) and the UDP-bound EcUMPK (PDB code: 2BND).
All diagrams of molecular structures have been generated using VMD [79].
|
10.1371/journal.pntd.0003755 | Deletion of Fibrinogen-like Protein 2 (FGL-2), a Novel CD4+ CD25+ Treg Effector Molecule, Leads to Improved Control of Echinococcus multilocularis Infection in Mice | The growth potential of the tumor-like Echinococcus multilocularis metacestode (causing alveolar echinococcosis, AE) is directly linked to the nature/function of the periparasitic host immune-mediated processes. We previously showed that Fibrinogen-like-protein 2 (FGL2), a novel CD4+CD25+ Treg effector molecule, was over-expressed in the liver of mice experimentally infected with E. multilocularis. However, little is known about its contribution to the control of this chronic helminth infection.
Key parameters for infection outcome in E. multilocularis-infected fgl2-/- (AE-fgl2-/-) and wild type (AE-WT) mice at 1 and 4 month(s) post-infection were (i) parasite load (i. e. wet weight of parasitic metacestode tissue), and (ii) parasite cell proliferation as assessed by determining E. multilocularis 14-3-3 gene expression levels. Serum FGL2 levels were measured by ELISA. Spleen cells cultured with ConA for 48h or with E. multilocularis Vesicle Fluid (VF) for 96h were analyzed ex-vivo and in-vitro. In addition, spleen cells from non-infected WT mice were cultured with rFGL2/anti-FGL2 or rIL-17A/anti-IL-17A for further functional studies. For Treg-immune-suppression-assays, purified CD4+CD25+ Treg suspensions were incubated with CD4+ effector T cells in the presence of ConA and irradiated spleen cells as APCs. Flow cytometry and qRT-PCR were used to assess Treg, Th17-, Th1-, Th2-type immune responses and maturation of dendritic cells. We showed that AE-fgl2-/- mice exhibited (as compared to AE-WT-animals) (a) a significantly lower parasite load with reduced proliferation activity, (b) an increased T cell proliferative response to ConA, (c) reduced Treg numbers and function, and (d) a persistent capacity of Th1 polarization and DC maturation.
FGL2 appears as one of the key players in immune regulatory processes favoring metacestode survival by promoting Treg cell activity and IL-17A production that contributes to FGL2-regulation. Prospectively, targeting FGL2 could be an option to develop an immunotherapy against AE and other chronic parasitic diseases.
| In larval E. multilocularis infection causing alveolar echinococcosis (AE) in humans as well as mice, immune tolerance and/or down-regulation of protective immunity is a marked characteristic of this chronic disease. Our study provides a comprehensive evidence for a major involvement of the recently identified CD4+ CD25+ Regulatory T Cell Effector Molecule FGL2 to the outcome of AE. Our major findings are as follows: 1) FGL2 is mostly secreted by Tregs and partly contributes to their functions; 2) FGL2 can down-regulate the maturation of DCs, suppress Th1 and Th17 immune responses, and support Th2 and Treg immune responses, and finally 3) IL-17A contributes to FGL2 secretion. Based on the present findings in mice, we will investigate FGL2 as a potential marker of progression of AE in human patients, or as a potential immunotherapeutical target. Early prediction of parasite regression (currently not yet possible) would allow clinicians to plan for withdrawing benzimidazole treatment, which is currently administered for life. Then, FGL2 should be investigated as a target for an anticipated immunomodulatory treatment of patients with progressive AE, especially of those who are non- or low-responders to benzimidazole treatment, or who suffer from side-effects due to chemotherapy.
| Alveolar echinococcosis (AE) is a very severe zoonotic helminthic disease in humans, exhibiting a fatal outcome if remaining untreated [1]. AE is characterized by chronic and progressive hepatic damage caused by the continuous proliferation of the larval stage (metacestode) of Echinococcus multilocularis [2], that behaves like a slowly growing liver cancer, progressively invading host tissues and organs [3].
During E. multilocularis infections in humans, a Th2-oriented immunity is basically associated with increased susceptibility to disease leading to chronic AE, while Th1 cell activation has been linked to protectivity, which may even yield aborted ("died-out") forms of infection [2,3]. Experimental murine AE is characterized, as studied in spleen or lymph node cells, by an initial Th1 response during the early stage of infection (till 1 month p.i.) that gradually switches to a more dominant Th2-biased response during the chronic phase of AE (2–4 months p.i.). Nevertheless, this mostly mixed Th1/Th2 profile, characterized by the concomitant presence of IL-12α, IFN-γ and IL-4 at the very early stage of E. multilocularis infection [4], is associated with the expression of pro-inflammatory cytokines in the periparasitic granuloma and partial/relative protective immunity (restriction of parasite growth) through fibrosis and necrosis [5]. It has been previously reported that CD4+CD25+ T regulatory cells (Tregs) play a critical role in human AE by blunting immune responses to specific antigens, or by suppressing the secretion of proinflammatory cytokines, especially through interleukin (IL)-10 and transforming growth factor beta1 (TGF-β1) [6]. Moreover, increased CD4+CD25+ Tregs were also observed in peritoneal cells of mice intraperitoneally (i.p.) infected with E. multilocularis, a finding that concurred with other findings demonstrating that E. multilocularis antigens promote T cell differentiation into Treg cells [7].
Previous microarray analyses showed that expression of mRNA coding for the fibrinogen-like protein 2 (FGL2) were significantly up-regulated in the liver of mice perorally infected with E. multilocularis eggs [8]. FGL2, a member of the fibrinogen-related superfamily of proteins secreted by T cells, has recently been reported by a number of groups to be highly expressed in Tregs. Its role was associated to Treg effector functions [9,10]. It was shown that FGL2 could inhibit dendritic cell (DC) maturation through binding to the low-affinity FcgammaRIIB receptor, and thus contribute to Treg activity [11]. There is evidence that FGL2 exerts an immunosuppressive effect on T cell proliferation. Thus, FGL2 seems to play an important role both in innate and adaptive immunity, by the fact to be expressed by activated CD4+ and CD8+ T cells, and reticulo-endothelial cells as well [12–17]. FGL2 has been propagated as a novel cancer biomarker, and was shown to be involved in the pathogenesis of inflammatory disorders such as allo- and xenograft rejection [12,18–22] and cytokine-induced fetal loss [23], as well as in the pathogenesis of infectious diseases, such as viral hepatitis [14,17]. However, nothing is known about FGL2 and its potential role in parasite-induced immunotolerance.
The major aims of this work were thus: 1) to study the role of FGL2 in T cell reactivity as well as its effect on the maturation of DCs in an early time-point and a late stage of E. multilocularis infection in fgl2 knock-out (fgl2-/-) mice; 2) to elucidate how parasite components, i.e. metabolites represented by those expressed in the VF of the E. multilocularis metacestode, affect the immune response in fgl2-/- mice; 3) to explore how FGL2 is secreted during the course of E. multilocularis infection; and 4) to provide a comprehensive picture of the various cell and molecular components involved in the regulation of the peritoneal periparasitic immune cell infiltrate, and likewise in the spleen as a key immune organ. To achieve these goals, Th1/Th2-related and Treg/Th17 related cytokines, the maturation of DCs, and the generation of Tregs and their functions were studied at the different disease stages in an experimental model with active or knocked-out FGL2-expression.
The animal study was performed in strict accordance with the recommendations of the Swiss Guidelines for the Care and Use of Laboratory Animals. The protocol was approved by the Commission for Animal Experimentation of the Canton of Bern (approval no. BE_103/11).
8-week-old female C57/BL6 (wild type [WT]) and C57/BL6 fgl2-/- mice [24] were bred and housed in specific-pathogen-free (SPF) facilities according to recommendations of FELASA, and monitored by daily assessment of health status, putative weight loss or gain during the experiments. E. multilocularis metacestodes (clone KF5) were maintained by serial passages (vegetative transfer) in C57BL/6 mice [25] and injected intraperitoneally as previously described [26,27]. Each experimental group included 6 animals unless otherwise stated. Control mice (mock-infection) received 100 μL of RPMI-1640 only. Mice were sacrificed at 1 or 4 month(s) post-infection, corresponding to early and late stage of disease, respectively. Parasite tissues were surgically recovered and, if present, fat and connective tissues were carefully removed for subsequent wet-weight determination of the parasite mass.
Total RNA was extracted from parasite tissue previously put into TRIzol (Invitrogen) according to the manufacturer’s instructions. cDNA was synthesized using the Omniscript Reverse Transcription kit (Qiagen, Hilden, Germany). SYBR-Green Mix-based qRT-PCR was carried out on a Rotor-Gene 6000 qPCR detection system (Corbett) with the FastStart Essential DNA Green Master (Roche, Basel, Switzerland) following the manufacturer’s instructions. PCR cycling was performed in triplicates in final volumes of 20 μL containing 2 μL cDNA and 10 pM of each primer (Cycle scheme: initial denaturation at 95°C—15 min, 45 cycles of 95°C—15 s, 55°C—30 s and 72°C—30 s). Fluorescence was measured in every cycle, and a melting curve was analyzed after the PCR by increasing the temperature from 55 to 95°C in 0.5°C increments. The primers used were described earlier [28], and em14-3-3 mRNA levels were quantified relative to the mRNA level of a parasite housekeeping gene, the β-actin homologue E. multilocularis. Respective mean values from triplicate determinations from 6 individual mice in each group were taken for the calculation of relative emII/3 and em14-3-3 mRNA levels in relation to em-β-actin mRNA levels).
Peritoneal exudate cells (PEC) and spleen cells were collected by peritoneal rinsing or grinding in 5 mL RPMI-1640 (Gibco, Basel, Switzerland) and incubation of PEC or spleen cell suspensions in 15 mL RPMI-1640 +20%FCS in a petridish for 2 h at 37°C 5%CO2, as described earlier [25]. The non-adherent cells were collected, and highly (N99%) enriched iTreg cells were obtained by MACS (magnetic cell-separation) using the mouse CD4+CD25+ T cell Isolation-Kit (Miltenyi Biotec, Germany) followed by FACS.
In vitro suppression assays were carried out with cultures of 2×104 CD4+CD25- T effector (Teff) cells from WT-mice as responder cells, together with 8×104 irradiated spleen cells as APCs and titrated numbers of CD4+CD25+ Treg cells from either E. multilocularis-infected AE-fgl2-/- or AE-WT mice as suppressor-cells, compared with non-infected controls. For rFGL2/antibody blockade studies, 1 μg/mL mouse rFGL2 or anti-FGL2- (monoclonal IgG2a; Abnova, Luzern, Switzerland) were added to the cell cultures at a 1:1 Treg:Teff ratio in the presence of APCs and ConA (2μg/mL). Cell proliferation was assayed using the colorimetric BrdU cell proliferation ELISA kit (Calbiochem, Merck, Switzerland) according to manufacturer’s instructions.
Spleen cells were cultured at a density of 2×106 cells/mL in RPMI-1640 +10%FCS. For assessment of the effects of stimulation by recombinant FGL2 (rFGL2) or anti-FGL2 monoclonal antibodies (anti-FGL2-MAb), (both from Sigma-Aldrich, Basel, Switzerland) they were incubated with 1 and 5 μg/mL rFGL2 or 1 μg/mL anti-FGL2-MAb for 48h in the presence of a protein transport inhibitor cocktail (Ebioscience, San Diego, CA, USA) for cytokine staining. Negative control reactions were performed without rFGL2 or without anti-FGL2-MAb. The effects of recombinant IL-17A (rIL-17A) anti-IL-17A antibodies (both from Sigma), were assessed by stimulation of cells with 0.5, 1, 2 and 4 μg/mL rIL-17A or 1 μg/mL anti-IL-17A for 48h, while negative control reactions were performed without rIL-17A or anti-IL-17A antibodies. FGL2-levels in the serum and supernatant from cell cultures were measured by sandwich ELISA (Biolegend, San Diego, CA) according to manufacturer’s instructions. Spleen cell cultures were also stimulated with 2 μg/mL ConA for 48h, or with 10μg/mL of VF for 96 h, in the presence of protein transport inhibitor cocktail for cytokine staining. The same cell reactions performed without VF were used as negative controls.
PEC or spleen cells were incubated with 1 μg of purified anti-CD16/CD32 for 20 min in the dark to block non-specific binding of antibodies to the FcγIII/II receptors, cells were then stained with surface markers separately for 15 min with 1 μg of primary antibodies: FITC-labeled anti-CD80, anti-CD86, anti-CD25; PE-labeled anti-CD11b, anti-CD11c, PECy 5.5-labeled anti-CD4. For intracellular staining, the cells were first incubated with Inside-Fix (Miltenyi, Bergisch Gladbach, Germany) for 20 min at room temperature then stained with PE-labeled anti-IFN-r, anti-IL-4, anti-IL-17A, anti-IL-2, anti-IL-10 and anti-Foxp3 in Inside-Perm (Miltenyi, Bergisch Gladbach, Germany) for 15 min in the dark. Corresponding fluorochrome-labeled isotype control antibodies were used for staining controls. For each sample, a minimum of 500,000 cells were acquired using a FACS LSRII flow cytometer and analyzed using FlowJo software (Tree Star, OR, USA), employing the gating strategy shown in S1 Fig. All antibodies were purchased from BD Pharmingen (Palo Alto, USA).
All data were analyzed by SPSS 17.0. The results are presented as means ± SD. Normality of data was assessed by D’Agostino & Pesrson and Shapiro-Willk test. For normally distributed groups of data, One-way-ANOVA followed by Bonferroni’s post-test or unpaired two-tail Student’s t-test were used to compare the differences between groups, and two-tail Spearman’s rho was used to analyze the correlation coefficient. Significance was defined as P<0.05 for all tests, except those subsequently corrected by Bonferroni.
The levels of FGL2 expression in sera of E. multilocularis infected (AE-WT) mice were significantly higher, both at 1 and at 4 month(s) post-infection, when compared to non-infected WT-controls (Fig 1A). Spearman correlation coefficients indicated a positive correlation between serum IL-17A level and FGL2 (r = 0.435, P = 0.045) in AE-WT-mice. To examine whether IL-17A contributes to FGL2-secretion, spleen cells from non-infected WT-mice were co-cultured either with recombinant IL-17A (rIL-17A) as an external stimulus, or with anti-IL-17A antibodies, and FGL2 levels were quantitatively analyzed in respective culture supernatants by ELISA. As shown in Fig 1B, addition of rIL-17A lead to increased FGL2-secretion in a dose-dependent manner, while the addition of anti-IL-17A had no effect (Fig 1B).
To characterize the role of FGL2 in the control of parasite growth, E. multilocularis-infected fgl2-/- mice and control WT littermates were analyzed after 1 and 4 months p.i. with respect to parasite weight and the expression of em14-3-3 as a marker for cellular proliferation activity [24]. At the late stage of infection (4 months- p.i.), fgl2-/- mice exhibited a significantly lower parasite load compared to WT mice (Fig 1C and 1D), and 14-3-3 expression levels in AE-fgl2-/- mice were significantly lower those in AE-WT mice (Fig 1E). Moreover, the parasite invaded the liver (a marker of pathogenicity) in only 33.3% of the AE-fgl2-/- mice compared to 94.4% of AE-WT mice.
To explore the cellular source of secreted FGL2, CD4+ effector T cells (Teffs), CD8+ T cells, CD4+CD25+ Tregs, antigen presenting cells (APCs) were FACS sorted from spleen cell suspensions of AE-WT mice, and non-infected control mice. Quantitative RT-PCR showed that fgl2 mRNA-levels were significantly increased in CD4+CD25+ Tregs derived from AE-WT mice, when compared to non-infected controls, while no significant changes were evident with respect to CD4+ Teffs, CD8+ T cells. A slight decrease in FGL2 mRNA-levels were noted in APCs from AE-WT mice compared to those in non-infected mice (Fig 2A).
The contribution of FGL2 in the generation and maintenance of Tregs in AE-WT and AE-fgl2-/- mice was analyzed ex vivo, and in vitro by either addition of rFGL2 to spleen cells or treating cultures with anti-FGL2 antibodies. At 4 months p.i., an increased frequency of Foxp3+/CD4+ CD25+ cells could be observed in PECs as well as spleen cells from AE-WT mice, compared to respective preparations in AE-fgl2-/- mice (P<0.05) (Fig 2B and 2C). In addition, expression levels of Foxp3 and IL-10-transcripts were significantly increased in PECs from AE-WT mice (Fig 2D). Moreover, when in vitro cultured PECs from AE-WT mice were exposed to anti-FGL2-MAbs and stimulated with VF, the frequency of CD4+CD25+Foxp3+ cells was decreased compared to PECs cultured in the absence of anti-FGL2-MAbs (Fig 2E). This indicated that in the absence of FGL2 E. multilocularis metabolic components might exert immune-modulatory activities.
We then assessed the effect of the targeted deletion of fgl2 on the ability of Treg cells to suppress the proliferation of Teffs. Treg cells from either non-infected or AE-fgl2-/- mice, were less efficient in suppressing normal CD4+ Teff cell proliferation when compared to Treg cells from WT-mice (S2 Fig). To further study the role of FGL2 regarding Treg functions, spleen cells from AE-WT mice were exposed to either anti-FGL2-MAb or rFGL2. In response to rFGL2, Tregs from AE-WT mice inhibited ConA-induced CD4+ Teff proliferation; conversely, the same Tregs were not able to inhibit CD4+ Teff proliferation in the presence of anti-FGL2-MAbs (Fig 2F).
To further explore the effects of FGL2 on the immune response during E. multilocularis infection, T cell functions, in AE-WT and AE-fgl2-/- mice were comparatively assessed. Purified splenic CD4+ T cells from AE-fgl2-/-mice exhibited an increased proliferation in response to ConA, as compared to splenic CD4+ T cells from AE-WT mice (P<0.01) (S3A and S3B Fig). To further study the role of FGL2 in Tcell proliferation, splenic CD4+ T cells from WT mice were cultured in the presence and absence of rFGL2. CD4+ Teffs showed a pronounced proliferation in response to ConA stimulation, which was inhibited by the addition of rFGL2 (S3C Fig). Furthermore, T helper (Th) cells from AE-fgl2-/- mice appeared oriented towards a lower Th2 response at early stage of infection (1 months p.i.), and a stronger Th1-response at late stage of infection (4 months p.i.) (Fig 3A–3C and 3E). This dichotomic polarization was, however, not confirmed by in vitro cultivation of the cells in the presence of ConA stimulation (Fig 3D). Respective flow cytometric analyses of CD4+ T cells in spleen from AE-fgl2-/- mice and AE-WT mice showed that there was no difference in expression of IFN-γ and IL-4 at 48 h after of exposure to ConA (Fig 3D). However, expression levels of IL-17A and IL-2 were significantly higher in CD4+ T cells in spleen cell cultures obtained from AE-fgl2-/- mice at 4 months p.i. when compared to AE-WT mice.
The role of FGL2 in the maturation of different subsets of DCs, i.e. CD11b+ and CD11c+ DCs, was investigated. First, the maturation levels in both PECs and spleen cells from infected AE-fgl2-/-mice, AE-WT mice, and non-infected controls were studied. Among CD11b+ DCs, the frequency of the maturation marker CD80 in PECs and spleen cells was higher at 4 months p.i. in AE-fgl2-/- mice than in AE-WT mice (Fig 4A and 4B). The same was found for CD11c+ DCs (Fig 4G and 4H). However, there was no difference in CD86 frequency in both subpopulations of DCs (Fig 4D and 4E). Subsequently, we assessed the same parameters, but upon in vitro cultivation and stimulation with ConA for 48 h. Findings revealed that CD80 frequency, not CD86, in both DC subpopulations from AE-fgl2-/- mice at 4 months p.i. was significantly higher than in those from AE-WT mice (Fig 4C,4F and 4I).
Flow cytometry of spleen cells exposed to VF for 96 h showed that the expression of IL-2 was significantly higher in CD4+ cells from AE-fgl2-/- mice obtained after 4 months p.i. than in the corresponding cell population isolated from AE-WT mice. However, there was no difference in expression of IFN-γ, IL-4 or IL-17A between CD4+ cells from AE-fgl2-/- mice and AE-WT mice (Fig 5A). To further explore the role of VF on Tregs and FGL2 secretion, respectively, spleen cells from AE-WT mice and non-infected WT controls were each cultured in the presence of VF (10 μg/mL). FGL2 levels in the supernatants were determined by ELISA. An identical experiment was performed with CD4+CD25+ Tregs and CD4+ Teffs instead of spleen cells, in the presence of APCs. No differences in FGL2 levels in supernatants of sorted AE-WT Tregs, nor in CD4+ Teffs, could be detected in response to VF, when compared to cultures from non-infected animals (S4 Fig).
For DCs at 96 h after exposure to VF, the CD80 frequency, both in CD11b+ and CD11c+ DCs from AE-fgl2-/- mice at 4 months p.i., was significantly higher than in DCs from AE-WT mice. However, there was no difference in CD86 frequency in both subpopulations of DCs from AE-fgl2-/- mice after exposure to VF, compared to DCs from AE-WT mice (Fig 5B).
To further confirm the role of FGL2 on T cell functions and DC maturation, spleen cells from non-infected WT mice were cultured in the presence or absence of rFGL2 or anti-FGL2-MAbs. Flow cytometry showed that IFN-γ expression was specifically increased in response to VF in the presence of anti-FGL2 MAbs, and IL-17A expression was decreased in the presence of high concentration of rFGL2 (5 μg/mL) and there was no influence on IL-17 expression upon exposure to anti-FGL2-MAb (Fig 6A). The expression of CD62L, a cell adhesion molecule that is abundant on naïve lymphocytes, but not expressed on effector memory T-lymphocytes [29], was found to be increased on CD4+ T cells in the presence of rFGL2 (1 μg/mL), but decreased in the presence of anti-FGL2 antibodies (Fig 6A). However, this was not the case upon analysis of total spleen cell populations (S5 Fig), which indicated that FGL2 might play an important role in down-regulating lymphocyte co-stimulation and effector T cell production.
For DCs, the expression of CD86 on CD11c+ DCs was decreased in the presence of rFGL2, but increased in the presence of anti-FGL2 MAbs. However, expression of CD86 on CD11b+ DCs showed the opposite (Fig 6B).
During infection with E. multilocularis metacestodes, immune tolerance and/or down-regulation of immunity is a marked characteristic that becomes more pronounced at the later stage of chronic disease in both humans [30] and in experimentally infected mice [5]. In the present study, we identified FGL2 as an important mediator of susceptibility to E. multilocularis infection in mice, and demonstrated for the first time that FGL2 a) partially contributes to Treg functions; b) has the capacity to down-regulate the maturation of DCs; c) suppresses Th1- and Th17-type immune responses; d) promotes Th2-biased and Treg immune responses; and finally that e) IL-17A contributes to FGL2-secretion.
Earlier microarray-based data obtained from E. multilocularis-infected liver tissue have shown that the peri-parasitic area is characterized by an increased FGL2 expression [7]. Since immunomodulation is a hallmark of AE, and Treg cells are one of the key immune subsets to mediate this effect, our objective was to study the role of Treg-expressed FGL2 in the outcome of E. multilocularis infection. We originally hypothesized that upregulated FGL2-expression at the host-parasite interface promotes parasite proliferation. Comparative analysis of E. multilocularis growth in fgl2-/- mice versus WT mice at the late stage of infection showed clearly that AE-WT mice exhibited a significantly higher parasite load as compared to AE-fgl2-/- mice. In parallel, the parasite proliferation potential, assessed by determination of the E. multilocularis em14-3-3 gene-expression-level, was significantly higher in AE-WT mice, and respectively lower in AE-fgl2-/- animals. In previous experiments we had already documented that em14-3-3-expression significantly increased upon reduced protective immunity such as encountered in athymic nude mice, or decreased when the parasite growth was hindered as e.g. by albendazole therapy [28,31]. AE-WT-mice also exhibited higher FGL2-levels in the serum as compared to non-infected mice, elevated fgl2 mRNA expression levels in respective PEC and spleen cells, and a concomitantly increased number of Tregs, which represent one of the major Treg producers [9]. Taken together, these results support the hypothesis that Treg-expressed FGL2 contributes to the pathogenesis of E. multilocularis infection. In other infection models, such as viral MHV-3 infection [10], similar results have been demonstrated, while no such effects were found during protozoan (Toxoplasma gondii), bacterial (Yersinia enterocolitica, Listeria monocytogenes, and Mycobacterium tuberculosis), and other types of viral infections such as murine gamma-herpesvirus-68 and Sendai infections [32]. The successful survival of pathogens depends mainly on evading the host immune response by, for example, varying their surface antigens, eliminating their protein coat, and/or modulating the host immune response. Immunosuppression is sometimes directly caused by pathogen-derived metabolic products, and sometimes involves antigenic mimicry. One of the most sophisticated mechanisms of immune evasion is the selective activation of a directly targeted subset of T helper cells. Different pathogens may target different regulative pathways of the host immune response, such as Th2, Tregs or Th17. It is known that the regulation of such pathways is not dependent on single genes or molecules. Rather, a complex immuno-network is usually involved, which requires a sophisticated modulatory process via specific modulatory molecules from E. multilocularis to become effective, some may even act in combination or exert together a synergistic effect. Already know E. multilocularis bioactive molecules include e.g. Alkaline Phosphatase [33], Em492 [34], EmTIP [7], among others; if these are involved in e.g. FGL2-induction needs to be further explored.
How FGL2 influences and/or modulates the outcome of experimental AE is still largely unclear, although some of our study data provide first indications. In experimental AE [5,35,36], the control of metacestode proliferation appears to be predominantly T cell-dependent, as assessed upon use of different immune-compromised mouse models [27,31,37], and confirmed by observations in human AE-patients with immune suppression-associated conditions [38–41]. It is therefore conceivable that the proliferating metacestode itself specifically activates and concurrently modulates the immune response to its own advantage. The metacestode appears to secrete immuno-active metabolites to suppress the protective Th1-type response, and to promote a secondarily reinforced Th2-type response through functional induction and maintenance of Treg cells [2, 25]. Tregs, which over-express a subset of regulatory cytokine genes including those coding for IL-10 and TGF-β, play an important role in promoting immune tolerance in a number of parasitic disease models [42]. In AE, the expression of both cytokines appears up-regulated in mice as shown by different studies [39], and other studies strongly suggested that they are also up-regulated in human AE-patients [43].
Various molecular and cellular mechanisms have been proposed to explain how Tregs suppress immune responses. These include cell-to-cell contact-dependent suppression, cytotoxicity, and immunoregulatory cytokine secretion such as IL-10 and TGF-β [44]. However, the importance of these cytokines remains controversial, as several reports have demonstrated that antibodies against IL-10 and TGF-β failed to block Treg suppressive function, and Tregs from TGF-β–deficient mice retained normal suppressive activity in vitro [44]. In addition, the ambiguous role of TGF-β, which is both, a strong inducer of immune tolerance and an activator of the pro-inflammatory IL-17 cytokine system, remains puzzling [45,46].
Undoubtedly, FGL2 represents an alternative candidate that could, at least partially, support regulatory, and thus immunosuppressive, functions of Tregs. In Tregs of WT BALB/c mice, FGL2 encoding mRNA is constitutively expressed at a high level, and expression even increased after MHV-3 infection, and adoptive transfer of WT Tregs into MHV-3 resistant fgl2-/- mice suggested that FGL2 might be an important Treg effector molecule [10]. Previous studies on FGL2 had shown that rFGL2 suppressed T cell proliferation induced by anti-CD3/28 MAbs and ConA [47,48]. In this study, we now demonstrated that rFGL2 suppressed CD4+ effector T cell proliferation in response to E. multilocularis antigenic metabolites present in the metacestode VF. We also showed that FGL2 inhibited the maturation of DCs, suppressed Th1 and Th17 immune responses, and FGL2 polarized an allogeneic immune response towards a Th2-oriented cytokine profile, both in vivo and in vitro. Conversely, in fgl2-/-mice, Th1 cytokine levels and activity of DCs and T cells were all increased when compared to WT-animals, and FGL2-serum-levels correlated with IL-4 expression in WT-mice before and after E. multilocularis infection. The development of a Th2-oriented immune response in WT-mice during the course of E. multilocularis infection corroborated the generally known effect of FGL2 to promote a Th2 cytokine production, with a concomitant inhibition of Th1- and Th17-oriented immunity [47]. Furthermore, increased serum levels of IL-17A correlated with high FGL2 serum levels, suggesting for the first time that IL-17A could contribute to FGL2-secretion. This was confirmed in vitro by the finding that recombinant IL-17A promoted the production of FGL2 in spleen cells.
We also investigated whether FGL2 would affect the functional activities of Treg cells, by directly assessing the effects of rFGL2 and of an anti-FGL2-MAb on Treg activities in vitro. The presence of recombinant FGL2 promoted Treg function, while addition of anti-FGL2-MAb completely abrogated Treg activity. Further evidence for the role of FGL2 as a molecule that promotes Treg function, and thus immunosuppression, was given by the observation that fgl2-/- mice exhibited both decreased Treg numbers and impaired Treg function. The mechanism by which FGL2 mediates its immunosuppressive activity is currently under investigation. Previous studies have indicated that FGL2 binds to the inhibitory FcγRIIB receptor (CD32) expressed primarily on APCs. Both CD80 and CD86 ligands are found on APCs and are known to provide efficient costimulation, however, in our experiments, distinct functions may be attributed to CD80 and CD86. The differential functions of these molecules have already been the subject of considerable studies, with most data suggesting that the two ligands share substantially overlapping functions. There is, however, also an increasing evidence that supports the view that CD80 may be a more effective ligand for CTLA-4 than CD86 [49]. This FGL2-FcγRIIB interaction through CD80 was shown to induce apoptosis in B cells, and to inhibit DC-maturation [11]. In E. multilocularis infection, several immune subsets may express the FcγRIIB receptor, such as macrophages (including the ‘epithelioid cells’ that line the ‘immuno-modulating’ laminated layer), and also the numerous CD8+ T cells present in the periparasitic infiltrate [50]; CD8+ T cells have actually been shown to express the FcγRIIB receptor in a murine model of Trypanosoma cruzi [51]. Taken our recent data on the course of cytokine expression by the periparasitic immune infiltrate in E. multilocularis infection [4] combined with the results from this study we propose that in E. multilocularis infected mice the following events occur: (a) TNF-α, IFN-γ and IL-17A are released by the host at early stage of infection; (b) these cytokines, and especially IFN-γ as demonstrated previously [32] but also IL-17A as shown in the present study, contribute to FGL2-secretion by Tregs and other cells; and (c) secreted FGL2 can bind to FcγRIIB receptor, leads to a late-stage immune suppression by down-regulating the maturation of DCs, decreases co-stimulation of effector T cells, suppresses Th1 and Th17 immune responses, and accelerates Th2 immune responses. Some parasite's specific molecules may be involved in the modulation of FGL2, as discussed earlier. Overall, this will lead to a periparasitic immune suppressed (anergic) status that favors the continuous “tumor-like” progression of the parasite (Fig 7). Direct inhibition of macrophage and/or mast cell functions could also be induced by this interaction [24].
These findings on the role of FGL2 in E. multilocularis infection now open the door for more applied approaches. For instance, the experimental treatment of E. multilocularis infected mice with anti-FGL2-MAb could provide a means of converting the immunological anergy during chronic disease into a more pro-active, Th1-oriented immunity, with potentially fatal consequences for the metacestode. This is already under investigation. Furthermore, our findings may provide a rationale for studying FGL2 as a target for an immunomodulatory treatment option in patients with progressive AE. In addition, FGL2 may be proposed also as a serum marker indicative for progression of E. multilocularis infection, may be useful to assess the clinical status of AE-patients and the course and outcome and/or parasite activity in human AE.
|
10.1371/journal.pntd.0001927 | Appraisal of a Leishmania major Strain Stably Expressing mCherry Fluorescent Protein for Both In Vitro and In Vivo Studies of Potential Drugs and Vaccine against Cutaneous Leishmaniasis | Leishmania major cutaneous leishmaniasis is an infectious zoonotic disease. It is produced by a digenetic parasite, which resides in the phagolysosomal compartment of different mammalian macrophage populations. There is an urgent need to develop new therapies (drugs) against this neglected disease that hits developing countries. The main goal of this work is to establish an easier and cheaper tool of choice for real-time monitoring of the establishment and progression of this pathology either in BALB/c mice or in vitro assays. To validate this new technique we vaccinated mice with an attenuated Δhsp70-II strain of Leishmania to assess protection against this disease.
We engineered a transgenic L. major strain expressing the mCherry red-fluorescent protein for real-time monitoring of the parasitic load. This is achieved via measurement of fluorescence emission, allowing a weekly record of the footpads over eight weeks after the inoculation of BALB/c mice.
In vitro results show a linear correlation between the number of parasites and fluorescence emission over a range of four logs. The minimum number of parasites (amastigote isolated from lesion) detected by their fluorescent phenotype was 10,000. The effect of antileishmanial drugs against mCherry+L. major infecting peritoneal macrophages were evaluated by direct assay of fluorescence emission, with IC50 values of 0.12, 0.56 and 9.20 µM for amphotericin B, miltefosine and paromomycin, respectively. An experimental vaccination trial based on the protection conferred by an attenuated Δhsp70-II mutant of Leishmania was used to validate the suitability of this technique in vivo.
A Leishmania major strain expressing mCherry red-fluorescent protein enables the monitoring of parasitic load via measurement of fluorescence emission. This approach allows a simpler, faster, non-invasive and cost-effective technique to assess the clinical progression of the infection after drug or vaccine therapy.
| Leishmaniasis is a parasitic disease that is far from eradication. The lack of an efficacious vaccine and treatment failures are major factors in its intractable worldwide prevalence. A non-invasive imaging technique using genetically engineered parasites that expressed fluorescent proteins could give to researchers a quantitative and visual tool to characterize the parasite burden in experimental infections. In addition, it can be useful for determining the efficacy of candidate vaccines or drugs using High Throughput Screening methods that allow the testing of libraries of compounds in an automated 96-well plate format. Herein, we demonstrate that there is a good correlation between fluorescence emission and the parasite load, thus permitting the use of this output to monitor the progression of the disease. In order to validate this tool we have immunized mice prior the parasite challenge with a red-emitting parasite strain, confirming the scientific suitability of this approach as a valuable alternative model.
| Leishmania major is the main cause of cutaneous leishmaniasis (CL) in the Old World. Parasites are transmitted by Phlebotominae sandflies whilst blood feeding on infected mammalian hosts. CL is widely spread in the developing world, affecting people in 88 countries with 1.5 million new cases reported each year. CL usually produces ulcers on the exposed parts of the body that often leave disfiguring scars, which in turn, can cause serious social prejudice [1].
Conventional in vivo animal models for the study of parasite-host relationships involve large number of animals. These animals are required to be slaughtered at different time points in order to identify both anatomical distribution and parasite numbers in organs and tissues. Furthermore, this approach has some important limitations that must be overcome: i) post-mortem analysis of animals makes it impossible to track the space/time progression of the pathogen within the hosts; ii) spread of the pathogen to unexpected anatomic sites can remain undetected; iii) in order to achieve precise and relevant data, it is necessary to kill large numbers of animals. Recent real-time in-vivo imaging techniques with genetically modified pathogens represent a valuable complementary tool. They can be used for conventional studies of pathogenesis and therapy as long as the modified pathogen retains the virulence of the parental strain. Moreover, this has led to an increased number of reports concerning genetically modified parasites that express bioluminescent and/or fluorescent reporters. This was principally developed for in-vitro infection studies and to monitor diseases in living animals [2], [3]. Bioluminescent pathogens expressing the sea pansy Renilla reniformis luciferase have been used in experimental murine infections of Toxoplasma gondii [4] as well as in the rodent malaria parasite Plasmodium berghei [5]. A recent study has allowed scientists to identify the liver stages of firefly luciferase-expressing parasites in living animals [6].
This approach has also been successfully implemented in trypanosomatids. Lang and co-workers showed that a luciferase expressing L. amazonensis strain was useful for rapid screening of drugs in infected macrophages [7]. Further studies used the same techniques with L. major [8], L. infantum [9] and in in-vivo murine experimental infections [7], [10]. Besides, the use of Trypanosoma brucei expressing R. reniformis luc gene has permitted scientists to find unusual colonizing niches during the progression of African trypanosomiasis [11].
Fluorescent imaging offers several benefits: i) unlike light-emitting proteins, fluorescent reporters do not require specific substrates: ii) the fluorescence emitted is very stable over time and iii) this approach is useful when studying tissue harvested from infected animals since parasites can be individually identified [12].
The first transgenic Leishmania species expressing the green fluorescent protein (GFP) was reported by Beverley's group [13]. Episomally transfected Leishmania spp. with GFP or enhanced GFP (EGFP) have enhanced High Throughput Screening (HTS) methods in free-living promastigotes [14]–[17] and amastigotes [18]–[22]. However, only recently, the stable transfection of the EGFP reporter has been found suitable for both in vitro and in vivo infection studies [21]–[24]. Although native GFP produces significant fluorescence and is extremely stable, the excitation maximum is close to the ultraviolet range, which can result in damaging living cells.
Red-fluorescence labelled parasites have been used to determine the early stages of CL pathogeny at the infection site (revised by Millington and co-workers [25]). By combining a L. major Red Fluorescent Protein 1 (RFP)-expressing strain and dynamic intravital microscopy, the site of sandfly bites has been identified in vivo in a mouse model. The study reveals an essential role for both neutrophils and dendritic cells that converge at localized sites of acute inflammation in the skin following pathogen deposition [26], [27]. Using mCherry-L. infantum chagasi – responsible of visceral leishmaniasis in the New World – researchers have been able to report the recruitment of neutrophils and their role in non-ulcerative forms of leishmaniasis [28]. In addition, to study the mechanism regulating dentritic cell recruitment and activation in susceptible BALB/c [29] and resistant C57BL/6 mice [30] DsRed labelled parasites were used
Fluorescent parasites have been used to explain some aspects of Leishmania biology. L. donovani lines stably expressing either EGFP or RFP have been used to identify hybrid parasites produced during the early development of the sandfly [31]. In addition, a L. major strain, which episomally expressed the DsRed protein, was used for quantifying the infectious dosage transmitted by a sandfly bite [32].
Based on the improved photostability as well as suitability for intravital imaging, mCherry was considered the best choice for our studies in comparison to other red fluorescent proteins [33]. mCherry is a protein derived from the coral Discosoma striata RFP. It has a maximum emission peak at 610 nm with a 587 nm excitation wavelength. Despite the fact that it is 50% less bright than EGFP, it is more photostable and it has higher tissue penetration [12]. Because of this, it is the best-suited choice in applications of single-molecule fluorescence or multicolour fluorescent imaging [34]. In this report we describe the use of a stably mCherry-transfected L. major strain as a valuable tool to both in vitro assays for drug screening and in vivo pre-clinical vaccine studies in real-time.
The animal research described in this manuscript complied with Spanish (Ley 32/2007) and European Union Legislation (2010/63/UE). The used protocols were approved by the Animal Care Committee of the Centro de Biología Molecular and Universidad Autónoma de Madrid (Spain).
Female BALB/c mice (6–8 week old) were purchased from Harlan Interfauna Iberica S.A. (Barcelona, Spain) and maintained in specific-pathogen-free facilities for this study.
L. major LV39c5 (RHO/SU/59/P) strain was used for generating mCherry transgenic promastigotes. Parasites were cultured at 26°C in M199 supplemented with 25 mM HEPES pH 7.2, 0.1 mM adenine, 0.0005% (w/v) hemin, 2 µg/ml biopterin, 0.0001% (w/v) biotin, 10% (v/v) heat-inactivated foetal calf serum (FCS) and antibiotic cocktail (50 U/ml penicillin, 50 µg/ml streptomycin). Attenuated Δhsp70-II (Δhsp70-II::NEO/Δhsp70-II::HYG), used as candidate vaccine [35], is a null mutant for the hsp70-type-II gene, generated by targeted deletion in L. infantum (MCAN/ES/96/BCN150) strain [36]. Δhsp70-II promastigotes were grown in RPMI 1640 (Sigma-Aldrich) culture medium supplemented with 10% (v/v) FCS, 50 U/ml penicillin and 50 µg/ml streptomycin.
The 711-bp mCherry coding region was amplified by PCR from pRSETb-mCherry vector, a kindly gift from Dr Roger Y. Tsien – Departments of Pharmacology and Chemistry & Biochemistry, UCSD (USA) – [37] with the primers RBF634 and RBF600 (Table 1). PCR product was cut with appropriate restriction enzymes and ligated into the BglII and NotI sites of the pLEXSY-hyg2 expression vector (Jena Bioscience GmbH, Germany). Parasites expressing mCherry Open Reading Frame (ORF) were obtained by transfection of L. major with the large SwaI targeting fragment derived from pLEXSY-mCherry by electroporation and subsequent plating on semisolid media containing 200 µg/ml hygromycin B (Sigma-Aldrich) as previously described [38]. Correct integration of mCherry ORF into the 18S rRNA locus of the resulting transgenic clones (mCherry+L. major) was confirmed by Southern blot and PCR amplification analyses, using the primers of Table 1. The fluorescent of stable-transfected mCherry clones was confirmed by both flow cytometry (BD FACSCantoII) and confocal microscopy (Nikon Eclipse TE2000E).
Starch-elicited peritoneal macrophages were recovered from BALB/c mice and then 5×104 cells were plated on black 96 wells plates with clear bottom. Macrophages were infected at a ratio of five metacyclic promastigotes per macrophage. Metacyclic mCherry+L. major promastigotes were isolated from stationary cultures (4–5 days old) by Ficoll gradient centrifugation [39]. Briefly, 2 ml of parasite suspension in M199 containing approximately 7×107 stationary-phase promastigotes were layered onto a discontinuous density gradient in a 15 ml conical tube consisting of 2 ml of 20% (w/v) Ficoll stock solution made in distilled water and 2 ml of 10% (w/v) Ficoll diluted in M199 medium.
Metacyclic parasites were opsonized with 4% (v/v) C5− mouse deficient serum (The Jackson Laboratory, USA) at 37°C for 30 min and resuspended in RPMI containing 10% (v/v) FCS [40]. The infection was synchronized by centrifugation (330×g, 3 min at 4°C) and infected macrophages were incubated at 37°C in a humidified 5% CO2 atmosphere [41]. Cells were washed extensively with phosphate buffer saline (PBS) to remove the free parasites and overlaid with fresh medium, which was replaced daily thereafter. After one day incubation, to allow differentiation into amastigotes, drugs (miltefosine, amphotericin B and paromomycin) were added to the appropriate wells in a threefold dilution series in RPMI (Sigma-Aldrich) with 10% (v/v) FCS and cells were further incubated at 37°C for a further incubation of 72 h. Plates were read in a fluorescence microplate reader (Synergy HT; BioTek) (λex = 587 nm; λem = 610 nm).
Metacyclic promastigotes were isolated from stationary cultures (4–5 days old) by negative selection with peanut agglutinin for mouse infections. Briefly, promastigotes were resuspended in PBS at 108 cells/ml, and peanut agglutinin (Vector laboratories) was added at 50 µg/ml; the sample was incubated for 25 min at room temperature. After centrifugation at 200×g for 10 min, the supernatant contained the non-agglutinated metacyclic promastigotes [42].
The virulence of L. major parasites was maintained by passage in BALB/c mice by injecting hind footpads with 106 stationary-phase parasites. After 6–8 weeks, animals were euthanized and popliteal lymph nodes were dissected, mechanically dissociated, homogenized and filtered. L. major amastigotes were isolated from murine lymph nodes by passing the tissue through a wire mesh followed by disrupting the cells sequentially through 25G1/2 and 27G1/2 needles, and polycarbonate membrane filters with pore size of 8, 5 and 3 µm (Isopore, Millipore) [43]. Isolated amastigotes were transformed to promastigote forms by culturing at 26°C in Schneider's medium (Gibco, BRL, Grand Island, NY, USA) supplemented with 20% (v/v) FCS, 100 U/ml penicillin and 100 µg/ml streptomycin. For infections, amastigote-derived promastigotes with less than five passages in vitro were used.
BALB/c mice were injected with several inocula (104, 105 and 106 promastigotes/mouse) during the setting up of the model. For protection studies mice were vaccinated intravenously (tail-vein injection) with the Δhsp70-II mutant strain (2×107 promastigotes/mouse) [35], or injection of PBS (control group), and four weeks post-vaccination, were infected with 2×105 mCherry+L. major metacyclic promastigotes. The infections were performed by injection of parasites in 50 µl PBS in the right hind footpads. The growth of the lesion was monitored by fluorescence emission detection (see below). The contralateral footpad of each animal represented the negative control value. Footpad swelling was measured using a Vernier calliper and data were represented as the increment of the lesion size respect to the not infected footpad.
Fluorescence emission was measured using an intensified charged coupled device camera of the In Vivo Imaging System (IVIS 100, Xenogen). Wild type- and mCherry+L. major-infected animals were lightly anesthetized with 2.5–3.5% isoflurane and then reduced to 1.5–2.0%. Anesthetized animals were placed in the camera chamber, and the fluorescence signal was acquired for 3 s. Fluorescence determinations, recorded by the IVIS 100 system, were expressed as a pseudocolour on a gray background, with red colour denoting the highest intensity and blue the lowest. To quantify fluorescence, a region of interest was outlined and analyzed by using the Living Image Software Package (version 2.11, Xenogen).
The total number of living parasites invading the target organs (popliteal lymph node draining the injected site) was calculated from single-cell suspensions that were obtained by homogenization of the tissue through a wire mesh. The cells were washed and cultured in Schneider's medium containing 20% (v/v) heat-inactivated FCS, 100 U/ml penicillin and 100 µg/ml streptomycin. The cell suspensions were serially diluted and dispensed into 96-well plates. The plates were incubated for 10 days and then each well was examined and classified as positive or negative according to whether or not viable promastigotes were present. The number of parasites was calculated as follows: Limit Dilution Assay Units (LDAU) = (geometric mean of titer from quadruplicate cultures)×(reciprocal fraction of the homogenized organ added to the first well). The titer was the reciprocal of the last dilution in which parasites were observed [44].
Aimed to create a L. major fluorescent strain we electroporated wild-type promastigotes with the lineal 5874 bp SwaI-SwaI fragment containing the ORF encoding mCherry as well as the hyg selection marker of the pLEXSY-mCherry plasmid. After selection on semisolid plates containing 100 µg/ml hygromycin B, individual colonies were seeded in M199 liquid medium supplemented with 10% FCS and hygromycin B. Genomic DNA isolated from these cultures was used to confirm the correct integration of the target sequence into the 18S rRNA locus of L. major genome. Figure 1A shows a schematic representation of the planned integration. Genomic DNA from wild-type strain and two hygromycin B resistant clones were digested with NdeI and hybridized with a labelled external probe (EP). As shown in the Southern analysis of Fig. 1B, wild-type DNA digested with NdeI yielded an 8.4-kb hybridization band, whereas in the two-hygromycin B resistant clones an additional 3.8-kb hybridization band was observed; this band is generated by the integration event (Fig. 1A) corresponding to the expected size. Further confirmation of the correct planned replacements was confirmed by PCR (Fig. 1C) using the set of primers depicted in Fig. 1A.
The mCherry expression in stable-transfected L. major promastigotes (mCherry+L. major) was monitored by flow cytometry. Cell populations of mCherry+L. major strain and a parasite line containing the pLEXSY-mCherry episome emitted strong red fluorescence when they were excited at wavelength of 587 nm (Fig. 2A). Clones with integrated mCherry gene had an average 10-fold higher fluorescence than the ones expressing the gene episomally. This is an expected result given that the mCherry gene was integrated under the control of rRNA promoter, which is known to present high-level transcription rates. As shown in Fig. 2B, both strains (episomal and integrative) were strongly more fluorescent than untransfected parasites.
In order to establish the correlation between parasite number and fluorescence intensity, different number of procyclic and metacyclic promastigotes as well as freshly isolated amastigotes from infected animals were placed in 96-well plates and their fluorescence intensity was measured spectrofluorometrically. A clear correlation between fluorescence intensity and the number of the three parasitic forms was observed (Fig. 2C).
The stability of mCherry expression was monitored over a period of 6 months after transfection and no change was observed in fluorescence intensity during this period, even in the absence of hygromycin B.
Once the infectivity of the mCherry+L. major parasites was recovered through mouse infections, the amastigotes obtained from cutaneous lesions were differentiated back into promastigotes. These cells were grown up to stationary phase and used to infect freshly isolated BALB/c peritoneal macrophages at a 5∶1 multiplicity in 24-well plates. Figure 3 shows fluorescence images of either promastigotes or amastigotes internalized in macrophages. A strong red fluorescence emission from free-living mCherry+L. major promastigotes was observed by confocal microscopy (Fig. 3B). Similarly, round-shaped red fluorescent emitting amastigotes were observed inside parasitophorous vacuoles in the cytoplasm of the infected macrophages (Fig. 3F). The course of the in vitro infection was followed over a period of 48 h by measuring the absolute fluorescence of the infection and the percentage of infected macrophages. These experiments were carried out in parallel with others using the classical Giemsa staining to determine parasite load in vitro (data not shown). No differences between both methods were accounted thus pointing to the suitability of fluorescence analyses to assess the infectivity of mCherry+L. major strain on mouse macrophages. A major application of a fluorescent Leishmania model would be its usefulness to perform HTS of potential leishmanicidal compounds in vitro. To assess the suitability of our mCherry+L. major for this goal, current drugs used in the treatment of human leishmaniasis (miltefosine, paromomycin and amphotericin B) were assayed in Leishmania-macrophage infections at different concentrations over a 72 h-span. Absolute fluorescence emitted by mCherry+L. major infected macrophages was plotted vs. drug concentrations, obtaining the dose-response curves of Fig. 4. Nonlinear regression analysis of the curves, fitted by the SigmaPlot statistic package, reached IC50 values of 0.12±0.03 µM for amphotericin B, 0.57±0.12 µM for miltefosine and 9.20±3.59 µM for paromomycin. In all the cases the difference in fluorescence emission corresponded to a difference in the percentage of infected cells, also observed microscopically. These findings clearly showed that the mCherry+L. major strain is a useful tool for in vitro drug screening.
In order to determine whether mCherry+L. major parasites could be detected in vivo using whole-body imaging, 104, 105 and 106 metacyclic forms of the fluorescent-transgenic strain were injected subcutaneously into the hind footpads of six BALB/c mice per group. Lesion progression monitored both by direct measuring of fluorescence emission by mCherry+L. major amastigotes (recorded in an IVIS 100) and by the development of hind-limb lesions assessed by measuring the thickness of the footpads with a Vernier calliper. Animals were examined every seven days for a total of eight weeks (except the group infected with 106 parasites that were sacrificed at 6th week post-infection because the appearance of ulcerations in the footpads).
Figure 5 shows the fluorescence intensity recorded weekly from the footpads of representative mice of each inoculum group (104, 105 and 106 metacyclic promastigotes per mouse). The fluorescent signal (estimated as average radiance: p/s/cm2/sr) was plotted against the infection time of each inoculum (Fig. 6A). Fluorescent signal was detected after the first week post-infection in mice infected with 106 metacyclic parasites (radiance = 0.26×108 p/s/cm2/sr), reaching a radiance of 5.0×108 p/s/cm2/sr five weeks later. In the group of mice injected with 105 metacyclic parasites, the fluorescence was detected the third week after inoculation (radiance = 0.8×108 p/s/cm2/sr), reaching similar intensity than the mice group injected with 106 parasites at the 8th week of inoculation. Finally, the fluorescence signal of mice injected with 104 metacyclic mCherry+L. major parasites was not detectable until the 5th week (radiance = 0.35×108 p/s/cm2/sr), reaching the maximum intensity (radiance = 1.14×108 p/s/cm2/sr) at the end of the experiment.
The success of infection defined as the lesion onset and its development in the inoculated footpads over the time, was observed in every mice. Although there was a good correlation to lesion size, fluorescence was more sensitive to evaluate the progression of infection. Figure 6B shows that lesion emergence was dependent on the size of pathogen inoculum and it was hardly measurable during the first weeks after infection. Lesion size in mm was 0.34, 0.94 and 0.57 measured at the third, fifth and seventh week, respectively corresponding to 106, 105 and 104 metacyclic promastigotes per inoculum. It is remarkable that a weak but measurable fluorescence signal from infected hind limbs was detectable two weeks prior to visible and measurable injury took place in all dose groups.
As expected and since BALB/c mice have a predisposition to develop an anti-inflammatory Th2 response, the lesions appearing after infection setup were non-healing without treatment. A correlation analysis of the fluorescence emitted by the lesions toward the end of the 8-week period from mice infected with 104 and 105 metacyclic mCherry+L. major parasites (Fig. 6C) shows significant differences between both groups (P<0.001) using unpaired t-Student test. These differences were also found when the lesion thickness of the footpads was compared using the traditional calliper-based method (Fig. 6D). There was a clear correlation between both parameters in both dosing groups with an estimated Pearson coefficient of 0.94. At the end of the experiments, animals were sacrificed and the popliteal lymph nodes draining the lesions were dissected under sterile conditions. The parasite load of these organs was determined by the limit dilution method. Figure 6E shows the number of promastigote-transforming amastigotes estimated in the animals of both dosing groups, showing significant differences (P<0.001) and correlating highly with both size lesion and fluorescence (Pearson coefficient = 0.79).
The suitability of this in vivo approach was assessed for the evaluation of an experimental vaccination protocol against CL that had been previously shown to be effective on a L. major–BALB/c infection model [35]. In previous studies, it was established that intravenous inoculation with Leishmania promastigotes, lacking both alleles of the hsp70-II gene (Δhsp70-II line), confers a partial protection against L. major infection in mice. For this study, we inoculated a group of six mice with 2×107 promastigotes of Δhsp70-II mutant and four weeks later, mice were challenged with 2×105 metacyclic forms of the mCherry+L. major strain into mouse footpads. In parallel to the vaccinated group, a control group was injected with the same inoculum of mCherry-expressing transgenic parasites.
Red fluorescence emission in the footpad of mice infected with mCherry+L. major parasites was followed over the time in both groups (Fig. 7A). Fluorescence signal was detected in both groups four weeks after challenge; however, fluorescence signal was higher in control group mice than in vaccinated animals. By the end of the 8th week animals were euthanized, the poplyteal lymph nodes dissected, homogenized and the parasite load determined as above. Figure 7B shows an 80% reduction (P<0.001) in the parasite load of popliteal lymph nodes of vaccinated group related to the control group. Therefore, since reproducible results were obtained with both parasite quantification and fluorescence emission methods, we conclude that the murine model of CL established with the mCherry+L. major fluorescent strain might be a suitable system for testing antileishmanial therapies both in vitro and in vivo.
Transgenic parasites expressing reporter proteins are valuable tools to perform robust HTS platforms [45] and to understand the underlying mechanisms of pathogenesis [3]. GFP is one of the most commonly used reporters among fluorescent proteins. Several mutants derived from native GFP have been developed to cover longer wavelengths of the spectrum. Reporter molecules, whose emission peak is in the red spectral range, the same as mCherry, are excellent candidates for these kinds of studies. Furthermore, light absorption by tissues in the red and far-red spectra is reduced and consequently, the penetration is higher [46]. Moreover, mCherry is the best general-purpose red monomer due to its superior photostability compared to mStrawberry and DsRed, which is inadequately folded at 37°C [13].
The integration of the reporter gene into the 18S rRNA locus of L. major represents an efficient and effective strategy to guarantee a stable expression when the parasites need to be grown in the absence of selection drugs for both in vitro screenings and in mice infections [7], [22]–[25], [47], [48].
mCherry fluorescence was detected in the different stages of the L. major life cycle. Lesion-derived amastigotes showed two times less activity than metacyclic promastigotes. In turn, these were three times less fluorescent than logarithmic promastigotes. Similar results were reported in promastigotes of different Leishmania species, in which luciferase expression was much higher than that of amastigotes from animal lesions and experimentally infected macrophages, respectively [7], [47]. However, Mißlitz and co-workers [23] showed that EGFP expression levels were 2–10 times higher in amastigotes than in promastigotes of both L. mexicana and L. major. Although these species were stably transfected by the integration into the 18S rRNA locus; they differed in the downstream region of the reporter gene. Whilst no specific 3′ untranslated region implicated in the stage-specific regulation was included downstream on the luc gene [47]; the intergenic calmodulin A region was configured into the pLEXSY plasmid ([7] and the present work). In a similar way, the cysteine proteinase intergenic region (cpb2.8) was included in the studies conducted by Mißlitz [23]. Intergenic sequences responsible for a high transcription rate in amastigotes should be included in future vectors for regulating the reporter's expression. In this sense, technologies such as RNA sequencing can provide a complete transcriptome that could be used to improve the expression technology in both promastigote and amastigote forms [49], [50].
Assays designed to simplify rapid and large-scale drug screenings are not performed on the clinically relevant parasite stage, but on promastigotes instead. Axenic amastigotes have also been screened by means of HTS platforms [9], [51], [52]. However, expression arrays comparing both axenic amastigotes and those isolated from infected macrophages have shown metabolic differences, impaired intracellular transport and altered response to oxidative stress [53].
The suitability of mCherry+L. major transgenic strain is an important tool for bulk testing of drugs in the intracellular amastigote stage. This was demonstrated further by using three drugs in clinical use against leishmaniasis: amphotericin B, paromomycin and miltefosine.
Most of the drug screening assays attempted to analyze intracellular parasites using GFP-tranfected Leishmania spp. Theses methodologies clearly showed that there was not enough sensitivity to enable a precise and reliable microplate screening. Consequently an in-depth flow cytometric analysis is required [54]. Recently, a novel method for assessing the activity of potential leishmanicidal compounds on intracellular amastigotes through the use of resazurin (a fluorescent dye with emission wavelength in the red spectrum) has advanced to microplate analysis [55]. Unlike the GFP-expressing parasites, mCherry emission is also found in the same spectral range as resazurin. This level of sensitivity was sufficient to detect 104 amastigotes isolated from lesions. This means that mCherry reporter provides several benefits over fluorescent proteins for performing HTS into microplate format.
Other advantages of fluorescent proteins are that they allow a dynamic follow up (kinetic monitoring) of the drug efficiency using a single plate. Drugs must be maintained in the culture medium for a time long enough for them to take effect. On the contrary, multiple plates are required if a specific substrate is added, requiring one for each recorded time interval.
Through our research we want to raise the importance of the source of host cells used for experimental infections when drug-screening assays are carried out. Several differences in the host-parasite interactions have been pinpointed when comparing primary macrophages with immortalized human macrophage-like cell lines [56]. Most of the current multiwell-screening methods involve established-macrophage cell lines since it is quite difficult to scale-up a procedure based on primary macrophages [57], [18], [58], [20], [22], [59]. Accordingly, a well-planned combination of different approaches (promastigote/intracellular; cell line/primary cultures) would help us to identify lead compounds through large-scale drug screening [60], [61].
The manipulation of large numbers of potential drugs not only requires easy-to-use, repeatable and readily quantifiable tests, but also it needs to mimic natural conditions within the host cell. Because of the profound influence of the host's immune response on the treatment of leishmaniasis, new approaches should include the whole immunopathological environment found at the host-parasite interaction site. However, only one alternative approach has been used in order to transfer this immunological concept to HTS systems [43], [61].
The main advantages of mCherry-transfected parasites are automation and miniaturization. As experiments are performed in 96-well plates, reducing costs of reagents, and time of analysis is of great importance. Besides, we can also eliminate tedious steps such as staining or cell lysis. In addition this allows a dynamic follow up as cells remain viable after each reading time interval.
As the stable integration of the gene encoding reporter proteins represents a valuable tool for assessing whole-body imaging in laboratory mice [47], [7], [10], [62]–[64], we decided to use the same mCherry-transfected strain for in vivo applications. Experimental infections with L. major in BALB/c mice footpads resulted in a non-healing and destructive chronic lesion at the site of injection. The mCherry in vivo model developed in this study clearly allowed the fluorescence signal in the first week post-inoculation with 1×106 stationary parasites, a dose used in leishmanial research to induce the rapid development of CL. Similar models in BALB/c mice with EGFP, used 10- and 200-fold parasite doses and the fluorescence signal was visualized afterwards [21], [24]. The lymph node draining the lesion was not detected in this study, probably because the lower inoculum used or because of the shorter time of testing when the animals were killed. Previously, reports detected the fluorescence or luminescence signal emitted by the lymph node a long time after post-infection (2.5–10 months) [7], [21], [23].
In order to evaluate the eligibility of our fluorescent tool for the monitoring of in vivo treatments, we applied this approach by evaluating an experimental vaccine against leishmaniasis that had been previously shown to be effective. Our previous studies showed that a L. infantum strain lacking the hsp70-II gene (Δhsp70-II line) conferred resistance to a subsequent infection with L. major [35], [36]. We found that the progression of the infection was efficiently and effectively observed by recording the mCherry signal through real-time imaging. Vaccination of infected mice for a period of 8 weeks with the vaccine reduced the infection when compared with the control group. Further to this, Mehta and co-workers successfully used a similar vaccination approach to assess the efficiency of a real-time imaging platform using an engineered strain of L. amazonensis expressing the egfp gene [24].
In conclusion, we have developed a valuable fluorescence-emitting L. major transfected strain. This strain allows us: i) actual imaging, which is important when studying tissue harvested from an infected animal because parasites can be individually identified; ii) to easily develop new, fast and efficient platforms for the screening of potential leishmanicide drugs testing thousands of compounds in Leishmania amastigote-infected macrophages; iii) to reproduce the infection in real-time due to the virulence of L. major-transfected strain, which in turn increases the sensitivity of detection especially at the earlier phases of the process. Furthermore, this avoids the unnecessary slaughter of large amounts of animals at different time-points owing to direct imaging and fluorescence testing, which can be performed without traumatic handling to the animals.
|
10.1371/journal.pntd.0001647 | Lower Richness of Small Wild Mammal Species and Chagas Disease Risk | A new epidemiological scenario involving the oral transmission of Chagas disease, mainly in the Amazon basin, requires innovative control measures. Geospatial analyses of the Trypanosoma cruzi transmission cycle in the wild mammals have been scarce. We applied interpolation and map algebra methods to evaluate mammalian fauna variables related to small wild mammals and the T. cruzi infection pattern in dogs to identify hotspot areas of transmission. We also evaluated the use of dogs as sentinels of epidemiological risk of Chagas disease. Dogs (n = 649) were examined by two parasitological and three distinct serological assays. kDNA amplification was performed in patent infections, although the infection was mainly sub-patent in dogs. The distribution of T. cruzi infection in dogs was not homogeneous, ranging from 11–89% in different localities. The interpolation method and map algebra were employed to test the associations between the lower richness in mammal species and the risk of exposure of dogs to T. cruzi infection. Geospatial analysis indicated that the reduction of the mammal fauna (richness and abundance) was associated with higher parasitemia in small wild mammals and higher exposure of dogs to infection. A Generalized Linear Model (GLM) demonstrated that species richness and positive hemocultures in wild mammals were associated with T. cruzi infection in dogs. Domestic canine infection rates differed significantly between areas with and without Chagas disease outbreaks (Chi-squared test). Geospatial analysis by interpolation and map algebra methods proved to be a powerful tool in the evaluation of areas of T. cruzi transmission. Dog infection was shown to not only be an efficient indicator of reduction of wild mammalian fauna richness but to also act as a signal for the presence of small wild mammals with high parasitemia. The lower richness of small mammal species is discussed as a risk factor for the re-emergence of Chagas disease.
| The classical methodology of mapping works with discrete units and sharp boundaries does not consider gradient transition areas. Spatial analysis by the interpolation method, followed by map algebra, is able to model the spatial distribution of biological phenomena and their distribution and eventual association with other parameters or variables, with a focus on enhancing the decision power of responsible authorities. Acute Chagas Disease outbreaks are increasing in the Amazon Basin as result of oral transmission. This scenario requires a new approach to identify hotspot transmission areas and implement control measures. We applied a geospatial approach using interpolation and map algebra methods to evaluate mammalian fauna variables related to these outbreaks. We constructed maps with mammalian fauna variables including the infection rates by Trypanosoma cruzi, in dogs and small wild mammals. The results obtained by visual examination of the maps were validated by statistical analysis. We observed that high prevalence of T. cruzi infection in dogs and small wild mammals was associated with mammal lower richness. Monitoring of T. cruzi infection in dogs may be a valuable tool for detecting the fauna lower richness of small wild mammals and elucidating the transmission cycle of T. cruzi in the wild.
| The causative agent of Chagas disease, Trypanosoma cruzi (Chagas, 1909), is a multi-host parasite capable of infecting almost all tissues of more than one hundred mammal species [1]. Dozens of species of insects from the Triatominae subfamily can act as its vector. Except for the epimastigote form, all other T. cruzi evolutive forms can infect mammals by oral and congenital pathways as well as by contamination of the mucosae and skin abrasions by infected triatomine feces. The biological plasticity of T. cruzi results in transmission cycles that are characterized by being multivariate and complex on unique temporal and spatial scales [1], [2].
Classically, Chagas disease was characterized as prevalent in rural populations, where houses were heavily infested by domiciliated triatomine species, mainly Triatoma infestans (Klug, 1834). The campaigns launched by the “Cone Sul” Intergovernmental Commission to eliminate the domiciliary vectors succeeded in that Brazil and other countries in South America are currently considered free from domestic transmission of Chagas via Triatoma infestans [3]. However, extradomiciliary vectorial transmission, domiciliary or peridomestic transmission by non-domiciliated vectors and oral transmission by ingestion of food contaminated by feces from infected insects (the principal method of current transmission), pose new challenges. In fact, mainly in the northern part of the Brazil, the number of Chagas disease outbreaks due to the ingestion of food contaminated by infected triatomine feces are increasing [4]–[6]. This is currently considered a new epidemiological scenario, demanding systematic surveillance methods that consider all components of the transmission cycle as well as the landscape and ambient conditions in which transmission is occurring.
In several reports, the maintenance of biodiversity has been pointed to as a strong buffering system and regulator in the dispersal of parasites; this has been named the “Dilution Effect”. Such a dilution effect has already been demonstrated to be of importance in the transmission of West Nile encephalitis, Hantavirus, Lyme disease and Schistosomiasis [7], [8]. Despite the demonstration of this effect, studies on the impact of biodiversity variation on the T. cruzi transmission cycle in the wild mammals using a Geographical Information System (GIS) have been scarce up to now [9]–[12]. The destruction of an ecosystem imposes important area and food restrictions onto wild mammal populations and may promote their greater contact with humans. The consequence of this process is the increased opportunity for contact among humans, domestic animals and wildlife [2]. In this scenario, the transmission of T. cruzi may be increased due to the following: (i) positive selection of generalist species with high transmissibility competence such as Didelphids and some caviomorph rodents that undoubtedly adapt and survive in degraded habitats, (ii) the consequent amplification of the parasite's transmission cycle due to higher abundance of competent reservoir species and (iii) the increased prevalence of infected bugs. In addition, the scarcity of food sources for triatomines (i.e., loss of wildlife due to destruction of the environment) led them to invade human dwellings and annexes [13]. Also, the quantitative and spatial patterns of the landscape and artificial lighting in human dwellings play a fundamental role in domiciliary invasion [14]. Human residences acting as light-traps for insects has significant epidemiological importance, as species with high rates of T. cruzi infection are drawn to human dwellings [15]. In this scenario, peridomestic mammals are more frequently exposed. Thus, their infection usually precedes the human infection. Hence, dogs have been proposed as being suitable sentinel hosts for T. cruzi transmission in areas at risk for human infection [16], [17].
Dogs can be important reservoirs of this parasite. They display both a high prevalence of infection and high parasitemia as evidenced in Panama [17], Argentina [18], Venezuela [19], Mexico [20], and the United States [21]. In contrast, in Brazil, a high serum prevalence in dogs has also been described in several areas, but the importance of dogs as a reservoir species has been described as negligible because no high patent parasitemia has been observed in these animals [16], [22].
Geospatial analysis based on the fundamental concepts of landscape epidemiology [23] is a powerful tool in the study of the association between landscape- and vector-borne diseases such as Chagas disease, Schistosomiasis and American Visceral Leishmaniasis [9], [10], [24]. Geospatial analysis allows for the identification of disease risk areas and disease interactions with the environment [10], [24].
The classical methodology of mapping works with discrete units and sharp boundaries, and does not consider transition areas. Nevertheless, environmental and biological phenomena are typically continuous and exhibit a gradual transition from one characteristic to another. Unlike the classical methodology, spatial analysis by the interpolation method, followed by map algebra, is able to model the spatial distribution of the continuous biological phenomena, representing the distribution and association of these phenomena in a more realistic way. This modeling can enhance and facilitate decision making [25].
The present paper evaluates and compares T. cruzi infection rates of dogs from three Brazilian biomes, including areas where orally transmitted Chagas disease outbreaks were reported and areas where Chagas disease is endemic. Our objectives were to (i) assess the impacts of lower richness of small wild mammals on the prevalence of T. cruzi infection in dogs, (ii) discuss the role of dogs in the transmission cycle of T. cruzi and their putative role as sentinels and (iii) to assess the interpolation and map algebra method as a tool for the construction of potential Chagas disease risk area maps.
Dogs (n = 649) were sampled in 3 geographic Brazilian biomes: the Amazon, Caatinga and Pantanal, from 5 states and 13 municipalities (28 localities) (Figure 1 and Table 1).
Among these municipalities, samples were taken immediately after an outbreak of Acute Chagas Disease (ACD) from Redenção, Cachoeiro do Arari, Belém, Curralinho and Axixá do Tocantins. The other monitored areas were Abaetetuba, Monte Alegre, Augustinópolis, Esperantina, Jaguaruana, Russas and João Costa, while Corumbá (Midwest Brazil) was used as a control area. The areas included in our study reflect the locations where our laboratory has been developing research and reference services over the past few years.
The states of Piauí and Ceará display similar patterns: a high density of naturally T. cruzi infected Triatominae, which are the main vectors of disease in both regions. Despite that, no new vectorial transmission Chagas disease cases have been observed in the last decade and, as far as we know, only one outbreak has occurred due to oral transmission [2], [26].
The municipalities of Jaguaruana, Redenção and Russas, which are endemic for Chagas disease, are located in the mesoregion of lower Jaguaribe, in the northeastern state of Ceará. In Jaguaruana, the average annual temperature ranges from 23°C to 33°C. The collection area consists of clay and sandy soil plains, which are characterized as Caatinga, and include the typical vegetation of semi-arid areas. Redenção (ACD outbreak in 2006) is located in the Baturité mountain range region. The climate is semi-arid, and the average annual temperature ranges from 24°C to 35°C. The collection area, originally part of a tropical semi-humid forest, is currently characterized by secondary vegetation consisting of small trees (up to 6 m), rocky formations, and remnant patches of semi-humid forest near deforested areas occupied by monoculture plantations or unplanned households (slums). The municipality of Russas is located in the state of Ceará. The climate is semi-arid with average temperatures ranging from 18.8°C to 35.4°C. The vegetation comprise open scrub and savanna, with deciduous thorny forest areas. The municipality of João Costa is located in the southeast of the state of Piauí and is characterized as a megathermic semi-arid region. The average annual temperature ranges from 12°C to 39°C. The vegetation in this area displays the typical Caatinga features and residual semi-deciduous forest patches.
T. cruzi oral transmission in the Brazilian Amazon region has been reported since 1968 [5], [27], although this region was considered an endemic area for many years. Just after 2005, when the prevalence of Chagas cases in other parts of the country decreased and surveillance in the Amazonian region was improved, microepidemics of ACD began to appear regularly and frequently, mainly associated with the consumption of the palm-tree fruit açaí and other foods [4]–[6].
The municipalities of Abatetetuba and Belém (ACD outbreak in 2009) are located in the northeastern mesoregion of the state of Pará. Cachoeiro do Arari (ACD outbreak in 2006) and Curralinho (ACD outbreak in 2009) are located in the mesoregion of Marajó, while Monte Alegre is located in the lower Amazon mesoregion of Pará. The common climate is characterized as tropical humid, with regular rainfall and winds, and temperatures between 27°C and 36°C. The area is known as varzean, which is a freshwater swamp forest. In most of the collection areas, the original native vegetation (Amazonian forest) is being replaced by an extensive açaí fruit monoculture with a few remaining patches of the original vegetation at the river banks. The municipalities of Augustinópolis, Axixá do Tocantins (ACD outbreak in 2009) and Esperantina are located in the northwestern mesoregion of the Tocantins state, almost at the border of Pará. The climate of these cities is tropical subhumid, with maximum temperatures occurring during the dry season that reach 39°C.
This region presents an enzootic cycle of T. cruzi transmission; however, it is not considered an endemic area for Chagas disease, as human cases have never been recorded in the region. This region comprises a large natural environment with a multiplicity of habitats and a wide variety of biodiversity. Farms encompass an area located in the core of a biodiversity corridor in the Pantanal of Mato Grosso do Sul, Brazil.
The capture of small wild and synanthropic mammals was performed as follows: live traps were arranged in linear transects, and the capture points were established with Tomahawk (Tomahawk Live Traps, Tomahawk, WI) and Sherman (H. B. Sherman Traps, Tallahassee, FL) traps. The traps were baited with a mixture of peanut butter, banana, oat and bacon and set at 20-m intervals in all types of vegetation formations and habitats. The trapped animals were taken to a field laboratory ≤2 km from the capture point, where the remaining procedures were performed. The trapped animals were examined for the prevalence and pattern of T. cruzi infection, as previously described [2], [9], [28], [29]. Some data (from the municipalities of João Costa, Cachoeira do Arari, Redenção) used in this meta-analysis comprise both already published studies [2], [9], [22] and some unpublished data (from the municipalities of Jaguaruana, Russas, Abaetetuba, Belém, Monte Alegre, Curralinho, Axixá do Tocantins, Augustinópolis and Esperantina) were collected by our laboratory. The sampling efforts to capture mammals were similar in all 28 localities (820-1.100 traps/night), with 4 or 5 nights of capture each season, and the captures were performed in every season of the year (Table S1 and Figure S1).
The active search for dogs was conducted in the houses neighboring the linear transects where the small wild and synanthropic mammals were captured and in the houses where oral outbreaks of Chagas disease had occurred. Blood samples were collected from 649 dogs living in houses located in twelve municipalities from four Brazilian states. Dog blood was collected in three biomes; collections from the Caatinga (n = 188) and Amazon (n = 422) biomes were compared to collections from the Pantanal biome (n = 39) (Table 1).
Herein, we considered that each dog represents one single event, even when related to the same house. This is due to: (i) dogs are separated individuals, differing each other in age, behavior, activities, etc… This fact is reflected in their different degrees of T. cruzi exposition; (ii) dogs have no pack behavior; and (iii) dogs are not confined in the intradomociliar area and have different and multiple opportunities to be infected during their activities. The interpretation of our results was based on different patterns of infection of the mammals. Fresh blood smears and hemoculture, when positive (especially the first due to its lower sensibility), show a high parasite load in the peripheral blood of the animals, which means a high chance of transmission to the vector, reflecting transmissibility. These tests are very specific but less sensitive – their importance lies in detecting infected animals that may represent a source of infection for the vector. Serological assays indicate infection of the animal. Therefore, serological positive and parasitological negative tests for a given animal demonstrates its infection with a low rate of parasite, this mammal is a host of the parasite, but is not involved in the amplification of parasite populations, i.e., its transmission potential to the vector is low.
Blood was collected from dogs in heparinized vacutainer tubes by puncture of the cephalic vein. To evaluate T. cruzi infection, four tests were conducted. Two of these tests were parasitological assays including (i) the examination of fresh blood smears (microscopic analysis) and (ii) a hemoculture assay, in which 0.2–0.4 mL of blood was cultured in two tubes containing Novy-Mc Neal-Nicole medium [NNN] with a liver infusion tryptose medium [LIT] overlay. When those tests were positive, the parasites were amplified for cryopreservation and DNA extraction for molecular characterization and two serologic diagnostic assays were performed: (iii) the Indirect Immunofluorescence Antibody Test (IFAT) as previously described [30] and (iv) the Enzyme-Linked Immunosorbent Assay (ELISA, Bio-Manguinhos, FIOCRUZ, Rio de Janeiro, RJ, Brazil). Disease diagnosis was based on serology by the ELISA (Cut-off: optical absorbance ≥0.200, mean±3 SD) and IFAT (Cut-off: titer of 1/40). Each microtiter polystyrene plate had 2 positive and 2 negative control sera. Animals were defined as seropositive when samples were reactive in both the IFAT and ELISA. Seropositive animals that displayed negative results in the parasitological assays were considered to have sub-patent infections. To evaluate possible cross-reactions and/or mixed infection by T. cruzi and Leishmania spp., dog sera were also assayed for Leishmania infantum using IFAT and the Rapid Test for Diagnosis of Canine Visceral Leishmaniasis (CVL) (TR DPP®, Bio-Manguinhos, FIOCRUZ, Rio de Janeiro, RJ, Brazil). The IFAT cut-off value adopted for T. cruzi infection was 1/40 when the IFAT result for L. infantum was lower than 1/40 and the DPP results were negative. When infected by L. infantum, dogs were considered positive for T. cruzi infection only when the IFAT titer was 1/80 or higher. For L. infantum infection, the adopted IFAT cut-off value was 1/40 when the infection was also confirmed by DPP and 1/80 when the DPP assay was negative. The interpretation of these tests in assemblage indicates the role played by the tested mammals in the transmission cycle.
DNA was extracted from logarithmic phase cultures and serum samples of dogs with patent parasitemia (positive blood slide smears) in the absence of hemocultivated parasites, using a phenol chloroform protocol [31]. PCR was performed using the primer pair S35 (5′-AAATAATGTACGGGGGAGATGCATGA-3′) and S36 (5′-GGGTTCGATTGGGGTTGGTGT-3′) [31]. Cyclic amplifications were performed with an initial denaturation of five minutes at 94°C, followed by 35 amplification cycles (94°C for 30 seconds, 55°C for 30 seconds, 72°C for 30 seconds) and a final ten-minute elongation step at 72°C. Each 25 µL total reaction volume contained 25 ng total DNA, 10 ρmol of each primer, 0.4 mM dNTPs, 2 mM MgCl2, and 2.5 U Taq polymerase (AmpliTaq®Gold, Applied Biosystems). PCR products were visualized on a 2% agarose gel after ethidium bromide staining. PCR resulted in a 330-bp amplicon for T. cruzi and a 760-bp amplicon and heterogeneous set of fragments ranging in size from 300 to 450 bp for T. rangeli.
The base map was acquired from the IBGE (Brazilian Institute of Geography and Statistics). The coordinates of all biological data were captured using a hand-held GPS (Global Positioning System) receiver (Garmim III GPS, Garmin International, Olathe, KS, USA) and recorded in the WGS 84 Datum (World Geodetic System 1984) geodetic coordinate system.
Maps representing the spatial distribution of T. cruzi infected dogs (response variable) and species richness, abundance and parasitological and serological prevalence of wild small mammals (covariables) were generated using the interpolation method of Inverse Distance Weighted (IDW) with the 12 nearest sampled data points selected. However, for map analysis, only the polygon of the studied municipality, (i.e., only a local analysis) was used. A variable radius was applied specifying the number of nearest input sample points (n = 12) to perform interpolation. After that, we used map algebra to find evidence of spatial correlation of the response variable with each covariable by the use of arithmetic operators (subtraction). The algebraic analysis of maps (spatial variables, response variable and covariables), represented by pixels, results a new map where the values in each geographical position was the result of subtraction (in our case) of the values of the variables associated with the geographical position.
The term “map algebra” was established by Dana Tomlin in the early 1980s [32] with the development of the “Map Analysis Package GIS”. Map algebra provides tools to perform spatial analysis operations and is based on a matrix algebra, which refers to the algebraic manipulation of matrices (as maps in raster data structures). Spatial data were analyzed in a GIS platform using ArcGis 9.3 software (Environmental Systems Research Institute, Redlands, CA, USA).
For the analysis of the proposed hypothesis (small wild mammal lower richness is associated with the increase of prevalence of T. cruzi infected dogs), we used the GLMs (Generalized Linear Models) with a Poisson link function [33]. For the response variable we used the infection of the dogs (based on serological assays – IFAT and ELISA, as described above). The following covariables were included: (1) Species richness of small wild mammals collected (DS): The richness was calculated as being the number of species captured in each area; (2) Abundance of small wild mammals collected (NM): The abundance for each localities was based on: n = total number of mammals captured. In the present model, aiming to evaluate the influence of both parameters (normally associated in ecological studies), these two covariables (DS*NM) were considered together and estimated by the “manual” selection method; (3) Prevalence of small wild mammals with positive T. cruzi parasitological assays (THC): That included (i) the finding of flagellates with typical T. cruzi morphology in fresh blood examination and (ii) the isolation and characterization of T. cruzi from blood in axenic medium – hemoculture; (4) Prevalence of small wild mammals with positive T. cruzi serological assay: based on the detection of specific anti-T. cruzi antibodies in the IFAT. The criterion of comparison between the models was based on the Akaike Information Criterion (AIC) and residual deviance [34], [35] to determine which model fits best considering the level of significance (p<0.05). For the analysis of normality, the Shapiro-Wilk normality test was performed. Each model is specified as a combination of covariables that can influence the probabilities of dogs becoming infected. The comparison between dogs from Chagas disease-outbreak and non-outbreak areas were calculated using 2×2 contingency tables along with a Chi-squared test. Each dog was considered as one independent event, even when living in the same house. Both analyses were performed using the software R (Version 2.11.1) [36].
All wild animal manipulation procedures were performed in accordance with the COBEA (Brazilian College of Animal Experimentation) following the guidelines of the Animal Ethics Committee (CEUA) protocol of FIOCRUZ (Oswaldo Cruz Institute Foundation), Ministry of Health, Brazil. All field workers who manipulated animals directly were adequately dressed with protective equipment, following protocols previously approved by the CEUA-FIOCRUZ Committees of Biosafety and of Bioethics (licenses: P0007-99; P0179-03; L0015-07; P0292/06). The wild animal captures were licensed by the Brazilian Institute of Environment and Renewable Natural Resources (IBAMA) licenses 068-2005 and 225-2006 (IBAMA/CGFAU/LIC). In all cases, consent from the dog owners was obtained. In addition, the owners also helped handle the animals during sampling to avoid incidents. A canine standard questionnaire was applied. For each dog, the questionnaire included name, sex, age, size, color and main phenotypic features, birthplace, age at which the pet entered the house, the dog's main function and movement areas.
Trypanosoma cruzi infection in dogs was sub-patent in the majority of the cases. Only five dogs from Monte Alegre displayed trypomastigote forms in fresh blood preparations (Table 1). These dogs with patent parasitemia displayed severe clinical symptoms (fever, pale mucous membranes, generalized edema, rigid abdomen and splenomegaly), and the disease was fatal for two of them. Hemocultures performed in the three dogs surviving two months after the first blood collection were negative; these dogs produced Chagas negative hemocultures 3 months after the first blood collection. Positive hemocultures were detected only in two dogs, one from Abaetetuba and one from Curralinho. From one of these two dogs, it was possible to isolate the T. cruzi on two occasions after an interval of seven months (Abaetetuba). Molecular characterization was performed on these isolates and on five serum samples of dogs from Monte Alegre. T. cruzi k-DNA was amplified from these six dogs, from both serum and culture samples, with the exception of one dog from Curralinho with Trypanosoma rangeli (Figure 2 and Table 1).
Overall, the distribution of T. cruzi infection in dogs was not homogeneous among houses, localities, municipalities and/or biomes (Figure 3), as shown by the high variation in T. cruzi prevalence in dogs in the different municipalities (11–89%, Table 1). Within the same biome, municipalities with high and low T. cruzi seroprevalence in dogs were observed. In the Jaguaruana and João Costa municipalities located in the Caatinga biome, dogs displayed 71% and 11% seropositivity, respectively. Similar differences in dog seroprevalence were also observed in distinct localities in the Amazon biome (22–89%). Even within the same municipality, seroprevalence in dogs was not homogeneous. In Abaetetuba (Amazonia biome), the seroprevalence in dogs ranged from 0 to 42% according to the locality. In the Pantanal, where Chagas disease is not reported, all dogs serologically examined were negative for T. cruzi infection (n = 39) (Table 1). Each dog was considered as one independent event, even when living in the same house due do the fact that different seroprevalence rates could be observed in dogs of one same house. The not homogeneous distribution of T. cruzi seroprevalence in dogs of one single house reflect that these dogs have not been equally exposed to T. cruzi infection.
Co-infection was observed in 17% and 16% of sera from dogs examined in the Caatinga and Amazon biomes, respectively, as evaluated by the IFAT and/or Quick Test for Diagnosis of Visceral Leishmaniasis (CVL) (TR DPP®, Bio-Manguinhos, FIOCRUZ, Rio de Janeiro, RJ, Brazil), indicating co-infection by both parasite species.
An analysis of the maps generated using the interpolation method indicated an inverse distribution correlation among T. cruzi infection in dogs and a decrease in the richness and abundance of small wild mammal species. This spatial correlation evidence was confirmed by map algebra and demonstrated that among the response variable and covariables there is also an inverse correlation, which indicated that in areas with greater richness and abundance of small mammal species, dogs were less prone to be infected with T. cruzi. Since the resulting map algebra was subtracted from inverse correlation with two variables our result shows a distribution that is not homogeneous. A more indirect indication was given by the rates of parasitological and serological T. cruzi prevalence in dogs (Figure 4A–D).
Statistical analysis confirmed the generated maps and demonstrated that the covariables DS (Species Richness) and THC (parasitological prevalence) influence the average response variable (T. cruzi infection in dogs), polygons of the studied municipalities. The estimated DS (−0.095596) indicates that in areas that present greater mammal biodiversity, dogs are less prone to infection by T. cruzi. The estimated TCH (0.009066) indicated that in areas that present higher parasitological prevalence of infection in the small wild mammals, dogs are more exposed to the T. cruzi infection. The estimated rate of T. cruzi infection in dogs was 0.909 (CI95% 0.870–0.949) for DS and 1.009 (CI95% 1.004–1.014) for TCH. The analysis of residuals versus fitted values indicated that the behavior of the variance of residuals and homoscedasticity presented random residuals. For confirmation of the normality of the data, the Shapiro-Wilk test was performed (W = 0.9814), P = 0.8714 at 5%. We found a significant difference in the T. cruzi infection rate between dogs sampled from areas that suffered a Chagas disease outbreak compared to dogs from non-outbreak areas (28/103 versus 217/546, Chi-squared 5.81, P = 0.01). In other words, this probability would rise 0.5 points in areas with Chagas disease outbreaks.
The sustainability of successful control of Chagas disease requires a more accurate knowledge of the environmental factors that underlie the transmission cycle of this parasite in the wild, mainly, if there are still unknown and undetermined aspects of the current epidemiology of this trypanosomiasis. This demands multidisciplinary and complex studies, as Trypanosoma cruzi is a multihost parasite that displays a huge intraspecific heterogeneity and a complex transmission cycle that may exhibit local peculiarities. Oral transmission of T. cruzi to humans was reported as sporadic until 2004, but in the following years, this epidemiological profile of transmission became increasingly important in the epidemiology of Chagas disease, particularly in the Amazon region [4]–[6]. Outside this region, oral transmission has also been responsible for recent outbreaks of ACD in several Brazilian states, mainly in the North [2], [3]. Such outbreaks have also been reported in other Latin American countries [37], [38].
Our results indicate that infection by T. cruzi in dogs is not homogeneous but focal, as demonstrated by the differences in seroprevalence among close localities; these differences may be due to landscape features. The seroprevalence observed in dogs could be associated with their proximity to forest and rural areas and with the loss of richness and abundance and rates of infection of the small wild mammal fauna. One aspect that distinguishes the present and previous data of our group from other studies is the scarcity of the number of dogs that displayed positive hemocultures [17]–[19]. In fact, dogs in Brazil are apparently only rarely involved in the amplification of T. cruzi and seem to play a minor role in the dispersion of the parasite. Even the dogs from Monte Alegre/PA seem not be of epidemiological importance because hemocultures were negative 3 months after the detection of T. cruzi in their blood smears.
The importance of a host species as a reservoir of a vector-borne parasite mainly depends on its prevalence of infection, capacity to infect the vectors, and the rate of host-vector contact [39]. A possible consequence of a local simplification of the small wild mammal fauna, where generalist mammals are favored at the expense of specialist species, is an increase in the rate of infection among the remnant mammalian fauna when the selected species are competent reservoirs of T. cruzi. As a result, the parasite population increases in the area, favoring vector infection and exposure of dogs to parasites, as reflected by their seroprevalence. This scenario suggests that the assessment of potential disease risk factors requires detailed knowledge of local, site-specific conditions. The small wild mammalian fauna diversity plays an important role in the profile of the enzootic infection patterns in a given area, as shown by the high transmission focus described in a previous study [22].
Overall, despite many remaining questions, the current evidence indicates that preserving intact ecosystems and their endemic biodiversity should generally reduce the prevalence of infectious diseases [7], [8].
The determination of the spatial distribution of the elements that compose the epidemiological chain of a parasitic disease is of pivotal importance for the determination of trends and risk evaluation. Moreover, it is worth mentioning that the attempts to control a given multihost parasite based on the control of one single vector or host species will always be insufficient because parasite transmission very rarely relies on a single system. The simplification of the mammalian host diversity, associated with an increase in the abundance of competent reservoir host species as described here is certainly one of the risk factors involved in the reemergence of Chagas disease [9]. Reduced disease risk with increasing host diversity is especially likely when pathogen transmission is frequency-dependent, and when pathogen transmission is greater within a species than between species, particularly when the most competent hosts are also relatively abundant and widespread [7].
Piauí and Ceará display similar patterns regarding the presence of a high density of naturally T. cruzi infected Triatominae, which are the main vectors in both regions. Our results demonstrate a high prevalence of T. cruzi infection in dogs from the Caatinga, as described in previous studies from our group [9], [22]. Despite this high prevalence, no new Chagas disease cases of vectorial transmission have been observed there in the last decade [13], [26]. This may reflect the effectiveness of the already long-lasting epidemiological surveillance campaigns exerted in these areas despite their lack of regularity. Local people are aware of the risk of disease and adopt local measures to avoid infection risk. Further, although dogs were exposed to the T. cruzi transmission cycle and are hosts of the parasite, they do not display high parasitemia (i.e., had negative hemocultures) and are therefore not involved in the amplification of parasite populations, so consequently, the potential for transmission from these dogs to the vector is low.
The high prevalence of seropositive dogs in the Amazon region can be attributed to the elevated rate of contact among these domestic animals and the wild environment and because the houses are practically located inside wild forest areas. In these areas, it is difficult to delimit the of peridomestic and wild areas, and many local inhabitants and dogs are involved in hunting activities. Empirical evidence indicates that habitat fragmentation can increase or decrease disease prevalence (and also T. cruzi infection among wild small mammals) within a host species, depending on the specific biology of the host–parasite relationship [28], [40]. Another important factor that should be taken into account is the importance of the definition of risk area based on the characteristics of the micro-regional management of domestic animals that are sometimes reared in semi-extensive ways. In this case, these animals are more exposed to the wild cycle of transmission and this is reflected by a high prevalence of infection. The presence of seropositive dogs in strictly domiciled habitats, as observed in a previous study in Navegantes, in the state of Santa Catarina, indicates, for example, the presence of a transmission cycle very close to the animal's home [2]. Moreover, the high prevalence of infection in domestic mammals reared in a semi-extensive way (such as pigs from Cachoeira do Arari/PA) indicates that transmission is occurring farther from homes but within the areas of interface between the peridomestic and wild environments [16].
Surveillance for canine Chagas disease should be a useful tool for the design of suitable epidemiological control programs in areas where sylvatic triatomines are responsible for human infection, as in many rural endemic areas [17].
The geospatial analysis approach involving interpolation and the map algebra method are a powerful tool in the study of the association between lower richness and areas with high transmission rates in small wild mammals and the risk of exposure of dogs to T. cruzi infection. Dogs are important sentinels and efficient indicators of areas at risk for Chagas disease outbreaks, lower richness in wild mammalian fauna diversity and selection of suitable T. cruzi reservoir hosts.
Therefore, the monitoring of domestic animals can and should be used as a first measure in the diagnosis of areas with elevated risk of T. cruzi transmission. Dogs, in particular, are easy to handle and have a generally accessible traceability. The collection of blood from these hosts and serologic testing (the sending of material to a central diagnostic institute) does not require great cost and infrastructure. Moreover, blood samples can be easily obtained in areas where dogs are routinely collected and tested for Leishmania sp. Or the anti-rabies vaccination campaigns can be used to collect blood from a representative sample of dogs in a given area. The presence of seropositive dogs reflects exposure to T. cruzi and points to the transmission of the parasite in areas where these animals roam. Once this measure is implemented, we should have an efficient indicator of areas at risk for human Chagas disease that require particular epidemiological investigation, implementation of control measures and health education.
|
10.1371/journal.pntd.0004637 | Mapping the Distribution of Anthrax in Mainland China, 2005–2013 | Anthrax, a global re-emerging zoonotic disease in recent years is enzootic in mainland China. Despite its significance to the public health, spatiotemporal distributions of the disease in human and livestock and its potential driving factors remain poorly understood.
Using the national surveillance data of human and livestock anthrax from 2005 to 2013, we conducted a retrospective epidemiological study and risk assessment of anthrax in mainland China. The potential determinants for the temporal and spatial distributions of human anthrax were also explored. We found that the majority of human anthrax cases were located in six provinces in western and northeastern China, and five clustering areas with higher incidences were identified. The disease mostly peaked in July or August, and males aged 30–49 years had higher incidence than other subgroups. Monthly incidence of human anthrax was positively correlated with monthly average temperature, relative humidity and monthly accumulative rainfall with lags of 0–2 months. A boosted regression trees (BRT) model at the county level reveals that densities of cattle, sheep and human, coverage of meadow, coverage of typical grassland, elevation, coverage of topsoil with pH > 6.1, concentration of organic carbon in topsoil, and the meteorological factors have contributed substantially to the spatial distribution of the disease. The model-predicted probability of occurrence of human cases in mainland China was mapped at the county level.
Anthrax in China was characterized by significant seasonality and spatial clustering. The spatial distribution of human anthrax was largely driven by livestock husbandry, human density, land cover, elevation, topsoil features and climate. Enhanced surveillance and intervention for livestock and human anthrax in the high-risk regions, particularly on the Qinghai-Tibetan Plateau, is the key to the prevention of human infections.
| Anthrax is a worldwide zoonosis affecting mostly grazing herbivores, with occasional spillover to humans who have contact with infected animals or contaminated animal products. We characterized the distributional patterns of both human and livestock anthrax in China from 2005 to 2013, and identified agro-ecological, environmental and meteorological factors contributing to the temporal and spatial distributions of the disease. We found that the spatial distribution of human anthrax in China was mainly driven by densities of cattle, sheep and humans, coverage of meadow, coverage of typical grassland, elevation, pH level of topsoil, concentration of organic carbon in topsoil, and meteorological factors. We also identified the regions with higher probabilities for the occurrence of human cases. Our findings provided a clear qualitative and quantitative understanding of the epidemiological characteristics and risk recognition of anthrax in China, and can be helpful for prioritizing surveillance and control programs in the future.
| Anthrax is one of the ancient zoonoses caused by Bacillus anthracis [1]. It is primarily a disease in herbivores and sometimes sparks outbreaks in human with potentially serious consequences [2]. It is enzootic in most countries in Africa and Asia as well as some countries in Europe and America [3]. The disease occurs worldwide with an estimate of 20,000 to 100,000 new human cases each year [4]. According to the World Health Organization (WHO), developing countries in Africa and those in central and southern Asia have the highest human incidences of naturally occurred anthrax [5]. Because of its wide distribution and its potential use for bioterrorism, anthrax is considered as a global public health threat [6]. Concerns have been heightened by the persistent existence of human anthrax cases and outbreaks across continents in recent years, e.g., Zambia, Zimbabwe and Ethiopia in Africa [7–9], India, Bhutan, Bangladesh and Georgia in Asia [10–14], and Turkey, Greece and Serbia in Europe [15–17]. In addition, the emergence of “injectional anthrax” among heroin users in Europe highlights the possibility of new routes for the spread of human anthrax [18, 19].
Bacillus anthracis, the causative agent of anthrax, is a sporulating Gram-positive bacterium that manifests a particular bimodal lifestyle: the vegetative phase and the spore phase [2]. Bacteria in the vegetative phase are shed by infected animals and may die rapidly in most environmental conditions. Once sporulated from the vegetative cells, the bacteria can survive in soil for decades [20]. It has been speculated that levels of pH and calcium cation in the soil play an important role in the process of germination or in maintaining spore’s viability. Besides, the organic matter in the soil affects spore adhesion [5, 21, 22]. As a result, topsoil conditions may geographically regulate the distribution of anthrax infections. Some other environmental factors including climatic conditions could also be associated with anthrax infection in herbivores and humans [21, 23]. Herbivores, the primary hosts of this pathogen, are usually infected anthrax by ingestion of spores while grazing or browsing [5]. Human infection was usually a result of contacting ill animals during agricultural activities or processing contaminated animal products [24, 25]. Limited person-to-person transmission has been reported. Human anthrax cases are classified into three forms according to the transmission route: the cutaneous form accounts for about 95% of all human cases worldwide, the gastrointestinal form, and the inhalational form.
More than 112 thousands human cases have been reported in China from 1956 to 1997 with three large-scale outbreaks in the years of 1957, 1963 and 1977, respectively [26]. In the recent decades, human and livestock anthrax outbreaks have been reported in many provinces across the nation, such as Liaoning, Inner Mongolia Autonomous Region, Jiangsu, Guizhou, and Xinjiang Autonomous Region [27–31]. Previous studies mainly focused on local case-report, outbreak investigations, or the spatial and temporal distribution of cutaneous anthrax of human cases in China [27–33]. Using the national surveillance data of human and livestock anthrax from 2005 to 2013, we conducted a comprehensive and in-depth retrospective epidemiological study on the spatiotemporal dynamics and risk determinants of anthrax in mainland China.
The National Health and Family Planning Commission of China considers the collection of data from human and livestock anthrax cases as part of its routine surveillance, and such data collection is therefore exempt from approval by the institutional review board.
In China, inhalational anthrax is managed as a class A infectious disease, while other forms of human anthrax are listed as one of the class B infectious diseases. Cases diagnosed at medical institutions were reported to the Chinese Centre for Disease Control and Prevention (CCDC) through the web-based national Notifiable Infectious Diseases Reporting Information System (NIDRIS) (http://www.cdpc.chinacdc.cn/UVSSERVER2.0/login?fromSmp=true&fromCDC3=true&service=http%3A%2F%2Fwww.cdpc.chinacdc.cn%2Fportal%2FcasAuthUser%3Fvsite%3Dguojia). All clinically-diagnosed and laboratory-confirmed cases during 2005–2013 were included in this study. Routine surveillance of livestock anthrax is conducted by the Ministry of Agriculture of the People's Republic of China. The surveillance data are published monthly on the Official Veterinary Bulletin (http://www.moa.gov.cn/zwllm/tzgg/gb/sygb/), from which we extracted the monthly numbers of livestock cases and outbreaks as well as affected species in each province during the study period. The case definitions for human and livestock anthrax are stated in the S1 Text.
Data concerning agro-ecological, environmental and meteorological factors were collected for exploring potential determinants for the temporal and spatial distributions of human anthrax in mainland China (S1 Table). Raster-typed data with a 5 km2 resolution regarding the density of cattle, sheep and goats were obtained from the Food and Agriculture Organization of the United Nations (http://www.fao.org/AG/againfo/resources/en/glw/GLW_dens.html). The human population size and the annual number of livestock including cattle, sheep, goats, pigs and horses were obtained from the National Bureau of Statistics of China (http://www.stats.gov.cn/). The land cover data with a 1 km2 resolution in 2005 was collected from the Data Sharing Infrastructure of Earth System Science (http://www.geodata.cn). The elevation data were obtained from Global digital elevation data products (http://www.geodata.cn/data/datadetails.html?dataguid=201519481253546&docid=1301). The soil-related variables including pH level, concentration of organic carbon, and concentration of calcium in topsoil with a 1 km2 resolution were derived from the Harmonized World Soil Database (http://webarchive.iiasa.ac.at/Research/LUC/External-World-soil-database/HTML/). The climatic variables including monthly average temperature and relative humidity, as well as monthly accumulative rainfall and sunshine hours during the study period were obtained from the China Meteorological Data Sharing Service System (http://data.cma.cn/).
Using the spatial analytic methods in the ArcGIS 9.2 software (ESRI Inc., Redlands, CA, USA), we extracted the following 15 variables for each county: densities of cattle, sheep, goats and human, percentage coverages of meadow, typical grassland and alpine steppe, average elevation, percentage coverage of topsoil with pH > 6.1, average concentration levels of organic carbon and calcium in topsoil [21], monthly average temperature, relative humidity, and yearly accumulative rainfall, sunshine hours during the study period.
Monthly numbers of human cases were plotted to display the seasonal dynamic of the disease. Annual incidence rates in human and annual numbers of livestock cases were plotted to show the overall temporal trend. The bar charts of average age- and gender-specific incidences were created during the study period, and the proportions of human cases by occupation were calculated. Demographic data from the 2010 census were used to calculate the average annual incidences for each county, standardized by the national distribution of age (10-year age-group categories) and sex (male and female).
The Kulldorff retrospective spatial-temporal and spatial-only scan statistics were used to detect human anthrax clustering areas at the county level (SaTScan software, version 9.3, http://www.satscan.org) [34]. The spatial-temporal statistic was calculated by forming the cylindrical windows with a geographic circle up to 50% of the population at risk and a time span up to 90% of the study period. The spatial-only scan statistic was calculated for each year using the maximum spatial cluster size of 10% of the population at risk. Clustering areas were identified using the log likelihood ratio (LLR) test based on a Poisson model, and the significance levels were evaluated using 999 Monte Carlo replications. A P value ≤ 0.05 was considered statistically significant.
A thematic map of the standardized average annual incidences of human anthrax was created, and the cumulative numbers of livestock anthrax cases and the identified spatial-temporal clustering areas were overlapped on the thematic map. To illustrate the spatiotemporal dynamics of human anthrax cases from 2005 to 2013, a bar chart of the annual numbers of human anthrax cases over the study period were shown for each province on the map. In addition, a map series for the spatial distribution of annual clustering areas and annual incidence rates of human anthrax were created. Spearman correlation was used to examine the association between monthly human anthrax incidence and each climatic variable (temperature, relative humidity, rainfall and sunshine hours) within the most likely clustering area using the Stata software, version 11.0 (StataCorp LP, Texas, USA) [35].
A boosted regression trees (BRT) model was applied to explore the potential determinants of spatial distribution of human anthrax at the county level. It is a method that combines the advantages of two algorithms, regression trees and boosting, and is able to accommodate non-linear relationships between outcomes and covariates and multiway interactions among covariates [36]. The weight of each variable estimated from all identified trees represents the influence of that variable in predicting the outcome. The BRT approach has been applied to the risk mapping for infectious diseases such as avian influenza A (H7N9), highly pathogenic avian influenza (H5N1) and dengue fever [37–40]. We used the data from 2005 to 2011 to construct the BRT model and the data from 2012 to 2013 to assess the model’s predictive power. We adopted a tree complexity of 5, a learning rate of 0.005 and a bag fraction of 75% to identify the optimal trees for each bootstrap data.
To address the issue of multicollinearity between climate variables, principal component analysis (PCA) was performed for these variables before the BRT modeling [41]. The principal component with the largest eigenvalue (>1.0) accounts for 84.61% of the total variability (S2 Table). We refer to this principal component as the meteorological index and included it as a covariate in the BRT model. Temperature, relative humidity, rainfall and sunshine hours contributed almost equally to the meteorological index, although only sunshine hours has a negative loading (S2 Table). The BRT modeling was carried out in sequential steps. First, 1,360 counties were randomly selected without replacement from all 2,650 counties without reported human anthrax cases throughout mainland China. These control counties were then combined with the 272 counties with human anthrax cases reported during 2005–2011 to form a bootstrap dataset with a 1-to-5 case-control ratio. Second, the bootstrap dataset was randomly portioned into a training dataset with 75% of the counties and a test dataset with 25%. Third, a BRT model was built using the training dataset and validated using the test dataset. Finally, the fitted model was used to predict the probability of human anthrax presence during 2012–2013 for all counties in mainland China. These steps were repeated for 50 times. Each time, a receiver-operating characteristic (ROC) curve was produced, and the area under the curve (AUC) was calculated to evaluate the predictive power of the model. A risk map of human anthrax infections in 2012–2013 was then created based on the average predicted probabilities of the 50 repetitions.
A total of 3,115 human anthrax cases were reported in mainland China during 2005–2013. Cutaneous anthrax accounted for 97.7% of all the cases. The majority of the cases (72.2%) occurred in summer and autumn, and the disease usually peaked in July or August (Fig 1A). Males had a higher average annual incidence than females in the age groups of 20 years and above (Chi-square test, P < 0.01). The highest age-specific incidences were found in the age groups of 30–39 and 40–49 years for both males and females (Fig 1B). Herdsmen and peasants accounted for 88.7% of all cases reported during 2010–2013, followed by children less than 6 years old (31 cases, 3.0%) (Table 1). During 2005–2013, a total of 2,261 livestock anthrax cases were reported, the majority of which were cattle, sheep, goats and pigs (S3 Table). The epizootic curve of livestock anthrax was more fluctuating than the human epidemic curve but still showed an overall decreasing trend (Cochran-Armitage trend test, P < 0.01). It was moderately correlated with the epidemic of human anthrax, with a Spearman correlation coefficient of 0.38 (95% CI: 0.20–0.54, P < 0.01).
Human anthrax cases were distributed in 299 counties of 19 provinces with an average annual incidence of 0.39 per 100, 000 person years (range: 0.01–51.98). About 82% of all cases were located in six provinces/autonomous regions of western and northeastern China, including Sichuan, Xinjiang, Guizhou, Gansu, Qinghai, and Inner Mongolia (Table 1). The remaining cases were distributed sporadically in 13 provinces/autonomous regions (Fig 2). Four provinces/autonomous regions, Gansu, Qinghai, Yunnan and Inner Mongolia, showed a rebound in the number of human cases in recent years, despite the overall decreasing trend in the whole China (S1 Fig). Qinghai is the province that suffered the most from livestock anthrax during the study period. The spatial distribution of human anthrax was mostly consistent with that of livestock anthrax, except for Sichuan Province (Fig 2, S1 Fig).
S2 Fig shows the dynamics of spatial clustering areas and standardized annual anthrax incidence of human anthrax. The most likely clusters were persistently located on the eastern Qinghai-Tibet Plateau, whereas the locations of secondary clusters varied over time. The spatiotemporal scan statistic identified one most likely cluster and four secondary clusters during the entire study period (2005–2013) (Fig 2). The most likely cluster consists of 34 counties on the junction of Sichuan, Gansu, Qinghai and Tibet provinces/autonomous regions, and spanned from January 2005 to January 2013, with a relative risk of 424.3 in comparison to counties outside the cluster. In the most likely cluster, monthly human anthrax incidence was positively correlated with monthly average temperature, relative humidity and monthly accumulative rainfall at lags of 0–2 months (S4 Table, S3 Fig). For the three climatic variables, the Spearman correlation coefficients ranged from 0.67 to 0.70 at lags of 0 or 1 month, and decreased with longer time lags. Sunshine hours at lags of 0–1 months were marginally correlated with monthly incidence of human anthrax (P = 0.048).
The BRT model found that the spatial distribution of human anthrax was significantly associated with the densities of cattle, sheep and human, coverage of meadow, coverage of typical grassland, elevation, coverage of topsoil with pH > 6.1, concentration of organic carbon in topsoil, and the meteorological index. All above potential predictors had weights (relative contribution) of more than 5.0 in the BRT models (Table 2). The maps of human anthrax incidence overlaid by these variables were also created to display the potential spatial association between them (S4 and S5 Figs). The probability of occurrence of human cases increased with higher values of densities of cattle and sheep, coverage of meadow, and concentration of organic carbon in topsoil. The risk rose quickly with higher elevations in the range of 500–1500 m, and plateaued or dropped for elevations above 1500m. In addition, the probability of occurrence of human anthrax cases was negatively associated with the density of human population, and the meteorological index (S6A Fig).
The estimated AUC value of 0.921 (95% CI: 0.899–0.943) indicates a decent predictive power for the probability of occurrence of human cases. The AUC estimates are 0.975 (95% CI: 0.966–0.984) and 0.892 (95% CI: 0.873–0.912) for the training and test dataset, respectively (S6B Fig). To avoid overfitting of the model, which is possible considering the enormous heterogeneity in all variables across the whole country, we performed a sensitivity analysis by restricting case counties and the sampling of control counties to seven provinces (Sichuan, Qinghai, Gansu, Guizhou, Inner Mongolia, Heilongjiang and Liaoning provinces /autonomous regions) where spatiotemporal clusters of the disease were located. The resulted ROC and AUC’s are similar to our final model based on the data of the whole country (S6C Fig). On the basis of the average predicted probability of occurrence of human cases for each county in 2012–2013, the high-risk areas of human anthrax were mainly distributed in four regions: (1) the central-west high-risk region that contains most of the Qinghai-Tibetan Plateau, and covers eastern Qinghai, northwestern Sichuan, southwestern Gansu, and central Tibet; (2) the southwest high-risk region that consists of Yunnan, Guizhou and western Guangxi provinces; (3) the northwest high-risk region that covers western and northwestern Xinjiang; and (4) the north high-risk region that covers central and eastern Inner Mongolia, western and eastern Heilongjiang, and Jilin provinces (Fig 3). By superimposing the locations of reported human anthrax cases on the predictive risk map during 2012–2013, we found that 93.1% of reported cases were located in the high risk counties each with a probability of occurrence of human cases more than 0.7. Coinciding with the most likely cluster, the eastern part of the central-west high-risk region has the highest risk of occurrence of human cases.
Our study provides a complete overview of spatiotemporal dynamics of human and livestock anthrax in mainland China from 2005 to 2013. It is the first national risk assessment of human anthrax occurrence in China. Our study identified five clustering areas of human anthrax cases and four potential high-risk regions for the occurrence of human cases. In addition, we quantified the relationship of climate factors to the temporal trend of human anthrax and the contribution of agro-ecological, environmental and meteorological factors to the spatial distribution of human anthrax.
We found that males had a higher incidence than females, in particular for adults, probably due to the occupational exposure. Among herders and farmers, men are usually more exposed to livestock than women by undertaking most agricultural activities such as pasturing and slaughtering [13,14]. Although incidences of human cases and livestock cases were significantly correlated, inconsistency was found in their spatial distributions in some provinces, e.g., Tibet and Sichuan, likely due to underreporting of anthrax case in livestock or the possibility that more than one person may contract the disease from a single animal [42]. A serological surveillance study carried out by the Chinese Institute of Epidemiology and Microbiology during 1990–1994 showed that 29.1% of human samples and 31.2% of livestock samples had detectable antibodies to the capsular of Bacillus anthracis at outbreak spots [43]. Since then, there have been very few studies on serological prevalence of anthrax both in human and livestock across the country. It is necessary to strengthen the surveillance of anthrax in livestock, especially in the four high-risk regions identified in this study.
Densities of cattle and sheep were identified as useful predictors for the risk of human anthrax. Suitable habitat conditions for these livestock were also important predictors, i.e., a higher coverage of meadow, a higher elevation or a lower human density. Unlike brucellosis (another zoonotic disease) [44], the presence of human anthrax was not found to be associated with the density of goats. This could be partially explained by the difference in the feeding habits between goats and other livestock. It was documented that cattle ingest lots of soil by pulling the plant out of the ground when grazing, whereas goats usually browse on grass only, which makes them less exposed to spores in soil [5, 21]. There were only two outbreaks of anthrax among goats during the study period, as compared to 107 outbreaks in cattle and 21 in sheep (S3 Table).
The ability of Bacillus anthracis to form long-lasting, highly resistant spores is the key to the persistence of anthrax in any area [22]. Certain soil characteristics, such as high levels of organic matter, pH or calcium, were thought to facilitate the survival of spores [5, 21, 22, 45], although the role of the calcium level was not always clear [46]. We found the evidence for the ecological association of human anthrax with the pH value of topsoil and the concentration of organic carbon in topsoil (Table 2, S5C and S5D Fig), but the concentration of calcium in topsoil was not picked by our model (Table 2, S5E Fig). In our data, the presence of human anthrax was largely driven by the distribution of livestock and suitable habitats for them, and the calcium level in the topsoil of the habitats is generally lower than that of non-epidemic areas, as shown in S5E Fig. However, the possibility of spatial heterogeneity in the effect of the calcium level on the presence of the disease cannot be ruled out.
Temperature, relative humidity and rainfall were positively correlated over time with human anthrax in the most likely clustering area (S4 Table, S3 Fig). Increased rainfall and temperature in the summer could unearth the anthrax spores and facilitate the breeding of vectors, such as tabanids and Stomoxys [5, 21, 47, 48]. Moreover, the transfer of Bacillus anthracis between the vegetative form and the spore was thought to be related to temperature and relative humidity [49]. However, the exact ecological mechanism of climatic influence on the seasonality of anthrax is not clear and may vary across geographic regions. For example, drought easily makes the herbivores more exposed to the spores in soil by inhibiting the growth of grass [5, 21]. Extreme temperatures may also depress innate immunity of the host, reducing the minimal dose of anthrax for infection [5, 47]. In contrast to the positive temporal association of the climate variables to the incidence of the disease, it is interesting that the spatial risk distribution of human anthrax was negatively associated with the meteorological index and thus negatively associated with average temperature, relative humidity and rainfall. The negative association in the spatial dimension and the positive association in the temporal dimension at any given location do not contradict each other. In addition, the spatial association was partially due to the relatively high elevations of livestock habitats where average annual temperature, relative humidity and rainfall are relatively low.
Our results should be interpreted with the following limitations in mind. First, human and livestock cases could have been under-reported as the surveillance was passive. The changes in the diagnostic criteria for human anthrax cases since 2008 might have affected the quantity of the reported data. Second, the BRT model is an ecological analysis of relative contributions of risk factors and offers no causal interpretation. The causality of the identified risk factors can be tested with appropriately designed studies in the future. Third, some relevant risk factors were not available to refine our exploration, including but not limited to seroprevalence in human and livestock, exposure level of people at risk, and the industrialization level of livestock production [21].
Although anthrax underwent an overall decreasing trend during the study period, incidences rebounded and outbreaks were reported in recent years in some provinces [27–31]. While imposing threats to both animal productivity and human health in affected communities, anthrax remains a largely neglected zoonosis [50]. Existing surveillance programs for anthrax should be improved and expanded to cover livestock, human, and environmental samples, a “One Health” approach [14, 51, 52]. In addition, vaccination of both livestock and human would be essential for disease prevention and can be prioritized for high-risk regions identified in our work [42, 53].
|
10.1371/journal.ppat.1003799 | Myeloid Dendritic Cells Induce HIV-1 Latency in Non-proliferating CD4+ T Cells | Latently infected resting CD4+ T cells are a major barrier to HIV cure. Understanding how latency is established, maintained and reversed is critical to identifying novel strategies to eliminate latently infected cells. We demonstrate here that co-culture of resting CD4+ T cells and syngeneic myeloid dendritic cells (mDC) can dramatically increase the frequency of HIV DNA integration and latent HIV infection in non-proliferating memory, but not naïve, CD4+ T cells. Latency was eliminated when cell-to-cell contact was prevented in the mDC-T cell co-cultures and reduced when clustering was minimised in the mDC-T cell co-cultures. Supernatants from infected mDC-T cell co-cultures did not facilitate the establishment of latency, consistent with cell-cell contact and not a soluble factor being critical for mediating latent infection of resting CD4+ T cells. Gene expression in non-proliferating CD4+ T cells, enriched for latent infection, showed significant changes in the expression of genes involved in cellular activation and interferon regulated pathways, including the down-regulation of genes controlling both NF-κB and cell cycle. We conclude that mDC play a key role in the establishment of HIV latency in resting memory CD4+ T cells, which is predominantly mediated through signalling during DC-T cell contact.
| Current antiretroviral drugs significantly prolong life and reduce morbidity but are unable to cure HIV. While on treatment, the virus is able to hide in resting memory T cells in a silent or “latent” form. These latently infected cells are rare and thus are hard to study using blood from HIV-infected individuals on treatment. Therefore, it is very important to have laboratory models that can closely mimic what is going on in the body. We have developed a novel model of HIV latency in the laboratory. Using this model we have shown that the presence of dendritic cells, an important type of immune cell that can regulate T cell activation, at the time of infection allows for the infection of resting T cells and the establishment of latency. We have demonstrated that this is predominantly mediated by direct cell-to-cell interactions. Further exploration of the mechanisms behind HIV latency could lead to new ways to treat and possibly eradicate HIV.
| Antiretroviral therapy (ART) for the treatment of HIV has led to a substantial reduction in morbidity and mortality; however, ART cannot cure HIV and life-long treatment is required. This is directly due to the persistence of long-lived latently infected cellular reservoirs, that include microglia, astrocytes, macrophages and naïve T cells [1]–[4], however, resting memory CD4+ T cells [5]–[7], are considered to be the major contributors. Latently infected resting CD4+ T cells are found in blood and tissue sites, including lymphoid tissue and the gastrointestinal tract [7]–[10]. The frequency of latently infected cells is up to ten times higher in tissue than in blood in HIV-infected patients or SIV-infected macaques on suppressive ART [8], [10].
It is unclear how latency is established in vivo. However, in vitro, latency can be established following survival of an activated CD4+ T cell that returns to a resting state carrying integrated virus [5], [11]–[13]. Alternatively, latency has also been established following direct infection of resting cells in the presence of chemokines or following spinoculation [14]–[19]. Dendritic cells (DC) are found throughout the body and interact closely with resting CD4+ T cells within lymphoid tissues. Therefore, given the high frequency of latently infected cells in lymphoid tissue, we hypothesised that latency in resting CD4+ T cells may result from interactions with DC as CD4+ T cells recirculate through lymphoid tissue. Using a novel model of resting CD4+ T cells co-cultured with primary DC, we demonstrate that myeloid DC (mDC) induce post-integration latency in resting memory CD4+ T cells, which required close DC-T cell contact.
Resting CD4+ T cells and syngeneic DC (including the two major blood DC subpopulations, plasmacytoid (pDC) and myeloid (mDC) DC) were sorted from the blood of healthy donors (Fig. S1). Seminaphtharhodafluor-1 (SNARF)-labelled resting CD4+ T cells were cultured either alone or co-cultured with DC at a DC: T cell ratio of 1: 10. Following 24 hours of culture, cells were infected with a CCR5-tropic enhanced green fluorescent protein (EGFP)-reporter virus, NL(AD8)-nef/EGFP (multiplicity of infection, MOI 0.5), and cultured for 5 days (Fig. 1 A). Cells were then analysed for expression of EGFP by flow cytometry to quantify productive infection (Fig. 1 B).
In the DC-CD4+ T cell co-cultures, we detected a spreading productive infection with the number of infected cells 5 days post-infection significantly greater (median (IQR) = 40 (31, 150) EGFP+ cells/104 cells; n = 5) compared to CD4+ T cells cultured alone (1.5 (1, 2.5) EGFP+ cells/104 cells, p = 0.03; Fig. 1 C). These results were consistent with previous work demonstrating enhanced productive infection of CD4+ T cells in the presence of DC [20], [21].
At day 5 post-infection, non-proliferating (SNARFhi) CD4+ T cells that were not productively infected (EGFP−) were sorted (purity was always >99%). Latent virus was quantified in the SNARFhiEGFP− CD4+ T cells upon stimulation with phytohaemagglutinin (PHA) in the presence of feeder peripheral blood mononuclear cells (PBMC) following a further 5 days of culture (Fig. 1 B). The number of EGFP+ cells following stimulation was, therefore, a surrogate measure for the number of latently infected cells in the SNARFhiEGFP− CD4+ T cells. When SNARFhiEGFP− CD4+ T cells were sorted from cultures infected in the absence of DC, few latently infected cells were detected (2 (1, 8.5) EGFP+ cells/104 cells; n = 5). In contrast, when SNARFhiEGFP− CD4+ T cells were sorted from the DC-T cell co-cultures a significant increase in the number of latently infected cells was observed (41 (28, 73) EGFP+ cells/104 cells; p = 0.03; n = 5; Fig. 1 D). Furthermore, when infections were performed in the presence of the protease inhibitor indinavir, there was no significant difference in the number of latently infected cells, as measured by EGFP expression following co-culture of SNARFhiEGFP− with PHA and feeder PBMC (Fig. 1 E). This confirms that a productive, spreading infection was not required to establish latency. Together, these results demonstrate that DC facilitate latent HIV infection in non-proliferating CD4+ T cells.
We next asked whether DC-T cell co-culture had activated the SNARFhi CD4+ T cells, which allowed for HIV entry. Sorted SNARFhiEGFP− CD4+ T cells that were co-cultured with DC for 5 days showed signs of early activation with increased expression of CD69 (1.5% (0.1, 2.1); p = 0.02; n = 4; Fig. 1 F). However, these cells did not express either HLA-DR or Ki67 (Fig. 1 F). As expected, resting CD4+ T cells that were cultured alone did not express any of the activation markers. These results confirmed that the sorted SNARFhiEGFP− cells were non-proliferating, partially activated CD4+ T cells.
To determine whether mDC or pDC were facilitating latency in resting CD4+ T cells, we next co-cultured sorted mDC and pDC with SNARF-labelled resting CD4+ T cells for 24 hours prior to infection, and experiments were performed as described above. While productive infection was enhanced in both the mDC and pDC-T cell co-cultures (Fig. 2 A), latent infection was only identified in the SNARFhiEGFP− CD4+ T cells that had been co-cultured with mDC (33 (19, 51) EGFP+ cells/104 cells; n = 5) following re-stimulation with PHA and feeder PBMC (Fig. 2 B) or following direct activation with anti-CD3/CD28, together with IL-7 and the integrase inhibitor L8, which allowed the detection of post-integration latency (Fig. 2 C). To further confirm that post-integration latency was established in resting CD4+ T cells co-cultured with mDC, we used a real time PCR assay to quantify integrated HIV DNA. Integrated HIV DNA was present in SNARFhiEGFP− CD4+ T cells sorted from the mDC co-cultures (1100 (686, 4960) HIV DNA copies/106 cells, n = 3; Fig. 2 D), but not in the CD4+ T cells sorted from the pDC co-cultures or the CD4+ T cells cultured alone (both <330 HIV DNA copies/106 cells). Similar results were observed with a nef competent EGFP-reporter virus (Fig. 2 E), demonstrating that the establishment of mDC-induced latency was not dependent on Nef.
Unlike experiments that utilised R5 EGFP HIV, when experiments were performed with an X4 EGFP reporter virus, latent infection was detected in the resting CD4+ T cells cultured alone (87 (51, 155) EGFP+ cells/104 cells; n = 4; Fig. 2 F). However, latency was still significantly enhanced in the non-proliferating CD4+ T cells in the presence of mDC (468 (213, 621) EGFP+ cells/104 cells) when compared to the T cells cultured alone.
In some experiments we added a low dose of Staphylococcus enterotoxin B (SEB; 10 ng/mL) to the mDC-T cell co-cultures to enhance productive infection and increase cognate interactions between mDC and T cells. In the presence of SEB we observed a significant increase in the level of productive infection; however, there was no difference in the level of latent infection (Fig. 2 G). Finally, we cultured mDC and T cells together at ratios ranging from 1∶10 to 1∶100 to determine the minimum interaction required between mDC and T cells to induce latency. We found that latency could still be established at a ratio of DC: T cells as low as 1∶100 (Fig. 2 H).
Taken together, these results demonstrated that in vitro mDC and not pDC facilitated post-integration latency in non-proliferating CD4+ T cells.
We have previously shown that memory CD4+ T cells and not naïve CD4+ T cells are susceptible to latent infection following chemokine exposure [14]. To determine whether mDC-induced T cell latency occurred in memory or naïve CD4+ T cells, we separated the SNARFhiEGFP− CD4+ T cells into CD45RO+ (memory) and CD45RO− (naïve) fractions prior to culture with feeder PBMC. In these experiments, latent infection was detected at significantly higher levels in the CD45RO+ memory CD4+ T cell fraction (146 (14, 197) EGFP+ cells/104 cells; Fig. 2 I).
A proportion of non-proliferating SNARFhiEGFP− cells sorted from DC-T cell co-cultures (Fig. 1 F and 3 A) expressed CD69. Therefore, in order to exclude the possibility that we were only detecting infection of the cells showing early signs of activation, we depleted CD69+ cells from the SNARFhiEGFP− T cells at day 5 post-infection prior to co-culture with PHA and feeder PBMC. We found no significant difference in the level of latency following depletion of the CD69+ cells (Fig. 3 B). CD69 expression can be transient, therefore, to confirm that we had not missed cells that expressed CD69, which had then been down-regulated; we measured the expression of CD69 over time following co-culture with mDC and infection with HIV. We demonstrated that CD69 expression peaked at day 2 and remained elevated out to day 5 post-infection (data not shown). These results demonstrate that the subpopulation of CD4+ T cells that were partially activated and expressing CD69 were not preferentially latently infected following mDC-T cell co-culture.
To determine why mDC and not pDC led to the establishment of latency, we compared cytokine levels in pDC-T cell and mDC-T cell co-cultures 5 days following HIV infection using bead arrays for known DC-secreted cytokines. Supernatants collected from HIV-infected mDC-T cell co-cultures compared to the HIV-infected pDC-T cell co-cultures had significantly increased expression of IL-6 (p = 0.002), IL-10 (p = 0.01) and CXCL9 (p = 0.002; Fig. 4 A). TNF-alpha was also up-regulated in the mDC-T cell compared to pDC-T cell co-cultures but the difference was not statistically significant (Fig. 4 A). As expected, IFN-alpha was detected at high levels in the pDC-T cell co-cultures but not in the mDC-T cell co-cultures (Fig. 4 A; p = 0.01), and latency was not established in pDC-T cell co-cultures even in the presence of neutralising antibodies to IFN-alpha (Fig. 4 B). Interestingly, when equal numbers of pDC and mDC were added to resting CD4+ T cells latency was reduced when compared to T cells cultured only with mDC (Fig. 4 C). This suggests that while pDC themselves do not induce latency, they are able to inhibit the establishment of latency mediated by mDC.
To determine whether the soluble factors that were differentially expressed in mDC-T cell co-cultures compared to pDC-T cell co-cultures were contributing to the establishment of mDC-induced latency, neutralising antibodies (nAb), to either the soluble factor or its receptor, were added to the T cells prior to co-culture with DC and the addition of HIV. Specific nAbs or an anti-IgG control were added to eFluor670 (alternative proliferation dye to SNARF)-labelled resting memory (CD45RO+) CD4+ T cells prior to co-culture with mDC, and again following infection, and latency determined as described in Fig. 1 (Fig. 4 D).
When nAbs to IL-6, the IL-10 receptor (IL-10R) or CXCR3 (CXCL9 and CXCL10 receptor) were added to the mDC-T cell co-cultures, no significant decrease in the number of latently infected CD4+ T cells was observed when compared to cultures where control anti-IgG was added. As we had previously shown that the chemokine CCL19 can condition resting CD4+ T cells allowing for enhanced entry and integration of HIV [14], we also added anti-CCL19, either alone or in combination with anti-CXCR3, to the mDC-T cell co-cultures. However, we did not detect a significant decrease in the number of latently infected cells. The activity of these nAbs was confirmed by their ability to block STAT3 signalling (aIL-6, aIL-10R) or chemokine induced migration (aCXCR3, aCCL19; Fig. S2).
To determine whether DC-T cell contact was required to establish latency in resting CD4+ T cells, we co-cultured mDC and resting CD4+ T cells separated by a 0.4 µm membrane transwell. Following 24 hours of culture, HIV was added to the mDC in the upper chamber and the CD4+ T cells in the lower chamber. Without mDC-T cell contact, the establishment of latency was significantly inhibited (<1 EGFP+ cell/104 cells) when compared to co-cultures without membranes (18 (10, 85) EGFP+ cell/104 cells; n = 5; p = 0.03; Fig. 4 E).
To further elucidate whether soluble factors other than those previously inhibited were involved in DC-induced latency, we added supernatant from infected mDC-T cell co-cultures to uninfected resting CD4+ T cells and then infected these cells with EGFP-HIV. Media changes were performed daily using supernatant from infected mDC-T cell co-cultures. Under these conditions the resting CD4+ T cells would be exposed to any soluble factors and free viral particles present in the mDC-T cell co-cultures but would not have any contact with the mDC. Latency was not detected in these cultures (Fig. 4 F), providing further evidence that DC-T cell contact, and not a soluble factor, was required for the establishment of DC-induced latency.
Direct DC-T cell signalling can occur following interactions between several cell surface receptors. In particular, interactions between lymphocyte function associated antigen-1 (LFA-1; composed of CD11a and CD18) on T cells and intercellular adhesion molecule 1 (ICAM-1) on DC are involved in DC-T cell adhesion [22] and subsequent T cell activation via formation of an immunological synapse [23]. To inhibit DC-T cell clustering, we used blocking antibodies to CD18 (10–20 µg/mL). Blocking of CD18 significantly inhibited, but did not eliminate, DC-T cell clustering, as observed by microscopy (data not shown). Following incubation with anti-CD18, there was no effect on the number of productively infected cells, however, we observed a significant decrease in the number of latently infected cells from the mDC-T cell co-cultures (20 (8, 34) latently infected cells/104 cells), when compared to cells cultured without anti-CD18 (25 (17, 48) latently infected cells/104 cells; p = 0.03) or in the presence of control anti-IgG (32 (22, 36) latently infected cells/104 cells; n = 6; p = 0.03; Fig. 5 A). However, when resting CD4+ T cells were stimulated with soluble ICAM-1 and anti-IgG there was no increase in latency observed (Fig. 5 B), suggesting that ICAM-LFA signalling alone does not induce latency in this model system. Furthermore, in the presence of anti-CD18 the number of latently infected cells from the mDC-T cell co-cultures remained greater than the CD4+ T cells cultured alone (<1 latently infected cell/104 viable cells; p = 0.01) suggesting that the effect of anti-CD18 was most likely due to the partial decrease in clustering/DC-T cell contact.
In order to determine whether mDC transfer of HIV was involved in the establishment of latency we performed experiments where we added virus to resting CD4+ T cells, washed off virus and added uninfected mDC to the CD4+ T cells (Fig. 5 C). Under these conditions we were still able to detect latency in the non-proliferating CD4+ T cells. Together, these results indicate that cell-cell contact plays a role in DC-induced T cell latency but that the mDC were not required to be infected and then transfer HIV to the resting CD4+ T cells.
To determine the effect of DC on gene transcription in latently infected resting CD4+ T cells, SNARF-labelled resting CD4+ T cells, from four independent donors, were cultured either alone or with syngeneic bulk blood DC at a 1∶10 ratio for 24 hours prior to infection with NL(AD8)-nef/EGFP. In these experiments, we included IL-7 (10 ng/mL) in all cultures to increase cell survival of the resting cells and infections were performed at an MOI of 5 to ensure a high frequency of latently infected cells. Mock infections were performed in parallel with media alone. Non-proliferating (SNARFhi) CD4+ T cells that were not productively infected (EGFP−) were sorted 5 days post-infection and lysed for either the detection of HIV DNA by real-time PCR or RNA for microarray studies (Fig. 6 A).
Infection was confirmed in the resting CD4+ T cells following co-culture with DC, in 4 independent experiments, by detection of HIV DNA (3×104 (7.4×103, 5.7×105) copies/106 cells; Fig. 6 B). Changes in gene expression were quantified in the sorted SNARFhiEGFP− CD4+ T cells using Illumina oligonucleotide microarrays. To identify genes expressed in DC-induced latency, we compared the expression profiles of non-proliferating, latently infected CD4+ T cells (HIV T (+DC)) to mock infected CD4+ T cells (Mock T (+DC)) that had been co-cultured with DC. In order to control for the effect of virus or DC alone, we first subtracted the gene expression profiles of control cells, which were T cells that had been cultured alone that were either uninfected (Mock T) or exposed only to virus (HIV T).
A scatter plot (Fig. 7 A), representing the common (genes that fall on the diagonal) and differentially expressed genes (genes that fall off the diagonal) from this comparison, highlighted the significant differences in gene expression between latently infected cells and controls (r = 0.77). Additionally, this plot showed that several of the genes that discriminate latently infected cells from control cells were genes downstream of type I interferons, including interferon-induced protein with tetratricopeptide repeats 1 (IFIT-1), interferon alpha-inducible protein 27 (IFI27), and 2′-5′-oligoadenylate synthetase 1 (OAS1).
Heatmaps of the top 100 genes (Fig. S3) confirmed the de novo induction of genes encompassing several biological and metabolic pathways in T cells exposed both to DC and virus. These included transcripts of the Interferon pathway, genes involved in the regulation of cell cycle entry and mitosis, as well as receptor and effector molecules of cell survival and apoptosis.
Network analysis (Fig. 7 B and C) showed that two major pathways were regulated in T cells following co-culture with DC exposed to HIV. Exposure of CD4+ T cells to HIV and DC led to the up-regulation of genes downstream of type I interferon by nucleotide sensors. Figure 7 B confirms the wide-ranging impact of the up-regulation of the type I Interferon pathway, as several molecules with antiviral activity were up-regulated, including ISG15 ubiquitin-like modifier (ISG15), and DEAD box polypeptide 58 (DDX58; also known as RIG-I). Genes involved in actin polymerisation, and the organisation of microtubules, were also up-regulated as a consequence of interferon pathway up-regulation. Additionally, the interferon pathway intercepted with the mammalian target of rapamycin complex 2 (mTORC2) pathway, which plays an important role in autophagy and T cell survival [24] (Fig. 7 B; Table S1).
Network analysis confirmed the negative impact of exposure of CD4+ T cells to DC and virus on the NF-κB pathway as well as several cellular metabolic pathways (fatty oxidation and glucose metabolism) regulated by peroxisome proliferator-activated receptor gamma (PPARG; Fig. 7 C). Inhibition of the NF-κB transcriptional network, which plays a significant role in HIV transcription, led to the down-regulation of protein kinase C alpha (PRKCA). This observation confirms the quiescence of these cells and may also be a step towards the induction of HIV latency. Triggering of a transcriptional program leading to T cell quiescence was confirmed by the increased expression of Kruppel-like factor 6 (KLF6), a gene with anti-proliferative functions [25], [26], as well as activating transcription factor 3 (ATF3) that has been recently been shown to negatively regulate activating protein 1 (AP-1)-mediated HIV transcription. Additionally, we observed a down-regulation of several molecules involved in DNA replication such as members of the minichromosome maintenance (MCM) complex (MCM4, MCM5, and MCM10) and the aurora kinase (Fig. 7 C). Pathways controlling pyrimidine and purine synthesis were also expressed at lower levels in cells exposed to virus and DC highlighting a reduced availability of nucleotides for cell division (Table S1). The down-regulation of NF-κB resulted in the decreased expression of several molecules that play a critical role in T cell survival, including CD27 (TNFRSF7), baculoviral IAP repeat containing 5 (BIRC5/survivin) and tumor necrosis factor receptor superfamily, member 6b (TNFRSF6B/DCR3), a decoy receptor that inhibits Fas ligand and LIGHT-mediated signalling [27]. In order to confirm the differential expression of genes in these different populations of cells, we used a highly quantitative PCR approach and showed a strong correlation between gene expression data measured by either gene array or PCR (Fig. 7 D and Table S2).
Taken together, results of transcriptional profiling highlighted the impact of two major transcriptional nodes in the inhibition of viral replication and the induction of latency. The up-regulation of Type I Interferons and their downstream target genes could trigger several genes endowed with antiviral activities and would also impact cell proliferation, survival and metabolic processes. Concomitantly, the down-regulation of NF-κB will lead to T cell quiescence and decreased levels of activation, both of which are required for HIV replication.
The study of latently infected resting CD4+ T cells ex vivo from HIV-infected patients on ART is greatly limited by the low frequency of latently infected cells and the lack of a distinctive surface marker to distinguish latently infected from uninfected cells. Here we demonstrate that latency can be efficiently established via direct infection of non-proliferating CD4+ T cells in the presence of DC. Using primary blood DC and resting CD4+ T cells we have demonstrated that: [1] co-culture of resting memory CD4+ T cells with DC can establish latent infection; [2] mDC but not pDC mediate this effect; [3] close cell-cell proximity is required between DC and T cells; and [4] multiple cell cycle genes were altered in non-proliferating CD4+ T cells, containing latently infected cells. These novel findings provide a potential pathway for the establishment and maintenance of latent infection in resting CD4+ T cells that recapitulates the likely events within lymphoid tissues in HIV-infected patients in vivo.
Previous studies have explored the ability of DC to enhance productive HIV infection within DC-CD4+ T cell co-cultures [28]–[31]; however, we are the first to present data demonstrating the ability of specific subpopulations of DC to induce latency in resting CD4+ T cells in these co-cultures. Using this model we clearly demonstrated that following co-culture of mDC with resting memory CD4+ T cells, post-integration latency was established. This was demonstrated by inducible virus (established both in the presence and absence of indinavir) and detectable integrated HIV DNA in T cells cultured with mDC but not those cultured alone following infection with an R5 EGFP reporter virus. While integrated R5 HIV DNA was only detected following co-culture with mDC in our model of latency, it was similar to that previously reported for resting CD4+ T cells infected in isolation with a wild type X4 NL4.3 virus [32]. Resting CD4+ T cells express very low levels of CCR5 in contrast to expressing very high levels of CXCR4. Additionally, we saw significantly higher levels of latency (∼100 fold) in our T cells cultured alone when we used an X4 EGFP reporter virus compared to an R5 EGFP virus.
Unlike memory CD4+ T cells, we were unable to detect latency in naïve CD4+ T cells following mDC co-culture. It is possible that the differential establishment of latency in resting naïve and memory T cells was due to differences in their cortical actin density and actin dynamics as previously suggested by others [33]. While mDC induced T cell latency in this model, pDC did not. Interestingly, pDC played an active inhibitory role in establishing latency, when co-cultures were performed with equal numbers of pDC and mDC (Fig. 4 C). One potential explanation for the difference between co-cultures of bulk DC and T cells (where latency was established) and DC-T cell co-cultures containing purified equal numbers of pDC and mDC (where latency was inhibited) could potentially be that the number of pDC present in the bulk DC (roughly 1 pDC to 3 mDC) was too low to inhibit latency. How pDC actively suppress the establishment of latency is unknown, but it does not appear to be mediated by IFN-alpha.
Establishment of mDC-induced latency was not dependent on DC-T cell transfer of HIV, as latency was still detected when T cells were infected in isolation and uninfected mDC added only after virus had been washed off. Nor was it dependent on the amount of virus replication, because while only mDC were able to induce latent infection, similar levels of productive infection were observed in both the pDC and mDC co-cultured CD4+ T cells (Fig. 2 A). Furthermore, we found that addition of SEB to the culture model enhanced productive infection but did not increase latent infection (Fig. 2 G). Together, these data provide evidence that the establishment of latency in the non-proliferating CD4+ T cells when co-cultured with mDC was not simply due to higher viral exposure in these cultures.
These results differ from a previous study that also looked at direct infection of resting CD4+ T cells, which concluded that DC had no effect on the integration levels of R5 or X4 virus in either naïve or memory CD4+ T cells [34]. However, in this study, although primary DC were used (defined as BDCA-1+ and BDCA-4+ cells), total DC were present at a frequency of only 0.89% and therefore the frequency of mDC may have been too low to demonstrate an effect of mDC on the infection of resting CD4+ T cells. Additionally, contrary to our data, a recent study has reported the ability of monocyte derived DC (MDDC) to activate latent infection in T cells [35]. A key difference was that latency in this study was unusually established in proliferating CD4+ T cells and not non-proliferating T cells as in our study. Furthermore, MDDC, as opposed to primary DC, were utilised in this study. MDDC have multiple significant functional and lineage differences to primary DC as we have recently demonstrated using detailed sorting and gene expression analyses [36].
In this study we utilised total blood CD11c+ mDC, which consist of at least three different subsets, a major SLAN (6-sulfo LacNAc+), an intermediate CD1c+ (BDCA-1) and a minor CD141+ (BDCA-3) population, each with different functional properties [37]–[40]. However, it is currently unclear whether one or more of these subsets is responsible for inducing latency in resting CD4+ T cells.
We have previously demonstrated that multiple chemokines, including CCL19, CXCL9 and CXCL10, can condition resting CD4+ T cells allowing for the establishment of HIV latency [14], [17]. However, blocking CCL19 and CXCR3, the receptor for CXCL9, 10 and 11, had a minimal impact on DC-induced latency (Fig. 4 D). While it is possible that there may be involvement of chemokines other than those inhibited, given that latency was not detected in resting CD4+ T cells infected in the presence of infected mDC-T cell culture supernatants, this is unlikely (Fig. 4 F). Rather, our data supports an essential role for direct DC-T cell interactions or DC-T cell signalling as mDC-induced latency was prevented when the mDC were cultured in transwells above the resting CD4+ T cells (Fig. 4 E). Unlike our previous work that was performed in the absence of productive infection [14], latency following DC-T cell co-culture was established in the presence of productive infection, which may more accurately mimic the establishment of latency in acute infection in vivo. Therefore, in the presence of productive infection it is possible that there are alternative pathways that lead to the establishment and maintenance of latency in resting CD4+ T cells.
Interactions between ICAM-1, found on DC, and LFA-1, found on T cells, strengthen DC-T cell adhesion and play a key role in the formation of the immunological synapse [41]. We have shown that clustering, facilitated by ICAM-1-LFA-1 interactions, contributed to DC-induced T cell latency, as latency was significantly reduced, but not eliminated, when blocking antibodies to CD18/LFA-1 were added to the DC-T cell co-cultures (Fig. 5 A). However, interactions between ICAM-1 and LFA-1 alone were not sufficient to induce T cell latency in the absence of mDC (Fig. 5 B), therefore, the reduction in latency observed in the presence of anti-CD18 was most likely due to the reduction in DC-T cell clustering rather than specific LFA-ICAM signalling events. As there are numerous other molecules involved in DC-T cell clustering, such as LFA-3 and CD2, additional signalling pathways should be explored as potential mediators of DC-induced HIV latency. Interestingly, a recent paper has demonstrated enrichment of latency in CD2 expressing T cells from HIV-infected patients on ART [42].
Transcriptional profiling experiments served to highlight changes in cellular gene expression in resting non-proliferating CD4+ T cells that contained latently infected CD4+ T cells. We showed significant differences in gene expression between resting CD4+ T cells from HIV and mock infected DC-T cell cultures. However, it is important to note that, while all cells within our “latent” cell population were exposed to virus, only a proportion were actually infected (median of 3%). It is possible that some of the observed differences in gene expression may be due to uninfected cells that were exposed to HIV but not infected. This may include the genes downstream of type I interferons as our in vitro experiments have shown that pDC, the major producers of type I interferons, were not involved in the induction of T cell latency. Therefore, it is possible that type I interferons are necessary but alone are not sufficient to induce HIV latency. We have demonstrated significant differential expression of genes involved in cell cycle, in particular those associated with cell cycle arrest (Fig. 7 C). During DC-T cell interactions in the presence of HIV, differential expression of co-stimulatory and negative regulatory factors determines the fate of the interacting CD4+ T cell [43]. These interactions can result in active suppression of T cell cycle and as a result may inhibit post-integration steps in viral replication and promote the establishment of latency. Indeed, in HIV-infected patients on ART, HIV DNA is found at higher frequencies in CD4+ T cells expressing the negative regulator PD-1 [44].
Latency has also been shown to be triggered by the absence of certain transcriptional machinery in resting CD4+ T cells, such as NF-κB [45] and nuclear factor of activated T (NFAT) [11], [46]. In DC-induced latently infected CD4+ T cells we observed the suppression of multiple genes associated with the activation of NF-κB (Table S1), including protein kinase C alpha, PRKCA, which also plays a role in the activation of NFAT [47], [48]. Therefore, it is possible that the global suppression of genes associated with the activation of NF-κB and/or NFAT may also contribute to the maintenance of latency in DC- T cell co-cultures by preventing progression to productive infection in cells that contain integrated HIV. However, while this data provides insights into genes that may potentially be important for both the establishment and maintenance of latency, it will be important to conduct gene knockdown experiments within our model in order to determine the specific role of individual genes in establishing and maintaining mDC-induced T cell latency.
In summary, this study has demonstrated a novel pathway for the establishment of latency in resting memory CD4+ T cells that was dependent on close proximity to mDC. Efficient infection of resting CD4+ T cells in close contact with mDC and HIV could explain the rapid early establishment of the latent HIV reservoir. Additionally, if infectious virus persists in tissues such as lymph node in patients on ART, mDC may facilitate ongoing infection of resting T cells leading to replenishment of the reservoir.
PBMC were isolated from buffy coats obtained from the Australian Red Cross Blood Service (Melbourne, Australia). Resting CD4+ T cells were negatively selected using magnetic cell sorting and a cocktail of antibodies to CD8, CD11b, CD16, HLA-DR, CD19 and CD69, as previously described [17], [49]. Sorted cells were routinely negative for CD69, CD25 and HLA-DR (Fig. S1 A). In some experiments bulk resting CD4+ T cells were further sorted into CD45RA+ naïve and CD45RA− memory CD4+ T cells using phycoerythrin (PE)-labelled antibody to CD45RA and a FACSAria (BD Biosciences). DC were isolated from blood as previously described [50]. Briefly, DC were enriched using magnetic bead depletion and antibodies to CD3, CD11b and CD19. Enriched cells were then sorted using a FACSAria (BD Biosciences) to obtain a bulk cocktail− HLA-DR+ DC population, HLA-DR+CD11c+ mDC or HLA-DR+CD123+ pDC. The purity of sorted cells was always >98% (Fig. S1 B).
In all experiments except where noted we used an NL4-3 virus with EGFP inserted into the nef open reading frame at amino acid position 75 at the aKpnI (Acc651) site with a CCR5-tropic (AD8) envelope (NL(AD8)-nef/EGFP), alternatively we used this virus with a CXCR4-tropic (NL4-3) envelope (NL4-3-nef/EGFP; both kindly provided by Damian Purcell, University of Melbourne, Melbourne, Australia). In one set of experiments we used a Nef-competent EGFP reporter virus, kindly provided by Yasuko Tsunetsugu-Yokota (National Institute of Infectious Diseases, Tokyo, Japan) [51]. HIV stocks were generated by FuGene (Promega, Madison, WI) transfection of 293T cells as previously described [49], [50]. Cells were infected at 37°C for 2 hours at an MOI of 0.5 or 5, as determined by limiting dilution using the Reed and Muench method [52], followed by a wash step to remove unbound virus.
Resting CD4+ T cells were labelled with proliferation dye, either SNARF (10 µM; Invitrogen) or eFluor®670 (5 µM; eBiosciences, San Diego, CA), according to the manufacturer's instructions. SNARF/eFluor670-labelled resting CD4+ T cells were cultured in media supplemented with IL-2 (2 U/mL; Roche Diagnostics) for 24 hours, with or without syngeneic bulk DC or sorted DC subsets (DC: T cell ratio of 1∶10), in the presence or absence of SEB (10 ng/mL; Sigma). Cells were then infected using an EGFP-reporter virus and cultured for a further 5 days (Fig. 1 A). In some experiments, cells were cultured with and without the protease inhibitor Indinavir (0.1 µM final) for 30 minutes at 37°C prior to infection. At day 5 post-infection, cells were analysed by flow cytometry for productive infection by detecting EGFP+ cells. Subsequently, the non-proliferating (SNARFhi/eFluor670hi) CD4+ T cells that were not productively infected (EGFP−) were sorted using a FACSAria (BD Biosciences). In order to amplify any latent infection, the sorted CD4+ T cells were stimulated with PHA (10 ug/mL)/IL-2 (10 U/mL) in the presence of PBMC and cultured for a further 5 days. The number of EGFP+ cells following re-stimulation was used as a surrogate measure for the number of latently infected, non-proliferating CD4+ T cells in the original cultures (Fig. 1 B). In some experiments, we also stimulated the sorted SNARFhi/eFluor670hiEGFP− CD4+ T cells directly with plate bound anti-CD3 (Beckman Coulter; 5 µg/mL) and soluble anti-CD28 (BD; 5 µg/mL). Flat bottomed 96 well plates were coated with anti-CD3 (5 µg/mL) for 3 hours at 37°C. Unbound antibody was then removed and 1×105 SNARFhi/eFluor670hiEGFP− CD4+ T cells were plated per well in 200 µL of media containing soluble anti-CD28 (5 µg/mL), IL-7 (50 ng/mL) and the integrase inhibitor, L8 (1 µM final). The number of EGFP+ cells was determined following 72 hours of culture. As a control for the activity of the integrase inhibitor L8, we cultured SEB-stimulated PBMC with and without L8 for 30 minutes prior to infection, and productive infection was determined at day 5 post-infection (Fig. S2).
For phenotypic analysis of the CD4+ T cells before culture, we stained sorted resting CD4+ T cells with CD69-FITC, CD25-PE and HLA-DR-perCP (BD Bioscience) on ice for 25 minutes. To determine whether co-culture with DC had altered the activation state of the resting CD4+ T cells, in some experiments the sorted SNARFhiEGFP− CD4+ T cells were labelled with either CD69-FITC or HLA-DR-FITC (BD Biosciences). Intracellular staining was also performed on the sorted SNARFhiEGFP− CD4+ T cells to detect expression of the cell cycle marker Ki67. Cells were permeabilised with 500 µL of 1× FACS Permeabilising Buffer (BD Biosciences) in the dark at room temperature for 10 minutes, washed once with FACS wash and incubated with Ki67-FITC (5 µL/105 sorted CD4+ T cells; Dako) for 45 minutes on ice. Following incubation, cells were washed twice with FACS wash and resuspended in 1% FACS fix. We performed analyses on a FACSCalibur (BD Biosciences) and results were analysed using Weasel software (Walter and Elisa Hall Institute, Melbourne, Australia).
Cytokine bead arrays (eBioscience) were used according to the manufacturer's directions to determine the concentration of IL-1-beta, IL-6, IL-10, IL-12p70, TNF-alpha, CXCL9 and CXCL10 in the cell cultures.
In some experiments, nAbs to CD18 (10 µg/mL (prior to infection) and 20 µg/mL (post-infection); clone 7E4; Beckman Coulter), CCL19 (25 µg/mL), CXCR3 (20 µg/ml), IFN-alpha (5 µg/mL) or control IgG (R&D Systems, Minneapolis, MN); IL-6 or IgG1 (10 ug/mL; BioLegend, San Diego, CA); IL-10R or IgG2 (10 ug/mL; Biolegend) were used. In these experiments, both the DC and the resting CD4+ T cells were pre-incubated with nAbs for 15 minutes on ice prior to culture. The nAbs were added again to the co-cultures following infection. Neutralising activity of anti-CCL19 (25 µg/mL) and anti-CXCR3 (20 µg/ml) was confirmed using a chemokine-induced migration assay. Resting CD4+ T cells were added to the top chamber of a 3 µM pore transwell migration plate (Sigma) and either CCL19 (100 nM) or CXCL10 (300 nM) was added to the bottom chamber. In experiments using anti-CXCR3, cells were treated with nAb for 15 minutes at 37°C and washed off prior to chemokine treatment. In comparison, in experiments using anti-CCL19, nAb was added together with chemokine to the bottom chamber. Migrated cells in the bottom chamber were then counted in duplicate at 20 hours post addition of chemokine. Anti-IL-6 and anti-IL-10R were used at neutralising concentrations previously described [53], [54]. As positive controls for these nAbs we demonstrated that 10 µg/mL of anti-IL-6 and anti-IL10R or 5 µg/mL of anti-IFN-alpha efficiently blocked IL-6 (100 ng/mL), IL-10 (50 ng/mL) or IFN-alpha (50 ng/mL) mediated STAT3 phosphorylation respectively (Fig. S2).
In order to determine the role of ICAM-1 and LFA-1 interactions in mDC-induced latency, resting CD4+ T cells were cultured alone or with 10 ug/mL of ICAM-1fc together with 6 µg/mL of anti-IgG-fc (both from R&D Systems) for 24 hours prior to infection and maintained post-infection.
DC were cultured with resting CD4+ T cells in the presence and absence of 0.4 µm cell culture inserts (BD, Franklin Lakes, NJ) with DC in the top chamber and resting CD4+ T cells in the lower chamber. Following 24 hours of culture, both the DC and the CD4+ T cells were infected as described above. In other experiments, we added supernatant from infected mDC-T cell co-cultures to uninfected resting CD4+ T cells and then infected these cells. Media changes were performed daily using supernatant from infected mDC-T cell co-cultures. To determine the role of DC-T cell transfer, resting CD4+ T cells were infected in the absence of mDC and uninfected mDC were added back to the T cells only after virus had been washed off.
SNARFhiEGFP− CD4+ T cells cultured with DC, in the presence (latently infected) or absence (mock-infected) of HIV, were sorted 5 days following infection with NL(AD8)-nef/EGFP. In these experiments, all culture media was supplemented with 10 ng/mL of IL-7 (Sigma) instead of IL-2, in order to increase cell survival of resting cells, and infections were performed at an MOI of 5 to ensure high numbers of latently infected cells.
Microarrays were performed as previously described [55]. Briefly, cells were lysed and RNA extracted (Qiagen, Valencia, CA), amplified (Ambion Applied Biosystems, Austin, TX) and hybridised to an Illumina Human-Ref8 (v3) BeadChip (Illumina, San Diego, CA). Beadchips were scanned using an Illumina BeadStation 500GX scanner and Illumina BeadStudio (version 3) software (Ilumina). Illumina probe data was exported from BeadStudio as raw data and screened for quality. Samples failing chip visual inspection and control examination were removed. Gene expression data was analysed using Bioconductor (http://bioconductor.org/) [56], an open-source software library for the analyses of genomic data based on R, a language and environment for statistical computing and graphics (www.r-project.org). The R software package was used for pre-processing, first to filter out genes with intensities below background in all samples, then to minimum-replace (a surrogate-replacement policy) values below background using the mean background value of the built-in Illumina probe controls as an alternative to background subtraction (which may introduce negative values) to reduce “over inflated” expression ratios determined in subsequent steps, and finally quantile-normalise the gene probes intensities. Genes were then filtered by intensity and by variance filters to allow a reduction in the number of tests and a corresponding increase in power of the differential gene expression analysis. The resulting matrix showing filtered genes as rows and samples as columns was log2 transformed and used as input for linear modelling using Bioconductor's limma package, which estimates the fold-change between two predefined groups by fitting a linear model and using an empirical Bayes method to moderate standard errors of the estimated log-fold changes for expression values from each gene. P values from the resulting comparison were adjusted for multiple testing according to the method of Benjamini and Hochberg [57]. This method controls the false discovery rate, which was set to 0.05 in this analysis. Microarray data is available through the National Center for Biotechnology Information Gene Expression Omnibus (GEO), series accession number pending.
Ingenuity Pathway Analysis (IPA) software (Ingenuity Systems, www.ingenuity.com) was used to identify canonical signalling pathways and networks associated with the expression profiles of the non-proliferating CD4+ T cells cultured with DC in the presence (HIV T (+DC)) or absence (Mock T (+DC)) of HIV. Differentially expressed Illumina Probe IDs were imported into the Ingenuity software and mapped to the Gene Symbol from Ingenuity knowledge database. The significance of the association between the dataset and the canonical pathway was measured in two ways: 1) A ratio of the number of genes from the dataset that map to the pathway divided by the total number of genes that map to the canonical pathway; 2) Over-representation Fisher's exact test was used to calculate a p-value determining the probability that the association between the genes in the dataset and the canonical pathway is explained by chance alone. The pathways were ranked by −log p-value. This score was used as the cut-off for identifying significant canonical pathways (p value<0.05). IPA's networks are built from direct or indirect physical, transcriptional, and enzymatic interactions between the mapped genes (focus genes). Two genes are considered to be connected if there is a path in the network between them. Ingenuity's approach is based on a multi-stage, heuristic algorithm that iteratively constructs networks that greedily optimize for both interconnectivity and number of Focus Genes under the constraint of a maximal network size. Each individual IPA network has a maximum of 35 focus genes and is assigned a significance score (based on P value) representing the likelihood that the focus genes within the network are found there by random chance.
As previously described, full length viral DNA was quantified using primers specific for the HIV-1 long terminal repeat (LTR) and Gag [58], and integrated HIV-1 DNA was quantified using a nested Alu-LTR real-time PCR [15], [59], [60]. Results were normalised for total input DNA as determined by real-time PCR for the CCR5 gene [61]. The correlation between gene arrays and real-time PCR was performed using a Spearman correlation test.
Microarray expression data were validated in two donors by reverse transcriptase real-time PCR (RT-qPCR), as previously described [62]. Briefly, SNARFhiEGFP− CD4+ T cells were lysed for RNA extraction and DNAse treatment (Qiagen, RNAeasy mini kit). cDNA was generated using CellsDirect qRT-PCR mix (Invitrogen). After reverse transcription all target genes were pre-amplified (18 cycles) using Taqman primers (Roche Probe library) specific for the transcripts of interest, which were also used for quantification. qPCR were performed on a Roche Light Cycler 348II and analysed according to the ΔΔct method.
In all experiments, Wilcoxon signed-rank or student paired t tests (for n<5) were performed for comparisons between populations using Graphpad Prism 5.0 software. P values of less than 0.05 were considered significant. Statistical analyses for microarray data were performed with program R, according to the method of Benjamini and Hochberg [57].
|
10.1371/journal.ppat.1004908 | A Single Protein S-acyl Transferase Acts through Diverse Substrates to Determine Cryptococcal Morphology, Stress Tolerance, and Pathogenic Outcome | Cryptococcus neoformans is an opportunistic yeast that kills over 625,000 people yearly through lethal meningitis. Host phagocytes serve as the first line of defense against this pathogen, but fungal engulfment and subsequent intracellular proliferation also correlate with poor patient outcome. Defining the interactions of this facultative intracellular pathogen with host phagocytes is key to understanding the latter’s opposing roles in infection and how they contribute to fungal latency, dissemination, and virulence. We used high-content imaging and a human monocytic cell line to screen 1,201 fungal mutants for strains with altered host interactions and identified multiple genes that influence fungal adherence and phagocytosis. One of these genes was PFA4, which encodes a protein S-acyl transferase (PAT), one of a family of DHHC domain-containing proteins that catalyzes lipid modification of proteins. Deletion of PFA4 caused dramatic defects in cryptococcal morphology, stress tolerance, and virulence. Bioorthogonal palmitoylome-profiling identified Pfa4-specific protein substrates involved in cell wall synthesis, signal transduction, and membrane trafficking responsible for these phenotypic alterations. We demonstrate that a single PAT is responsible for the modification of a subset of proteins that are critical in cryptococcal pathogenesis. Since several of these palmitoylated substrates are conserved in other pathogenic fungi, protein palmitoylation represents a potential avenue for new antifungal therapeutics.
| Cryptococcus neoformans is a ubiquitous environmental yeast that kills over 625,000 people annually, mainly in developing countries. Healthy humans frequently inhale infectious particles without noticeable symptoms. However, in immunocompromised people, the initial lung infection can spread to other sites, particularly to the central nervous system where it causes lethal brain infection. The infected host responds by deploying immune cells to engulf and kill the yeast, but C. neoformans can survive this engulfment and even multiply within the host cells. To understand the interactions between the invading microbe and host cells we screened 1,201 fungal mutants to identify fungal factors that influence these processes. One mutant, lacking an enzyme that modifies proteins with the lipid palmitate, showed an increase in engulfment by the host along with dramatic defects in morphology, stress resistance, and virulence. We went on to identify the proteins this enzyme modifies and explain how its absence leads to altered cell wall synthesis, signal transduction, and membrane trafficking; these changes explain the behavior of the mutant. We also found that the mutant could not cause disease in an animal model. Our work shows that protein palmitoylation is critical for cryptococcal pathogenesis and presents a potential avenue for antifungal therapy.
| Cryptococcus neoformans is a fungal pathogen that causes over 625,000 deaths per year, mainly in severely immunocompromised individuals. Cryptococcosis is contracted by inhalation of infectious particles from the environment [1], which leads to a primary pulmonary infection. In healthy people this infection is typically cleared, but in immunocompromised hosts the organism can proliferate and disseminate, with a tropism for the central nervous system where it causes lethal meningoencephalitis. As a result, this pathogen is a major threat to AIDS patients and to the rapidly growing population of individuals with other immunosuppressive conditions [2–5]. Host phagocytes, mainly macrophages, are critical for initial control of this facultative intracellular pathogen [6]. However, as the flip side to their positive role as the first line of host defense, these cells may also serve as sites for replication and latency, or potentially as vehicles for yeast dissemination [1]. In line with these activities, several studies have demonstrated a correlation between poor patient outcomes and the capacity of clinical strains to be phagocytosed and/or to proliferate intracellularly [7, 8]. Understanding the opposing roles of macrophages in cryptococcal infection and their interactions with C. neoformans is key to our ability to influence such events in favor of the host. Despite the importance of these interactions to cryptococcal pathogenesis, the critical features of the host and fungus that govern them have not been determined.
We developed an image-based high-throughput screening (HTS) assay to probe fungal-host cell interactions [9] and evaluated a C. neoformans partial deletion collection [10] for altered engulfment by a human macrophage-like cell line. One ‘hit’ lacked a gene that encodes a protein S-acyltransferase (PAT), incriminating protein palmitoylation as a key pathway in cryptococcal pathogenesis. Protein palmitoylation, the reversible addition of palmitate to cysteine, can regulate the stability, localization, and function of target proteins [11]. The enzymes mediating this modification were first identified in the model yeast S. cerevisiae [12, 13] and are now recognized as important effectors in eukaryotic cells [11]. Although protein palmitoylation has been shown to influence infectivity in viruses [14], bacteria [15], and parasites [16–18], its role in fungal pathogenesis has not been explored. The importance of this lipid modification in fungal pathogens is supported by studies of Ras1 localization in Aspergillus fumigatus, C. neoformans, and Candida albicans [19–21], but no other proteins have been shown to be functionally palmitoylated in these organisms. Finally, no PAT has been characterized in a pathogenic fungus. Our studies demonstrate that a single PAT is a major determinant of cryptococcal pathogenesis and, by defining the relevant palmitoylome, we identify the cellular mechanisms by which defects of this fatty acid modification dramatically alter fungal morphology, host interactions, and virulence in vivo.
C. neoformans engulfment by host cells and subsequent intracellular proliferation has been implicated in dissemination, virulence, and ultimately in patient outcome [8, 22, 23]. However, the full complement of fungal genes that participate in these processes has not been defined, and how individual gene products modulate interactions with host phagocytes is not known. To address cryptococcal interactions with host cells, we used an automated high content imaging method [9] to quantify the interactions between a human monocytic cell line (THP-1) and mutant fungi from a deletion collection made in the highly pathogenic reference strain H99 [10]. Of the 1,201 mutants we screened, 56 (4.7%) showed significant alterations in phagocytic index (Fig 1A). These mutants (S1 Table) were roughly equally distributed between strains with decreased and increased engulfment (30 and 26, respectively); the ten most extreme in each category are shown in Table 1. An example data set from one plate of the mutant collection (Fig 1B) shows strikingly increased phagocytosis of three mutants, two of which, pka1 and rim101, are known to have altered cell surface structures that would explain this phenotype [24, 25]. Additionally, pbx1, the top hit of the high uptake mutants (Table 1), has defects in cell wall structure and capsule assembly that cause increased engulfment by macrophages [26]. These observations validated our strategy for probing the interactions between C. neoformans and macrophages and encouraged us to further pursue novel hits from our screen.
Another strain (2A12) that consistently demonstrated an elevated phagocytic index (Fig 1B and Table 2) lacks the uncharacterized gene CNAG_03981. This gene is highly homologous to S. cerevisiae PFA4, which encodes a palmitoyl acyltransferase (PAT), and was accordingly given the same name (following guidelines in [27]). PATs are DHHC zinc finger domain-containing enzymes that mediate the reversible addition of palmitate to proteins, thereby regulating their membrane association and biological function [11]. Eukaryotic cells often express multiple DHHC domain proteins, which have similar enzymatic activity but modify variably overlapping groups of substrates [28]. These enzymes play key roles in protein fatty-acylation and membrane targeting [11], but have never been studied in C. neoformans or any other fungal pathogen. There are seven putative PATs encoded in the H99 genome; four of these were deleted in the collection that we screened but only pfa4Δ differed significantly from wild-type cells (Table 2). This suggested that Pfa4 acylates at least one protein that both influences host cell interactions and is not modified by other PATs. Given the limited knowledge of protein palmitoylation in C. neoformans biology and pathogenesis, we chose this mutant for mechanistic study.
We first generated independent pfa4 deletions in C. neoformans reference strain H99 (used for the deletion collection) and its more genetically tractable derivative KN99 [29]. Like 2A12, both mutants showed consistent increases in adherence to and engulfment by macrophages compared to wild-type cells (Fig 1C), with the greater uptake readily visible by confocal microscopy (Fig 2A and S1 and S2 Videos). These phenotypes, which were independent of the method used to label the cells (S1A and S1B Fig), were all reversed by complementation of the mutant with the wild-type gene at the endogenous locus.
The extremely high numbers of internalized mutant cells (Figs 1C and 2A) could potentially alter intracellular trafficking of C. neoformans, which is usually delivered to lysosomes after endocytosis [30]. To test this we used confocal microscopy to assess the progression of pfa4Δ and wild-type cells through various intracellular compartments after their exposure to host phagocytes (S2A Fig). The distribution of wild-type and mutant fungi between the cell surface (adherent cells), early endosomes (marked with EEA1), and lysosomes (marked with LAMP-1) was similar at late time points. The only significant differences were observed soon (15 min) after assay initiation, when a greater fraction of wild-type cells remained surface-accessible (adherent) while more mutant cells had already been phagocytosed (although not yet associated with EEA1). Overall, although the mutant is more efficiently internalized, both strains reach EAA1 and LAMP-1 compartments with similar dynamics. It has recently been suggested that C. neoformans-containing lysosomes do not completely acidify [31]. To test whether acidification differed between lysosomes containing pfa4Δ and wild-type yeast, we performed a phagocytosis assay in the presence of Lysotracker Red, a dye that becomes trapped and fluorescent in acidified organelles. We found that both strains were similarly distributed between unstained phagosomes and lysosomes (positive for Lysotracker; S2B and S2C Fig)
In addition to an increased number of internalized pfa4Δ cells, our confocal studies revealed an unusual and dramatic change in their morphology (Fig 2A and S1 and S2 Videos). While wild-type cells are spherical, the mutant cells appeared to have collapsed in on themselves, manifested as membrane staining in either crescent shapes or double rings depending on cell orientation. This aberrant morphology occurred whether the fungi were inside macrophages (Fig 2A) or grown independently (Fig 2B), indicating that the alteration is intrinsic to the mutant rather than induced by the host cells. We tested other dyes to rule out the possibility that the shape change was due to the Lucifer Yellow (LY) stain used in our phagocytosis studies; in all cases we observed a similar phenotype (Fig 2B). Next, to eliminate the possibility that any compound that binds cell wall structures induces cell collapse, we imaged actively growing, unstained wild-type and pfa4Δ mutant cells by brightfield and differential interference contrast (DIC) light microscopy. Under these unstained, actively growing conditions we could easily detect the same aberrant shapes seen in pfa4Δ cells stained with various dyes (S3 Fig), indicating that they represent an intrinsic feature of this mutant. Finally, to get a detailed view of this morphological defect we examined the cells by scanning electron microscopy. Consistent with our light microscopy results, wild-type cells were globular and smooth while pfa4Δ cells were dramatically deformed (Fig 2C). Surprisingly, this has little effect on their ability to replicate at 30°C, where their growth rate is close to that of wild-type cells.
The pfa4Δ mutant showed altered initial interactions with host cells and aberrant morphology. One model that explains both observations is that the mutant has fundamental defects in cell wall structure that alter both surface molecule exposure and cell wall integrity. To probe cell wall organization, we used dye and lectin binding with flow cytometry to assess the accessibility of various cell wall components (Fig 3). We found that chitin accessibility, probed with calcofluor white (CFW), was not significantly altered in pfa4Δ, unlike the decreased signal in a chitin synthase mutant (chs3Δ) included as a control (Fig 3B). In contrast, probes of chitosan (Eosin Y; EoY) and mannans (Concanavalin A lectin; ConA) showed that these glycans were much more accessible in the pfa4Δ mutant (Fig 3B), supporting aberrant wall structure; this was also reflected in an altered staining pattern for ConA (S4 Fig). Similarly, LY and pontamine (Pont), also dyes that bind cell wall (although their specific targets are not defined), showed clear changes in binding the mutant compared to controls (Fig 3B). The abnormal exposure of chitosan and mannans at the surface of pfa4Δ cells could explain their greater recognition by macrophages (see Discussion).
We reasoned that the altered arrangement of cell wall components in the pfa4Δ mutant would threaten overall cell integrity. We tested this hypothesis by plating serial dilutions of pfa4Δ in the presence of various stressors. Compared to wild-type and the complemented mutant, pfa4Δ was sensitive to plasma membrane damaging agents (SDS and H2O2), osmotic stress (KCl and NaCl), cell wall binding dyes (CFW, CR, and LY), and elevated temperature (37°C) (Figs 4A and S5). Only temperature sensitivity could be rescued by sorbitol (Fig 4A), suggesting that the cell integrity defects and temperature sensitivity are caused by perturbation of different pathways. This experiment also indicates that Pfa4 is not absolutely required for growth at high temperatures; in support of this conclusion, the pfa4Δ cells continued to grow slowly at 37°C for over a day even in the absence of sorbitol (S5A Fig). The mutant was also hypersensitive to treatment with cell wall lysing enzymes (S5C Fig), an assay which probes cell wall stability as well as cellular response to cell wall damage [32]. In all cases genomic or plasmid complementation of pfa4Δ restored wild-type phenotypes.
The pleiotropic effects of PFA4 deletion suggested the dysfunction of one or more protein substrates of palmitoylation, which are not lipidated and therefore mislocalized, misfolded and/or degraded. To test whether the enzymatic activity of Pfa4 was indeed responsible for these phenotypes, we mutated its catalytic DHHC sequence to DHAS (S5D Fig); mutation of this cysteine abolishes PAT activity in other systems [12, 13, 16]. When both forms of the protein were expressed in pfa4Δ, only the wild-type rescued the mutant’s sensitivity to cell wall stress (S5E Fig), showing that the observed defects are due to a lack of PAT enzymatic activity.
The inability of pfa4Δ to withstand cell wall stress could reflect defects in cell wall structure, inability to respond to and repair a damaged wall, or both. To investigate cell wall structure we used transmission electron microscopy (TEM). The walls of wild-type strains and of pfa4Δ expressing wild-type PFA4 were fairly uniform in thickness, and showed the expected multilayered organization [33]: an electron-dense inner layer surrounded by a more electron-lucent layer and then an outer rim of capsule (Fig 4B; the capsule layer is thin because the strains were grown in rich medium). In these cells the inner layer was always ≥50% of the total wall width (example shown in Fig 4C, top image). In contrast, the cell walls of the mutant (with or without the catalytically-dead Pfa4AS) were generally thinner, primarily due to a reduction in the inner layer (Fig 4B and 4C, middle image). In ~80% of these cells the inner layer was <50% of the total wall width or was completely absent (Fig 4C, graph); in many of them the existing outer layer was also disorganized (Fig 4C, bottom image).
We next tested whether pfa4Δ cells have defects in cell wall stress signaling that render them unable to respond to environmental changes, by growing serial dilutions of wild-type, mutant, and mutants expressing either PFA4 or the catalytically-dead pfa4AS on media containing caffeine (S5E Fig). Caffeine stimulates the cAMP/PKA pathway, activating PKA1/2 and thereby mimicking cell wall stress. This chemical activation of the cell integrity pathway can be lethal if there is a preexisting defect in the pathway [34]. pfa4Δ could not grow under these conditions, consistent with a signaling defect in response to cell wall stress. Taken together, these results indicate that pfa4Δ cells have both altered cell wall structure and defective transduction of signals from the cell integrity pathway that would normally compensate for such changes. This results in a disordered wall with altered exposure of cell wall components, which in turn likely facilitates recognition by host cells (see Discussion).
A distinguishing feature and major virulence factor of C. neoformans is its polysaccharide capsule, which associates with the cell wall via α-glucan [33, 35]. We observed that pfa4Δ cells were clumpy in culture, a characteristic often seen in hypocapsular cryptococci that suggested these cells might have a capsule defect. Interestingly, this was not observed: the capsules of pfa4Δ cells were morphologically similar to those of wild-type under inducing conditions, although they were slightly smaller overall (Fig 4D).
C. neoformans survives and proliferates within macrophage phagolysosomes [31, 36, 37]. We assessed the behavior of pfa4Δ cells in this challenging environment and found that host phagocytes rapidly killed them (Fig 5A). In contrast, wild-type and complemented mutant cells showed robust growth in this context (Fig 5A), and even caused host cell numbers to decrease slightly (they were unperturbed by the pfa4Δ mutant).
We further tested the virulence of pfa4Δ in a mouse model of cryptococcosis, monitoring disease progression by weight loss. Infection with wild-type or the complemented mutant killed 50% of the mice in 16 days, with all animals steadily losing weight by about two weeks and succumbing to infection by day 18 (Fig 5B). In contrast, mice infected with pfa4Δ showed a modest (3–5%) and transient (days 8–14) weight loss early in infection, but recovered and grew normally until the study was terminated at day 45; no CFU were recovered from lung or brain at that time. This dramatic effect of a single PAT on fungal pathogenesis is unprecedented.
Despite the importance of palmitoylation to fundamental processes of cell biology [11], the palmitoylome of C. neoformans, like that of other fungal pathogens, has never been defined, with only one protein (Ras1) shown to be functionally palmitoylated [19]. The dramatic effects of PFA4 deletion, which our active-site mutation studies showed are due to lack of enzymatic activity, indicate that this lipid modification is crucial for cryptococcal cell integrity and virulence. To mechanistically explain these observations, we used fatty acid chemical reporter labeling and bioorthogonal chemistry proteomics to determine the specific set of proteins modified by Pfa4 [38–40]. In this method, cells are metabolically labeled with alk-16, a palmitic acid analog with an alkyne group, which is incorporated into proteins in place of the normal fatty acid (Fig 6A). Proteins modified with alk-16 can then be labeled with azide-functionalized reagents via ‘click chemistry’ for fluorescence detection or proteomic analysis (Fig 6A; [39–41]). We grew wild-type and pfa4Δ cells with alk-16, and then performed labeling reactions with azido-rhodamine for in-gel fluorescence detection (Fig 6B; [38]). The total protein profile of the mutant was similar to that of wild-type, although there were fewer species at high molecular weights; alk-16 labeling did not alter this pattern. The alk-16-labeled proteins of both strains also showed similar profiles, but with a slight decrease in the overall levels of modified proteins in the mutant (Fig 6B).
We next reacted alk-16-labeled proteins with azido-biotin, purified the biotinylated proteins with streptavidin beads, and evaluated Pfa4-specific alk-16-labeled proteins by comparative proteomics. Of the 427 proteins identified in two independent experiments with at least 2 unique peptides, 72 showed ≥5-fold enrichment in the wild-type compared to the pfa4Δ mutant in both studies (S1 File). High-confidence Pfa4 substrates included proteins that act in a variety of cell wall processes, including cell wall synthesis, membrane trafficking, signal transduction, and transport (Table 3). At the top of our list was chitin synthase 3 (Chs3), which has been characterized as a Pfa4 substrate in S. cerevisiae [42, 43]. Interestingly, a second chitin synthase (Chs1) is also a Pfa4 substrate.
A number of substrates from our Pfa4-dataset have homologs that are known to be palmitoylated in S. cerevisiae, although not necessarily by Pfa4; these include Sso1 and Sso2, Vac8, Gpa2, Yck1 and Yck2, and Env7 [28]. Some have homologs known to be palmitoylated in other systems, such as Rho11 [44] and Vac8 [16]. Notably, C. neoformans Ras1 was labeled by alk-16 independent of Pfa4 (S1 File). This is consistent with previous studies in budding and fission yeast demonstrating that Ras1 is also a substrate of the Erf2/4 PAT complex, which is intact in our mutant [28, 41, 45]. Together, our results indicate that Pfa4 does not significantly alter global levels of fatty-acylation in C. neoformans, but palmitoylates specific proteins central to stress resistance and consequently to virulence, despite the presence of six other probable PAT genes in the cryptococcal genome.
Chs3 is critical for normal wall synthesis and maintenance [32, 46]. The discovery that it is a major substrate of Pfa4 is consistent with the multiple cell wall-related defects we observed in the pfa4Δ mutant, and explains how Pfa4 influences cell morphology, integrity, and consequently virulence. To establish a direct link between Pfa4-mediated palmitoylation and Chs3 function, we generated strains expressing Chs3-mCherry from the endogenous locus in both wild-type and pfa4Δ backgrounds. Palmitoylated Chs3 localized to internal compartments and to the plasma membrane, seen as a homogeneous rim outlining the cells (Fig 6C, top panel). In contrast, in the pfa4Δ cells, Chs3 is restricted to internal membranes, with occasional staining of vacuoles, suggestive of degradation (Fig 6C, bottom panel). This mislocalization of Chs3 is consistent with lack of palmitoylation of this protein and the cell wall-related defects observed in pfa4Δ cells. Interestingly, pfa4Δ and chs3Δ cells do not have completely congruent phenotypes. For example, chs3Δ does not exhibit the increased phagocytosis that first brought pfa4Δ to our attention (Fig 7A), although both strains show increased sensitivity to some cell wall stressors (Fig 7B) and poor retention of melanin at the cell wall (Fig 7C–7E, and [32]). Differences between the two mutants are likely due to the redundancy of both chitin synthases and palmitoyltransferases in C. neoformans, as well as the reduced palmitoylation of other Pfa4 substrates in the pfa4Δ mutant (see Discussion).
Phagocytes play multiple roles in cryptococcal pathogenesis, destroying fungi under some circumstances but also potentially harboring them and enabling them to survive, proliferate, and disseminate [1, 36]. Some outcomes of cryptococcal interactions with macrophages, including fungal engulfment and intracellular proliferation, correlate highly with patient outcome [7, 8]. These observations make host-pathogen interactions a compelling area of study, and raise the question of whether they might present feasible targets for antifungal therapy. Pursuing this question, however, requires mechanistic understanding of these events from the vantage point of both host and pathogen.
As a first step in such investigations, we used a high-content imaging-based assay to screen 1,201 C. neoformans mutants (corresponding to ~17% of the genome). We found 56 mutants that showed significantly altered uptake by host cells, including 29 lacking genes of unknown function that have not previously been investigated. Many of the mutants showing increased engulfment had been reported to be defective in host-pathogen interactions in other models; this validated our screen and provided strong support for uncharacterized hits. The genes deleted in several of the high-uptake mutants encode proteins involved in synthesis or remodeling of the cell wall and/or capsule, surface structures that interact most directly with host cells. Others encode signaling molecules or transcription factors involved in the response to environmental changes, such as would be encountered during infection. Intriguingly, most of the hits with reduced engulfment, more than half of which encode proteins with no known homologs in S. cerevisiae, have never been investigated. Future studies defining their biological roles should increase our understanding of C. neoformans’ interactions with host cells. Notably, the level of engulfment has no simple relationship to overall virulence in animal models, perhaps illustrating the complex role of phagocytosis in cryptococcal infection [36, 47]. For example, two hypervirulent mutants [10] showed opposite uptake results, with one (9A12) very poorly internalized while the other (2G9; lacking RIM101) was avidly engulfed.
One mutant that demonstrated increased uptake by phagocytes lacks PFA4, which encodes a protein containing the well-characterized DHHC domain characteristic of PAT enzymes. PATs catalyze the post-translational addition of palmitate to proteins, a reversible modification that can influence the localization, stability, and/or function of their substrates. The C. neoformans H99 genome contains seven genes encoding DHHC-domain proteins, and functional redundancy is common in this family of enzymes. It was therefore surprising that deletion of PFA4 had such a dramatic effect on C. neoformans morphology, stress sensitivity, and virulence. This suggested that Pfa4 modifies specific substrates that are critical in cryptococcal biology. For this reason, and because of the recent attention to PATs as potential antimicrobial drug targets [48, 49], we investigated the mechanism(s) by which lack of Pfa4 causes these phenotypes.
We postulated that Pfa4 was the primary or sole PAT modifying important cryptococcal proteins required for cell integrity and virulence. Our proteomic analysis supported this hypothesis, identifying 72 proteins as preferentially palmitoylated by Pfa4 (Table 3 and S1 File). As in S. cerevisiae [42], Chs3 is a key Pfa4 target. This protein is one of eight cryptococcal chitin synthases and is responsible for synthesizing the majority of cellular chitin during vegetative growth [32, 46]. If Chs3 does not properly localize and act in pfa4Δ cells as a result of lacking palmitoylation, one would expect to see cell walls with reduced chitin and consequently impaired function. This is exactly what we observe: Chs3-mCherry in the mutant is mostly restricted to internal membranes and is depleted from the plasma membrane compared to in WT cells (Fig 6C); as a consequence, the inner layer of the cell wall, which corresponds to the layer containing chitin [33], is markedly reduced. Chs1, another class IV chitin synthase, is also preferentially modified by Pfa4 and may contribute to these cell wall defects.
Beyond altered chitin synthase activity, cell wall production is likely further compromised in pfa4Δ cells secondary to defects in intracellular traffic. Pfa4 substrates that we identified include several proteins involved in protein secretion that are known to be palmitoylated in S. cerevisiae (Table 3) or other organisms. Since multiple proteins involved in cell wall biogenesis are membrane proteins that travel to their site of action in secretory vesicles, dysfunction of SNARES or other proteins involved in this transport could alter cell surface composition via partial blockade or mislocalization of vesicle cargo.
Aberrant cell wall synthesis probably causes the dramatically altered morphology of pfa4Δ cells (Figs 2 and S3). Such changes were previously only seen in dying cryptococci that had been exposed to harsh conditions, such as digestion with lysosomal extracts in vitro or extended growth in infected animals [50, 51]. In contrast, pfa4Δ shows wall collapse even during normal growth in culture in the absence of any stains or exogenous compounds. Mutant cells are also hypersensitive to salt and sorbitol, suggesting defects in regulating turgor pressure. Regulatory disturbance is further suggested by the sensitivity of pfa4Δ to caffeine, which activates the cell integrity pathway. These phenotypes are consistent with our identification of several proteins that function in signal transduction as Pfa4 substrates (Table 3). These include Rho11, a GTPase that acts in cell integrity signaling via the MAP kinase pathway [52], and an uncharacterized protein similar to Rho GTPase activating protein (GAP) that may be the Rho11 GAP. Mislocalization of these proteins would likely impair the cellular response to cell wall damage. We also identified the α subunit of the large G-protein Gpa1 as a Pfa4 substrate; this protein is upstream of cryptococcal cAMP signaling and is involved in pheromone and mating responses [53], which could explain the mating defects when pfa4Δ strains are crossed to each other (S6 Fig). Perturbation of multiple signaling pathways due to lack of Pfa4 severely limits the mutant cells’ ability to respond appropriately to changing environmental conditions, exacerbating the effects of defective wall synthesis and undermining mutant survival in the host.
We considered the possibility that the increased uptake of pfa4Δ cells by host phagocytes reflected inviability of the yeast. However, we ruled out this possibility by demonstrating viability of the mutant under conditions of our uptake assays (S5 Fig). Furthermore, killing fungi by treatment with heat, ethanol, or azide did not alter uptake of wild-type (S1C Fig) or mutant cells.
The combination of impaired cell wall synthesis and inability to appropriately respond to this condition results in weak and disorganized walls. This may impair other key virulence attributes of C. neoformans, such as the polysaccharide capsule, which associates with the cell wall. A perturbed wall, even in cells where the capsule is only slightly reduced in radius (as with pfa4Δ), may alter the capsule so that it cannot maintain its normal antiphagocytic role and exposes underlying wall components. This, combined with the changed wall arrangement, could explain our observation of abnormally high surface accessibility of specific cell wall components (Fig 3). These included cell wall mannoproteins [54, 55], which can interact with host phagocyte mannose receptors, and chitosan, which also interacts with macrophage receptors and induces a robust inflammatory response [56, 57]. Greater accessibility of these glycans could in turn explain the increased phagocytosis of pfa4Δ cells by macrophages. Once engulfed, these less robust cells, with defects in cell wall, signal transduction, and virulence factor expression, fare poorly (Fig 5A). Potentially reducing the pathogenicity even of cryptococci that remain outside of host phagocytes, we found important membrane transporters that are not correctly palmitoylated in the mutant. These include a putative carbohydrate transporter, a phosphate transporter, and NIC1 and SIT1, which transport nickel and siderophore-iron complexes, respectively [58, 59]. Because these metals are limiting during infection, incorrect processing or targeting of their transporters could influence pathogenesis. Furthermore, melanin, an important virulence factor in this pathogen, is poorly retained at the cell wall (Fig 7C–7E), a phenotype also seen in chs3Δ cells [32] and associated with reduced virulence. Ultimately, all of these factors combine to result in avirulence of the pfa4Δ mutant (Fig 8).
In contrast to our findings, the initial survey of cryptococcal deletion mutants [10] categorized the strain deleted for PFA4 (2A12) as normal in virulence. This may reflect the practical strategy used in that large-scale study, where pools of mutants were assayed, or the timing of those virulence studies. We did observe that animals infected with pfa4Δ initially show mild symptoms of disease, suggesting the initiation of a pathogenic process that might be interpreted as normal infectivity in short-term studies (as in ref. [10]), but that they subsequently clear the infection and recover completely. Consistent with our observations, 2A12 does show reduced virulence in recent studies of this deletion library using both an IV mouse model and invertebrate models of infection (performed at room temperature) [60]. The latter also supports our conclusion above that Pfa4’s contribution to virulence is temperature independent.
As well as demonstrating the key role of Pfa4-dependent palmitoylation in C. neoformans, our work provides valuable data sets to the community from both our screen and our palmitoylome analysis. It also bolsters a model that explains why multiple PATs have been retained during evolution despite the widespread redundancy of these enzymes: key PATs like Pfa4 may modify specific substrates that perform critical functions, in addition to sharing substrates with other PATs. This concept is supported by the recent identification of specific PATs that regulate central pathogenic processes in Toxoplasma and Plasmodium [16–18].
Our determination of the Pfa4-palmitoylome offers new insights into the role of an important regulatory lipid modification in the biology of C. neoformans. Pfa4 in C. neoformans is notable in modifying proteins that exhibit diverse modes of membrane association, including those that are otherwise predicted to be soluble or to have one or many membrane-spanning domains (Table 3). In contrast, its S. cerevisiae homolog appears to be restricted to modifying polytopic membrane proteins [28, 42]. Several cryptococcal Pfa4 substrates are also fungal-specific (e.g., the chitin synthases, vacuolar protein 8, the nickel and siderophore transporters, and the product of CNAG_02114). Finally, cryptococcal Pfa4 is unique among PATs studied to date in that it is essential for virulence in an animal model. The closest human homolog of Pfa4, DHHC6, is considerably distant in sequence homology, with the similarity restricted to the catalytic DHHC domain. These findings encourage efforts towards development of specific PAT inhibitors as novel avenues for therapeutics.
Strains used were C. neoformans serotype A strain H99α, its derivative KN99α, and deletions in these backgrounds (see below). The cryptococcal partial deletion collection in H99α [10] was purchased from the Fungal Genetics Stock Center (University of Missouri, Kansas City, MO) and H99α chs3Δ was a generous gift from Jennifer Lodge (Washington University). Fungal strains were maintained at -80°C and grown at 30°C on yeast peptone dextrose (YPD) with antibiotics as appropriate (100 μg/mL of nourseothricin (clonNAT, WERNER BioAgents, Germany) or 100 μg/mL G418 (Geneticin, Life Technologies, USA)).
The human monocytic cell line THP-1 (ATCC TIB-202) was grown in THP-1 complete media (RPMI-1640 with L-glutamine supplemented with 1 mM sodium pyruvate, 0.05% 2-mercaptoethanol, 10% FBS, and 100 units/mL Penicillin- 100 μg/mL Streptomycin solution as indicated) and differentiated with phorbol 12-myristate 12-acetate (PMA, from Sigma, St. Louis, MO) as described in [9]. THP-1 cultures were split every 3–4 days (inoculum of 105 cells/mL) and new batches were thawed every month.
All tissue culture plasticware and media were from BD Falcon and Sigma, fungal media components from Difco, PCR primers from Sigma, biolistic transformation reagents and materials from Bio-Rad, DH5α cells from Life Technologies, and restriction enzymes from New England Biolabs. Reagents for electron microscopy were from Ted Pella (Redding, CA) and Polysciences (Warrington, PA); antibodies for immunofluorescence were from Abcam (ab2900, anti-EEA1 rabbit polyclonal) or the Developmental Studies Hybridoma Bank (clone H4A3, University of Iowa); and antibodies for immunoblotting were from Sigma (clone HA-7 anti-HA mouse monoclonal and anti-FLAG rabbit polyclonal). Reagents for bioorthogonal labeling and click chemistry were from Sigma, except for azido-rhodamine, which was prepared as previously described [38].
To screen fungal mutants, THP-1 cells were seeded in 96-well plates (3.33 ×105 cells/mL, 100 μL), incubated for 48 hr (37°C, 5% CO2) in THP-1 complete media, washed three times with 150 μL RPMI-1640, and cultured for one day in serum-free media with antibiotics. In parallel a 96-pin replicator (Nalge Nunc International, Rochester, NY) was used to inoculate strains from the C. neoformans deletion collection into a Nunc Edge—96 well microplate containing 150 μL YPD per well. The microplates were incubated at 30°C overnight on a mini-orbital shaker (BELLCO Biotechnology, Vineland, NJ), followed by transfer of a 35 μL aliquot from each well into a new 96-well flat-bottom microplate (Costar 3904). The transferred cells were washed once with PBS (pH 7.5), once in Mcllvaine’s buffer (pH 6.0), and then resuspended in 100 μL of the same buffer containing 100 μg/mL Lucifer Yellow dye (Sigma L0144). After a 30 min incubation at RT with gentle agitation the cells were collected, washed once with PBS, and opsonized (30 min, 37°C) in 100 μL 40% human serum with gentle agitation. Serum was obtained from healthy donors with informed consent under a protocol approved by the Washington University in St. Louis Institutional Review Board. Finally, the cells were washed three times with PBS, resuspended in 150 μL RPMI-1640, and 35 μL from each well was diluted into 1 mL of pre-warmed RPMI-1640 in a deep-well 96-well plate (Nunc). To initiate the assay, the medium from each well containing THP-1 was aspirated and replaced by 100 μL of the cryptococcal suspension. After a 1 hr incubation (37°C, 5% CO2) the plates were washed vigorously four times with 150 μL PBS using a microplate washer (ELx405TM Select CW, Biotek, Winooski, VT). The samples were then immediately fixed in 150 μL 4% formaldehyde (20 min, 4°C), washed twice with PBS, and permeabilized for 20 min at RT with 0.1% saponin in PBS (150 μL). Samples were next washed twice with PBS, stained (15 min, RT, in the dark) with 2 μg/mL DAPI and 0.25 μg/mL HCS CellMask Deep Red (Life Technologies) in PBS, washed twice more with PBS, and 100 μL of 10 mM NaN3 in PBS was added to each well. Plates were either imaged immediately (on an IN Cell 1000 analyzer, GE, Piscataway, NJ) or stored at 4°C for later analysis. GE INCell Investigator Developer Software was used to identify host cell and fungal borders and calculate the overlap as described in [9]. Fungal cells that overlapped >50% with host cell bodies were considered internalized, ≤50% were considered adherent, and fungal cells with no overlap were not counted. In parallel with the screening assay, an aliquot of each fungal cell suspension was pipetted into empty 96-well plates for enumeration to allow normalization of results to fungal cell number (macrophage uptake of C. neoformans is linear in the range of fungal concentrations used in these assays [9]). The results were analyzed plate-wise (to reveal any systematic errors in different plates) before normalization and calculation of values relative to wild-type.
We used the split marker method [61] to delete PFA4 (CNAG_03981) in H99α and KN99α after amplifying NAT resistance split marker fragments from genomic DNA of strain 2A12 (pfa4Δ) from the Madhani deletion collection. For chromosomal complementation, we used a split marker approach to replace the NAT cassette of the mutant with wild-type genomic PFA4 sequences in tandem with a G418 resistance cassette. For endogenous tagging of the CHS3 gene (CNAG_05581) with mCherry, the last 1,674 bp of CHS3 were amplified with a BamHI site replacing the STOP codon and ligated to a BamHI/AvrII-cut fragment composed of mCherry followed by an HA epitope, a STOP codon, and 445 bp of the TRP1 terminator. The ligated fragment was cloned in front of a NAT resistance cassette and 616 bp of the CHS3 terminator (sequences immediately following the STOP codon) were subsequently cloned after the NAT cassette. The resulting plasmid was digested with BglII/MluI to release the 5’ fragment of the split marker and with XmaI/EcoRV to release the 3’ fragment of the split marker. Transformation was by biolistics (Bio-Rad PDS-1000/He) as described in [62].
For plasmid construction, a fragment encompassing the PFA4 coding locus and 226 bp of 3’ sequence was amplified so as to incorporate sequence that encodes 1.5X HA epitope tags in place of the first 2 codons. Fusion PCR was used to ligate this fragment to a second amplicon consisting of 900 bp of 5’ sequence (including the starting ATG) and sequence encoding 1.5X HA epitope tags, so as to reconstitute sequence encoding an N-terminal 3X HA-tagged Pfa4 sequence. This product (~3.5 kb) was cloned into ApaI/KpnI-digested pIBB103 [63] for expression and also used as template for mutagenesis of the DHHC motif into DHAS using overlapping primers containing the codon change. Plasmid transformation was as described in [63].
Cells were grown overnight at 30°C in YPD (with appropriate antibiotics if needed to maintain plasmids), diluted as for phenotyping, washed in PBS, and resuspended at 107/mL for staining as follows (all manipulations at RT): For LY and EoY (Sigma), cells were washed once in McIlvaines buffer, pH 6.0; resuspended in the same; and incubated for ~15 min with 250 μg/mL of the dye. For CFW (Fluorescent Brightener 28, Sigma), UV2B (Polysciences, Inc.) and Pont (Pontamine fast scarlet 4B, Bayer Corp., Robinson, PA), cells were stained in PBS with 100 μg/mL of CFW or UV2B or a 1:10,000 dilution of Pont (w/v). For ConA-FITC (Sigma), cells were stained with 30 μg/mL in Hepes-buffered saline, pH 7.0, containing 10 mM each MgCl2 and CaCl2.
For fluorescence microscopy, stained cells were washed twice, resuspended in the same volume of the corresponding buffer, mixed vigorously, spotted onto glass slides, covered, and imaged immediately on a wide field Zeiss Axioskop 2 MOT Plus with appropriate filters (DAPI for CFW and UV2B; FITC for LY, EoY, and ConA-FITC; and Texas Red for Pont). For the Chs3-mCherry strains, overnight cultures grown in YPD were washed twice with PBS, resuspended in 3 mL of PBS, and 6 μl were spotted on polylysine-coated glass slides and imaged immediately. For flow cytometry cells were washed three times, fixed in 3.7% formaldehyde/PBS (10 min; RT) or resuspended in PBS with 10 mM NaN3 and analyzed on an LSRII flow cytometer (Becton Dickinson, Franklin Lakes, NJ) for analysis using FlowJo software (Tree Star Inc., Ashland, OR).
For transmission EM, overnight cultures grown in YPD were diluted 10-fold, grown to OD600 = 0.2 (~107/mL), and washed twice in PBS. The cell pellet was resuspended in 1 mL of primary fixation mix (2.5% paraformaldehyde/2% glutaraldehyde in 100 mM cacodylate buffer, pH 7.2), incubated for 1 hr at room temperature (RT), washed in the same buffer, and post-fixed in 1% osmium tetroxide (Polysciences, Inc.) for 1 hr at RT. Samples were then rinsed in the same buffer, followed by dehydration in a graded series of ethanol and propylene oxide prior to embedding in Eponate 12 resin (Ted Pella, Inc.). Sections of 90 nm were cut with a Leica Ultracut UCT ultramicrotome (Leica Microsystems, Inc., Bannockburn, IL), stained with uranyl acetate and lead citrate, and viewed on a JEOL 1200EX transmission electron microscope (JEOL USA Inc., Peabody, MA) equipped with an AMT 8 megapixel digital camera (Advanced Microscopy Techniques, Woburn, MA).
For scanning EM, cultures were grown and fixed as above but in sodium phosphate buffer, then washed and 8.8 x 106 cells (4 x 106 cells/cm2) were added to wells of a 6-well plate containing a polylysine-coated plastic coverslip. After incubation at 4°C for 1–2 hr the coverslips were washed twice with DPBS, re-fixed in 2% paraformaldehyde, 2.5% glutaraldehyde in 0.1 M Sorensen’s sodium phosphate buffer (potassium-free, pH 7.4), and then sequentially rinsed in buffer and NanoPure Ultra-filtered deionized water. They were next post-fixed in 1% osmium tetroxide (aqueous) for 1 hr, rinsed with NanoPure Ultra-filtered deionized water, dehydrated in ethanol (30%, 50%, 70%, 80%, 90%, 3X 95%, and 3X absolute ethanol), critical point dried (Tousimis Samdri-780, Rockville, MD) via liquid carbon dioxide, mounted on aluminum stubs with double-sided adhesive carbon tabs, and sputter coated (Tousimis Samsputter-2a) with gold-palladium. Images were acquired using a Hitachi S2600 (Hitachi-hitec, Tokyo, Japan) instrument.
Strains to be tested were grown overnight in YPD, diluted to ~2 x 106/ml, and grown for two doublings. The cultures were then serially diluted (10-fold) and spotted (5 μL) onto buffered (pH 6.8 with 100 mM KPO4 buffer) synthetic dextrose medium with 1 mg/mL calcofluor white or onto YPD with 1.2 M NaCl; 1.2 M KCl; 0.01 and 0.03% SDS; 1, 3, and 5 mM H2O2; 0.1, 0.25, 0.5, 0.75, and 1 mg/mL caffeine; 1 mg/mL Congo red (stock prepared in 70% ethanol); or 25 μg/mL Lucifer Yellow. The same plates were also prepared containing various concentrations of sorbitol (0.5, 1, or 1.5 M). Plates were incubated at 30°C and 37°C for 3–4 days. Sensitivity to lysing enzymes was tested as in [64].
Cells were grown overnight in YPD, washed twice with DMEM, and 1 mL aliquots (106 cells) were pipetted into 24-well tissue culture plates (3 wells per strain) and incubated (37°C; 5% CO2) for 24 hr. The suspension was washed twice with deionized water (dH2O), resuspended in 24 μL dH2O, mixed with ~8 μL of India ink and visualized on a Zeiss Axioskop 2 MOT Plus microscope. At least 150 cells from each strain (50 per well) were analyzed with ImageJ (NIH) for capsule thickness ((outer capsule diameter minus cell wall diameter)/2).
For solid media assays, the cells were grown overnight in YPD medium at 30°C, diluted the next morning in 5 mL of YPD, grown to an OD600 of 0.2, washed twice in PBS, and adjusted to 107 cells/mL in PBS. 10-fold serial dilutions were made and 5 μL of each dilution spotted on L-DOPA (1 mM) plates. The plates were incubated at 25°C, 30°C, and 37°C for 3–4 days in the dark.
For assays in liquid medium, cells of each strain were grown similarly overnight, diluted in 25 mL of YPD, and allowed to grow for 2–3 generations. At that point, the cells were washed in PBS, resuspended in 2 mL glucose-free asparagine and salts media, and the cell density was quantified. The strains were adjusted to 5 x 107 cells/ml and incubated at 25°C for 18–24 hr in asparagine medium containing 1 mM L-DOPA. The cultures were spun down at 1000xg for 10 min, and photographed. To quantify the melanin in the media, the OD405 was measured for 100 μL aliquots of the supernatant fractions.
To assess fungal survival in macrophages, THP-1 cells grown in 12-well plates (250,000 cells per well), were washed with assay medium (RPMI + 1% FBS). In parallel, overnight fungal cultures (OD600 = 0.2–0.4; 1–2 x 107/mL) were washed twice with PBS, resuspended (108 cells/mL) in 40% human serum for opsonization (37°C; 30 min; with rotation), rewashed, resuspended in assay medium, and added to the THP-1 cells at an MOI of 0.1 or 1.0 as indicated. Plates were incubated for 1 hr, rinsed twice with 1 mL Dulbecco’s PBS (DPBS), and incubation continued for 0 hours (to measure initial association) or for the time indicated after addition of RPMI + 1%FBS. At the desired assay time points the medium was aspirated, wells were washed once with 1 mL DPBS, 1 mL of lysis buffer (0.05% SDS, 1 mM EDTA) was added, and the plate was shaken on a plate mixer for ~3 min. The resulting lysate was collected, vortexed vigorously, diluted, and spotted onto YPD media for determination of colony forming units (CFU).
To test virulence in mice, strains were cultured overnight in YPD, collected, washed in PBS, diluted to 106 cells/mL in PBS, and briefly sonicated (to disperse clumps seen in pfa4Δ). Sonication did not adversely affect mutant viability. Aliquots (50 μL) of the suspension were used to intranasally inoculate groups of ten mice (4–6 week-old female A/Jcr mice; National Cancer Institute) and dilutions of the suspension were plated immediately after infection to confirm inocula. Animals were monitored closely and sacrificed if they lost >20% relative to peak weight or at the end of the experiment (45 days). Homogenates of lungs and brains from 3 of the surviving mice infected with pfa4Δ were plated to determine organ burden.
Whole cell lysates (50 μg) were diluted with SDS buffer (4% SDS, 150 mM NaCl, 50 mM triethanolamine pH 7.4, Roche EDTA-free protease inhibitor cocktail) to 44.5 μL and then reacted with 5.5 μL freshly prepared click chemistry reaction cocktail containing azido-rhodamine (1 μL, 10 mM stock solution in DMSO), tris(2-carboxyethyl)phosphine hydrochloride (TCEP) (1 μL, 50 mM freshly prepared stock solution in deionized water), tris[(1-benzyl-1H-1,2,3-triazol-4-yl)methyl]amine (TBTA) (2.5 μL, 10 mM stock solution in DMSO/t-butanol) and CuSO4•5H2O (1 μL, 50 mM freshly prepared stock solution in deionized water)] for 1 h at room temperature. The click reactions were terminated by the addition of ice-cold methanol (1 mL). The mixtures were placed at −20°C overnight and then centrifuged at 18,000×g for 20 min at 4°C to precipitate proteins. The supernatants from the samples were discarded. The protein pellets were washed with methanol twice, allowed to air-dry for 10 min, resuspended in 35 μL of SDS lysis buffer, and diluted with 12.5 μL 4× reducing SDS-loading buffer (40% glycerol, 200 mM Tris-HCl pH 6.8, 8% SDS, 0.4% bromophenol blue) and 2.5 μL 2-mercaptoethanol. The resulting samples were heated for 5 min at 95°C and resolved on 4–20% SDS-PAGE gels (Bio-Rad). For in-gel fluorescence scanning, the gels were destained in 40% methanol, 10% acetic acid for at least 1 h, and then scanned on a GE Healthcare Typhoon 9400 variable mode imager with excitation and emission at 532 nm and 580 nm, respectively. After scanning, gels were also stained with Coomassie Brilliant Blue (Bio-Rad).
For affinity purification of alk-16-modified proteins, 2 mg of cell lysates labeled with alk-16 were subjected to Cu(I)-catalyzed click reaction as described above, except that azido-biotin was substituted for azido-rhodamine. Methanol-precipitated and washed protein pellets were resuspended in 200 μL of 4% SDS buffer (50 mM TEA, 150 mM NaCl, pH 7.4). Equal amounts of protein for each sample were diluted 1/4 by volume with 50 mM TEA buffer (150 mM NaCl, pH 7.4). 60 μl prewashed streptavidin agarose beads (Invitrogen) were added to each sample and the protein and bead mixtures were incubated for 1 h at room temperature on a nutating mixer. The beads were then washed once with PBS and 0.2% (w/v) SDS, three times with PBS and twice with 250 mM ammonium bicarbonate (ABC). Beads were resuspended in 500 μl 8 M urea, reduced with 10 mM DTT for 30 min, and then alkylated with 50 mM iodoacetamide in the dark for another 30 min. Finally, the beads were washed with 25 mM ammonium bicarbonate three times and digested with 0.5 μg of trypsin at 37°C overnight. The supernatant of each sample was collected, dried, and solubilized in 5% acetonitrile/1% formic acid for LC-MS analysis.
LC-MS analysis was performed with a Dionex 3000 nano-HPLC coupled to an Orbitrap XL mass spectrometer (ThermoFisher). Peptide samples were pressure-loaded onto a home-made C18 reverse-phase column (75 μm diameter, 15 cm length). A 180-minute gradient increasing from 95% buffer A (HPLC grade water with 0.1% formic acid) and 5% buffer B (HPLC grade acetonitrile with 0.1% formic acid) to 75% buffer B in 133 minutes was used at 200 nL/min. The Orbitrap XL was operated in top-8-CID-mode with MS spectra measured at a resolution of 60,000@m/z 400. One full MS scan (300–2000 MW) was followed by three data-dependent scans of the nth most intense ions with dynamic exclusion enabled. Acquired tandem MS spectra were extracted using ProteomeDiscoverer v.1.4.0.288 (Thermo, Bremen, Germany) and queried against the Uniprot complete Cryptococcus neoformans var. grubii H99 proteome (UP000010091) database concatenated with common known contaminants using MASCOT v.2.3.02 (Matrixscience, London, UK). Peptides fulfilling a Percolator calculated 1% false discovery rate (FDR) threshold were reported. The abundance of an identified protein was calculated based on the average area of its three most abundant peptides. For a protein to be considered a Pfa4-specific substrate, it had to be at least five-fold more abundant in the wild-type sample compared to the pfa4Δ sample as measured by protein abundance in both of the two independent experiments and be identified with at least two unique peptides.
|
10.1371/journal.pgen.1000006 | Differential Allelic Expression in the Human Genome: A Robust Approach To Identify Genetic and Epigenetic Cis-Acting Mechanisms Regulating Gene Expression | The recent development of whole genome association studies has lead to the robust identification of several loci involved in different common human diseases. Interestingly, some of the strongest signals of association observed in these studies arise from non-coding regions located in very large introns or far away from any annotated genes, raising the possibility that these regions are involved in the etiology of the disease through some unidentified regulatory mechanisms. These findings highlight the importance of better understanding the mechanisms leading to inter-individual differences in gene expression in humans. Most of the existing approaches developed to identify common regulatory polymorphisms are based on linkage/association mapping of gene expression to genotypes. However, these methods have some limitations, notably their cost and the requirement of extensive genotyping information from all the individuals studied which limits their applications to a specific cohort or tissue. Here we describe a robust and high-throughput method to directly measure differences in allelic expression for a large number of genes using the Illumina Allele-Specific Expression BeadArray platform and quantitative sequencing of RT-PCR products. We show that this approach allows reliable identification of differences in the relative expression of the two alleles larger than 1.5-fold (i.e., deviations of the allelic ratio larger than 60∶40) and offers several advantages over the mapping of total gene expression, particularly for studying humans or outbred populations. Our analysis of more than 80 individuals for 2,968 SNPs located in 1,380 genes confirms that differential allelic expression is a widespread phenomenon affecting the expression of 20% of human genes and shows that our method successfully captures expression differences resulting from both genetic and epigenetic cis-acting mechanisms.
| We describe a new methodology to identify individual differences in the expression of the two copies of one gene. This is achieved by comparing the mRNA level of the two alleles using a heterozygous polymorphism in the transcript as marker. We show that this approach allows an exhaustive survey of cis-acting regulation in the genome; we can identify allelic expression differences due to epigenetic mechanisms of gene regulation (e.g. imprinting or X-inactivation) as well as differences due to the presence of polymorphisms in regulatory elements. The direct comparison of the expression of both alleles nullifies possible trans-acting regulatory effects (that influence equally both alleles) and thus complements the findings from gene expression association studies. Our approach can be easily applied to any cohort of interest for a wide range of studies. It notably allows following up association signals and testing whether a gene sitting on a particular haplotype is over- or under-expressed, or can be used for screening cancer tissues for aberrant gene expression due to newly arisen mutations or alteration of the methylation patterns.
| Understanding the genetic causes of phenotypic variation in humans still remains a major challenge for human genetics. In hundreds of cases, a single DNA sequence polymorphism affecting a protein coding sequence has been linked to a clear simple Mendelian phenotype (see e.g. [1]) and, for a much smaller but increasing number of cases, to more complex phenotypes [2]–[4]. Recent developments in high-density genotyping technologies have led to the completion of several whole genome association studies that test hundreds of thousands of markers for a specific disease. While earlier studies essentially focused on variants in coding sequences and regions immediately surrounding candidate genes, whole genome scans interrogate, in an unbiased way, most of the human genome including large regions of non-coding DNA that had not been studied previously. Interestingly, some of the strongest signals observed in these association studies are located in non-coding regions, either in large introns (e.g. [5]–[7]) or far away from any annotated loci (e.g. [8] and references therein). The mechanisms connecting these polymorphisms to the etiology of the diseases are still unclear but regulation of gene expression remains an obvious candidate. It is thus becoming particularly important to have a powerful and reliable method to easily test the influence of DNA polymorphisms on gene expression. One of the approaches commonly used to identify regulatory polymorphisms is to look for statistical associations between variation in gene expression and individual genotypes [9],[10]. This method offers the advantage of simultaneously analyzing thousands of genes using gene expression arrays and has yielded fascinating results in yeast [11],[12] and mouse [13]–[16]. Its application in humans [17]–[24] suffers from relatively low statistical power due to potential inter-individual differences in a large number of causal variants involved in the regulation of a specific gene [25], their modest effects and the burden of the multiple testing correction necessary to take into account the large number of independent tests performed. In addition, since this approach requires extensive genotype information for all individuals, it is costly to apply to new samples. An alternative approach is to compare the relative expression of the two alleles in one individual: the effect of a polymorphism affecting in cis the regulation of a particular transcript can be detected by measuring the relative expression of the two alleles in heterozygous individuals using a transcribed SNP as a marker [26]–[32]. Several studies have used this approach in humans but have been criticized for their low throughput or the apparent high variability.
Here we describe a novel array-based method that allows high-throughput assessment of differential allelic expression. We used a modified version of the Illumina GoldenGate genotyping platform, the Allele-Specific Expression (ASE) assay, to assess the extent of differential allelic expression for over 1300 genes in more than 80 human lymphoblastoid cell lines (LCLs). Our analyses include 352 genes located in ENCODE regions and chromosome 21 that have been previously screened for cis-regulatory polymorphisms using total gene expression [19]. This allows us to directly compare the advantages and drawbacks of the two approaches in terms of range, sensitivity and robustness. We specifically address the issue of experimental noise and reproducibility of the findings and show that biology, not experimental variability, is responsible for the patterns observed. We discuss the relevance of our results for the identification of the molecular mechanisms regulating gene expression, as well as their implications for future genetic studies.
We first assessed the extent of differential allelic expression at 1,432 exonic SNPs using 81 individual LCLs with the Illumina ASE technology (Figure 1 and Table S1 for the composition of the Illumina ASE Cancer Panel). This technology uses primer extension assays with fluorescence-labeled allele-specific primers to measure the proportion of each allele separately at the genomic and transcriptomic levels (Figure 2). Five hundred and twelve SNPs (in 345 genes) displayed an expression level significantly higher than background in at least three heterozygous individuals and were further investigated (see Materials and Methods for details). The extent of differential allelic expression at each SNP was obtained by comparing the relative amount of each allele in RNA to the ratio observed in DNA.
As a first effort to determine if the assay could reliably be used to assess differential expression we generated spike mixes using varying proportions of total RNA extracts from two individuals. For 20 exonic SNPs located in expressed transcript, the two individuals are homozygous for the different alleles (i.e. respectively AA and BB), while for 192 SNPs one individual is heterozygous and the other homozygous (i.e. either AB and AA, or AB and BB). Since the expression of each gene may differ between the two individuals, one does not expect to observe an exact translation of the proportions of total RNA mixed to the “allelic” expression level. However, the allelic expression differences estimated for the different spike mixes should be the proportional to each other. For all homozygous/homozygous mixes (20 out of 20 SNPs) and 83% of the heterozygous/homozygous mixes (159 out of 192), we observed a significant linear correlation (p<0.05) between the proportion of mixed total RNAs and the “allele-specific expression” estimated by the assay (Figure 3 and Figure S1).
We then tried to assess the threshold above which differential allelic expression would be genuine: even if the two alleles are equally expressed in one individual, we expect the ratio of allelic expression measured at a given marker to deviate stochastically from 50∶50 due to experimental variability. In order to differentiate technical noise from biological signal (i.e. the differences in allelic expression due to differential cis-acting regulation), we evaluated the extent of experimental variability in the assay by comparing independent estimates of allelic imbalance for duplicates of individual RNA. We used duplicated measurements from 81 individuals at all SNPs expressed to determine a robust estimation of the experimental variability (N = 31,503 duplicates). After averaging duplicate differences for each SNP over all individuals, we observed than less than 3% of the SNPs show a population average variability greater than 10% (see Materials and Methods for more details and Figure S2). This level of experimental variability corresponds to a ratio of allelic expression of 60∶40 (i.e. 1.5-fold difference). Thus, population-average allelic expression ratio at any SNP lower than 60∶40 can be explained by experimental noise, while a SNP displaying a population-average differential allelic expression greater than this threshold most likely reflects a biological process affecting cis-regulation. We are interested in the present study in identifying loci with common allelic expression differences and we thus focused on population-average differential allelic expression: the average over all heterozygous individuals of the extent of allelic expression differences, regardless of which allele is over expressed (this is addressed later). The identification of a single individual with dramatic allelic expression difference is also possible using the same approach (but a different detection cut-off) but is beyond the scope of this paper. Among the 345 genes expressed in this first panel (512 SNPs), 72 (87 SNPs) displayed an average level of allelic imbalance larger than this 40∶60 cut-off and were thus considered to display significant differences in allelic expression (Figure 4).
These analyses rely on the observation of the three genotypes in the population (i.e. AA, AB and BB). To also include SNPs with a lower minor allele frequency for each it was not possible to observe homozygotes for the minor allele in our small sample, we designed a second analysis method using solely the heterozygous individuals. If the alleles are differentially regulated we expect to observe in some cases a very large variance in the ratio of the two alleles in the population. We used this approach to determine SNPs for which the heterozygous individuals harbor a variance of the allelic ration higher than expected using a Maximum expectation algorithm (see Materials and Methods for details). This approach does not allow us to quantity the overall extent of differential allelic expression but identifies 8 genes with differential allelic expression that were not identified by the previous method.
When one considers the estimates of allelic expression obtained using different SNPs in a same transcript showing significant differences in allele expression (i.e. with a population-average ratio greater than 60∶40), we note that 36 out of 44 correlations between individuals estimates are significant (for an average r2 of 0.83). Individuals showing a large allelic expression difference at one SNP display similar patterns at all heterozygous positions of the transcript (an example is shown on Figure 5). This observation supports our findings that the experimental variability is low in the Illumina ASE assay and that this assay allows quantitative assessment of differential allelic expression. Consequently, the population-average estimates of allelic imbalance obtained with different markers in the same transcript tend to be similar (Table S2) but can vary since different individuals will be included in the average (depending on whether they are heterozygotes at this marker).
To further assess the validity of our results, we randomly selected 25 genes tested on the Illumina ASE platform and used quantitative sequencing of RT-PCR products [33] to measure allelic imbalance for the same SNP in the same individuals (Figure S3). The selected genes consisted of eight autosomal genes with significant allelic imbalance and 17 genes for which the level of differential allelic expression did not reach our significance threshold. We analyzed the same 81 individual LCLs using RNA from the same extract as for the Illumina assay. Overall, we observed a strong correlation between the estimates of allelic imbalance obtained for each individual using the two methods for the genes with a ratio of allelic expression larger than our 40∶60 cut-off (r2>0.8 for 6 out of 8 genes, see Figure 6A as a example). The correlations were not statistically significant for the genes for which the average difference in allelic imbalance did not reach our significance threshold (16 out of 17 genes, Figure 6B): minor deviations observed in the allelic ratio for these genes likely correspond to random variations and are therefore not expected to be reproducible. The strength of the correlation (measured by Pearson's r2) for all 25 genes is shown on Figure 4. One SNP in CD44 (rs8193) displayed a low but significant correlation between our estimates of allelic imbalance obtained from the Illumina assay and those using quantitative sequencing (p<10−4, r2 = 0.4075), even though the average level of allelic imbalance was below our significance cut-off on the Illumina platform. Allelic imbalance at CD44 has been previously reported [30] and it is likely that the signal observed at that gene is real but corresponds to a low level of differential expression. Two genes (ABL2, XRCC1) showed significant allelic imbalance in the Illumina ASE assay (with a mean allelic ratio of, respectively, 70∶30 and 65∶35) but were not validated by quantitative sequencing. Manual inspection of the Illumina results for these genes revealed that the allelic expression ratios were estimated using a small number of homozygotes for the minor allele (respectively, 1 and 2 individuals) which led to an incorrect estimation of the expected dye ratio for heterozygotes and to a general over-estimation of allelic imbalance. For further analyses, we manually curated the list of all genes with significant differential allelic expression to remove potential false positives due to low number of homozygous individuals.
Our study uses lymphoblastoid cell lines (LCLs) and it remains controversial whether culture conditions could artefactually generate differential allelic expression. We therefore tested whether allelic imbalance is influenced by harvesting the cells after different numbers of passages. This allowed us to control the effect of changes in the culture environment including pH, nutrient concentration and cell density at the time of harvest. We compared our estimation of allelic imbalance for three genes with significant population-average allelic imbalance in 47 individuals using recently thawed LCLs harvested after the 2nd, 4th and 6th successive passages (respectively, “growths” 1, 2 and 3). The correlations between the allelic imbalance estimations are displayed in Figure S4 for the comparison of growths 2 and 3. The estimations of allelic imbalance after different passages were very similar to each other (r2>0.9), supporting the idea that differential allelic expression is little influenced by variations in culture environment.
An alternative approach to identifying cis-regulatory polymorphisms is to test for statistical association in a population between total gene expression measurements and the genotypes at markers in or surrounding the transcript. Interestingly, one of the genes showing the most marked difference in allele expression in our analysis is one of the 14 genes identified by Cheung and colleagues in a previous genome-wide study [17]. One comprehensive analysis was recently conducted for 512 RefSeq genes in ENCODE regions and chromosome 21 using LCLs from 60 unrelated individuals genotyped by the HapMap project [19]. In order to compare the respective strengths and weaknesses of total gene expression mapping and differential allelic expression, we designed a second panel that includes SNPs in the same genomic regions to analyze the same individual LCLs (Figure 1 and Table S1 for details). Using the information from the HapMap phase I (release 16) to select common exonic SNPs, we were able to include 228 and 124 genes from, respectively, ENCODE regions and chromosome 21, while Stranger and colleagues selected 321 and 191 genes (after screening for genes with high variance in their expression among individuals, see Materials and Methods for details). From the regions analyzed by Stranger and al., two-hundreds and ninety SNPs (in 170 genes) showed an expression level significantly higher than the background in three or more heterozygous individuals and were further investigated. Forty-nine out of 170 genes show significant level of differential allelic expression including 6 out of the 21 genes identified by Stranger and colleagues and present in our panel. Additionally, TTC3 which shows significant association between total gene expression and genotypes in the study by Stranger et al. shows patterns of allelic expression consistent with differential allelic expression (including a very high correlation between the extent of differential allelic expression estimated using different SNPs) on the Illumina ASE assay, even though it did not pass the significance cut-off. Overall in this second panel, 497 SNPs in 317 genes were expressed in three or more heterozygous individuals (out of 1536 SNPs in 674 genes) and 78 SNPs in 65 genes showed a significant level of differential allelic expression (Figure 1).
To test whether intronic SNPs could be used instead of exonic SNP, we included for each gene on the second panel one intronic SNP. In general, intronic SNPs were less successfully analyzed and passed our expression threshold only for genes highly expressed in LCLs (Figure S5). This finding is consistent with previous observations [30] and the low proportion of unspliced mRNA (heteronuclear RNA) in cells relative to spliced transcripts. If the intronic SNP of a gene was detected in the RNA extract, it typically yields estimates of differential allelic expression very similar with those obtained using exonic SNPs.
Overall, 177 out of 1,009 expressed SNPs (in 140 out of 643 genes expressed, 22%) display population-average ratios of allelic expression larger than 40∶60 or an higher than expected variance in allelic expression among heterozygous individuals and are thus unlikely to result solely from stochastic variation in the experiment (Figure S6). Table 1 shows the 133 SNPs (100 genes) with significant allelic imbalance after manual curation to remove possible false positives due to a low number of individual homozygous for the minor allele (this list is likely over-conservative and the complete data is presented in Table S2).
Many of the genes with the highest extent of allelic imbalance in LCLs are located on the X-chromosome. While it is known that one allele at most X-linked genes is silenced in females by inactivation of one entire chromosome [34],[35], we would expect that a polyclonal cell population (in which half of the cells inactivate one X chromosome and the other 50% inactivate the alternate X chromosome) would give a similar level of expression for both alleles. However, all X-linked genes on our two SNP panels (22 SNPs in 12 genes) were among the top 5% of genes with most dramatic allelic imbalance patterns. The extent of allelic imbalance at a given gene varies among individual LCLs but interestingly, the patterns of allelic imbalance are very consistent across genes for a given individual (Figure S7). Additionally, the inheritance of the expressed allele (determined, when possible, using the pedigree information for the two families included in this study) appeared random. It has been previously proposed that the extent of clonality of a cell line could explain the patterns of allelic imbalance at genes with random mono-allelic expression [30]: clonal cells will all have the same X chromosome inactivated and thus display very high ratios of allelic imbalance. In contrast, cell-lines composed of a polyclonal population of lymphoblasts will have one or the other of their X chromosomes inactivated in different cells and thus an apparent expression of both alleles (i.e., a low extent of differential allelic expression). Our observations at X-linked genes are consistent with this hypothesis and the biased clonality of these LCLs, which were created over 20 years ago and passaged numerous times (see also [30]).
The two autosomal genes displaying the most dramatic allelic imbalance patterns have previously been shown to be imprinted in humans: PEG10 [36] and SNRPN [37]. In addition, KCNQ1, MEST and ZNF215 which are imprinted in humans [38]–[40] also show significant differences in allelic expression (Table 1). The mode of inheritance of the expressed allele also corresponds, in each case, to what has been described for the expression of these genes: for PEG10 and SNRPN, heterozygous individuals express the paternally-inherited allele (i.e. maternally imprinted) while for KCNQ1 the maternally-inherited allele is expressed. Our limited pedigree information is not conclusive for MEST and ZNF215. The only other known imprinted gene analyzable in our panel, PLAGL1 [41],[42] did not pass the significance threshold (i.e. an allelic ratio greater than 60∶40) but shows a population average allelic imbalance larger than 55∶45 and a high correlation between the two SNPs analyzable in the panel (rs2076684 and rs9373409); therefore it likely represents a significant difference in allelic expression.
The 83 remaining genes (103 SNPs) with significant population-average allelic imbalance included several genes for which allelic imbalance had been shown in previous studies (e.g. IL1A or IGF1 described in [30]). For some genes (e.g., CHI3L2), one allele/haplotype is clearly expressed more than the other in heterozygotes and the inheritance pattern in families supports a genetic cause for allelic imbalance. For other genes, neither the direction of allelic imbalance nor the pedigree analysis allowed us to easily differentiate the genetic/epigenetic cause of the differential allelic expression (Table 1). For 56 genes with significant differences in allelic expression we tested whether differential allelic expression could be statistically associated to one of the SNP in the vicinity of the gene genotyped by the HapMap project (see Materials & Methods for details). The results of these tests for SERPINB10 and ABCG1 are shown on Figure 7 and the strongest nominal association for each gene is displayed on Table 2. Twenty-three genes still display statistical significant associations after Bonferroni correction for multiple testing (highlighted in green on Table 2) showing a clear enrichment relative to the 2–3 associations expected by chance. Our power to detect a significant association between a HapMap SNP and the under-/over-expressing chromosome in this setting is low due to our reduced sample size (only the heterozygous individuals are taken into account in this analysis) and the number of regulatory haplotypes identified is thus likely underestimated. Additionally, many SNPs are tested for each gene and it is thus possible that some of the regulatory haplotypes result from spurious associations (i.e. they are false positives). One argument against a very high rate of false positive in our analysis is that imprinted genes such as MEST or PEG10 do not show any signal of association (Figure S8) consistent with the fact that the cis-regulatory mechanism at these genes is not encoded in the DNA sequence. To further investigate the validity of our association, we attempted to independently confirm these regulatory haplotypes by testing for the statistical association between one SNP in the regulatory haplotype and gene expression level. We used gene expression measurements performed at the Wellcome Trust Sanger Institute (kindly provided by M. Dermitzakis) on the same individual cell lines assayed by Illumina gene expression arrays. For each gene, we tested whether the homozygotes for the regulatory haplotype associated with low allelic expression in heterozygotes show a significantly lower gene expression level than the homozygous individuals for the regulatory haplotype associated with high allelic expression. We also performed locus-specific RT-PCR and quantified the level of gene expression using SYBR-Green for eleven genes for which differential allelic expression was significantly associated with allelic expression but for which expression data were not available (2 genes) or genes with strong association with a regulatory haplotype but that were not validated using the Sanger dataset (9 genes). Overall, out of the 47 genes with a significant association between a SNP (or several, defining the regulatory haplotype) and differential allelic expression at the nominal cut-off, 10 were confirmed using gene expression measurements while 5 other genes showed a trend but did not reach statistical significance (Table 2).
We analyzed differences in relative allelic expression (or allelic imbalance) at 1,380 human genes using 2,968 SNPs and more than 80 lymphoblastoid cell lines from individuals with European ancestry. Using quantitative sequencing we validated our results for a subset of genes and showed that the experimental variability in both settings is low and that the Illumina ASE assay and quantitative sequencing of RT-PCR products yield reproducible estimates of allelic imbalance consistent with each other. Overall, the experimental noise is much lower than the difference in relative allelic expression observed at many loci and therefore cannot be responsible for it. Additionally, the high concordance of the results obtained using different SNPs in the same transcript supports our findings that allelic imbalance, as we estimated it, is not an experimental artefact but reflects inherent biological differences in the relative expression of both alleles in heterozygous individuals. We also showed that lymphoblastoid cell lines, despite being simplified biological materials, are suitable resources to investigate mechanisms of gene regulation. Here, we demonstrated that our estimation of allelic imbalance is little affected by growth conditions and that LCLs harvested from different passages yield very similar results. Finally, the results efficiently recapitulate the consequences of the epigenetic mechanisms established in the individuals from which the cells have been derived (see also [43]). We are therefore confident that, overall, the patterns of allelic imbalance we observed are neither experimental artifacts, nor specific to the material studied, but represent a common biological phenomenon affecting human gene expression.
We showed that LCLs derived from female individuals still harbor the consequences of X-inactivation at all X-linked genes investigated, with one allele being transcriptionally silenced [34]. The extent of allelic imbalance detected at X-linked genes can vary among LCLs due to the various degrees of clonality of these cells but clonal LCLs consistently show complete silencing of one allele at all X-linked genes investigated (Figure S7). In addition, imprinting, established in the germ lines of the parents of the individuals from which the cells are derived [44], is also maintained in LCLs. In our experiments, PEG10, SNPRN, MEST and KCNQ1 show reduced or absent expression of one allele and, when the mode of inheritance can be determined, it corresponds to the imprinting mechanism described in the literature (i.e. PEG10 and SNPRN are maternally imprinted, KCNQ1 is paternally imprinted). We thus observe extensive differential allelic expression (i.e. allelic ratio larger than 70∶30) for all genes whose expression is known to be epigenetically regulated. This clearly shows that analysis of differential allelic expression is a suitable method for identifying the consequences of epigenetic mechanisms of gene regulation. The Illumina ASE assay would thus provide an efficient method to screen tumor tissues and identify patterns of differential allelic expression resulting from aberrant methylation or loss of imprinting that are known to be involved in the etiology of cancers [45]–[47].
Interestingly, IMPACT which shows significant extent of allelic imbalance at two SNPs (rs677688 and rs1053474) in our study, is known to be imprinted in mice [48] but not in humans [49]. The mode of inheritance of the over-expressed alleles could not be determined using the two families available in our study (i.e. the parents were always homozygous for the same allele). The attempt to map differential expression to a regulatory haplotype was not successful and is consistent with an epigenetic mechanism of gene regulation. More investigations are required to determine whether the pattern of allelic imbalance observed for IMPACT results from incomplete silencing of one allele following imprinting in the parental germ-lines or whether it results from random mono-allelic expression or another mechanism of gene expression regulation.
Our analysis of 643 genes expressed in LCLs shows that, for a large proportion of them (∼20%), the two alleles are differentially expressed in most heterozygous individuals. For 18 genes, differential expression resulted from a known epigenetic silencing of one of the two alleles, either through X-inactivation in females or imprinting. The mechanisms leading to allele-specific expression at all other genes could be driven by a polymorphism affecting the cis-acting regulation (e.g. a SNP in a transcription factor or a miRNA biding site) or simply result from random silencing of one of the two alleles. We tested 56 genes for association of differential allelic expression patterns observed with a cis-acting regulatory polymorphism using genotypes generated by the HapMap project (see Materials and Methods for details). For 23 of these genes we identified a region statistically associated with differences in allele expression that could indicate the existence of a regulatory haplotype (i.e., a region of one chromosome likely containing the polymorphism(s) causing the differential cis-regulation). These regions are often tens of kb long, consistent with previous descriptions of the linkage disequilibrium patterns in humans [50]. Although this approach does not identify the actual polymorphism(s) responsible for the differential cis-regulation, examination of these regulatory haplotypes provides some valuable insights on the mechanisms leading to differential expression and can guide future investigations. For example, the regulatory haplotype for GAS7 is almost exclusively restricted to the 3′UTR of the gene and may indicate that the patterns of allelic imbalance observed are due to differential mRNA processing, stability or the presence of a 3′ enhancer. In contrast, the regulatory haplotype identified for MGC33648 is located in the 5′ region and does not seem to overlap with the gene itself. This might be indicative of alternative promoter usage or differential transcription efficiency (e.g. due to differential transcription factor binding site affinity).
Several recent studies have used large-scale associations between gene expression and extensive genotype information to investigate gene regulation in humans, some of them using cell lines included in our study. In particular, Stranger and colleagues analyzed 630 genes located in ENCODE regions, on chromosome 21 and in one portion of chromosome 20. They found evidence of cis-acting regulation for 63 genes [19]. 2005). We were able to analyze 21 of these genes in our experiment. Six of them also showed evidence of cis-acting regulation (e.g. SERPINB10 or TSGA2) in our study while a seventh gene (TTC3) showed patterns consistent with differential allelic expression but did not reach our significance threshold. The remaining 14 genes did not show evidence of differential allelic expression in our analysis. Alternatively, we identified 10 new genes located in ENCODE region or chromosome 21 that showed significant level of differential allelic expression but were not detected in the Stranger study. Several non-exclusive reasons could explain the discrepancies between the results of the two approaches. First, it is worth noting that, even if the same individuals are analyzed by allelic-specific expression and gene expression association, the power to detect cis-acting effect differs depending on the allele frequency of the marker used: in gene expression association analysis all individuals are analyzed but the power in the regression analysis depends on their genotypes (e.g. the genotypes AA, AB and BB are encoded in the linear regression as 0, 1 and 2) while in allelic expression analysis only the individuals heterozygotes at the marker considered are analyzed. This can become particularly problematic to study differential allelic expression at some genes since it requires a relatively common exonic SNP to detect allelic imbalance. In this context, it is worth noting that intronic SNPs can successfully be used for genes that are highly expressed (see also [30]). Second, associations of gene expression to genotypes depends greatly on the linkage disequilibrium (LD) patterns and requires extensive genotype information from all the individuals in order to include one marker in LD with the regulatory polymorphism. Allelic expression, on the other hand, directly investigate cis effect directly at the gene level and thus only requires physical link between the gene and the regulatory polymorphism affecting it (i.e. they need to be on the same chromosome). Finally, the differences between allelic expression and gene expression mapping might indicate that some genes are also regulated by trans-acting mechanisms that differ among individuals: differential allelic expression is influenced only by cis-acting mechanisms of gene regulation while gene expression is influenced by cis- and trans-acting gene regulation. It is thus not unlikely that individual differences in trans-acting regulation swamp the signal from cis-acting polymorphisms. In this context, it is noteworthy that total gene expression mapping has been much more successful in mice and yeast for which the genetic heterogeneity is much lower and can be controlled (reviewed in [9],[10],[51]). In humans, or in any other outbred population, genetic heterogeneity greatly limits the identification of cis-acting mechanisms using gene expression data while measurements of differential allelic expression are unaffected.
We showed here that allelic expression assays are complementary from gene expression mapping and that the Illumina ASE assay overcomes two of the major limitations and criticisms of the former methodologies used to assess differential allelic expression: it allows a robust and high-throughput estimation of allelic imbalance: it is now possible to reliably screen hundreds of RNAs for several hundreds of genes in a couple of days. Additionally, when several SNPs can be used to assess differential allelic expression, the assay becomes very robust since each marker provides an independent estimation and one can test the correlation among estimates obtained at different positions. It is worth noting here that since this assay relies on the comparison of allelic ratio in DNA and RNA of each individual, it internally controls for the existence of polymorphisms in the primer sites or copy number variation encompassing the gene studied (that will affect equally DNA and RNA). Likely, the greatest advantage of the analysis of differential allelic expression over total gene expression is its flexibility. To identify differential regulation of gene expression using total gene expression, one needs extensive genotype information to test whether, at any polymorphic position, the gene expression differences among individuals segregate according to their genotype. This precludes a quick assessment of the expression of one locus in one cohort of particular interest or using a specific tissue. In contrary, differential allelic expression offers the advantage that any one gene can be quickly assessed in any cohort or tissue by simply comparing the expression of the two alleles in each individual (the amount of genetic information recently made available by the HapMap project allows a quick and easy selection of markers likely to be polymorphic for a given gene). The determination of regulatory haplotypes would still require extensive information concerning surrounding polymorphisms but the initial screening to determine whether one transcript is differentially cis-regulated can be done very efficiently with a handful of markers.
We showed that differential allelic expression is a robust approach to identify cis-acting mechanism of gene regulation. It complements gene expression association studies and offers additional perspectives, notably on epigenetic mechanisms of gene regulation. It could thus be particularly interesting to apply this assay to tumors to detect mis-regulated genes due to aberrant methylation patterns or loss of imprinting. In addition, our approach is applicable to any new cohort or tissue since it is self-sufficient to identify differential cis-regulation and does not require additional genotyping. It can be easily used to follow-up interesting non-coding regions associated to a particular disease and test if they are involved in the etiology of the disease through some regulatory effects on neighboring genes.
83 lymphoblastoid cell lines (LCL) derived from blood samples from the CEPH collection were selected for this project. They included 60 unrelated individuals obtained from Utah residents with ancestry from western and northern Europe for which DNA was genotyped for millions of SNPs covering the entire genome by the International HapMap Project. Additionally, 21 LCLs from CEPH pedigrees 1420 and 1444 were included to provide complete information on two three-generation CEPH families. Cells were grown at 37°C and 5% CO2 in RPMI 1640 medium (Invitrogen, Burlington, Canada) supplemented with 15% heat-inactivated fetal bovine serum (Sigma-Aldricht, Oakville, Canda), 2 mM L-glutamine (Invitrogen, Burlington, Canada) and penicillin/streptomycin (Invitrogen, Burlington, Canada). The cell growth was monitored with a hemocytometer and the cells were harvested when the density reached 0.8–1.1 × 106 cells/mL. Cells were then resuspended and lysed in TRIzol reagent (Invitrogen, Burlington, Canada). For all LCLs, three successive growths were performed (corresponding to the 2nd, 4th and 6th passages) after thawing frozen cell aliquots.
We estimated allelic imbalance at 1,380 genes (two panels of ∼1,500 SNPs, Figure 1) using the Illumina ASE assay (Figure 2). The experiment is similar to the one used for large-scale SNP genotyping [52] and gene expression profiling [53] except that DNA and RNA are independently assessed and compared to each other. RNA was first converted into biotinylated cDNA [53] while DNA was treated according to the usual GoldenGate assay protocol [52]. Biotinylated DNA (derived from genomic DNA or mRNA) was immobilized on paramagnetic beads and pooled SNP-specific oligonucleotides were annealed on the DNA. Hybridized oligonucleotides were then extended and ligated to generate DNA templates, which were amplified using universal fluorescently-labeled primers. Finally, single-stranded PCR products were hybridized to a Sentrix Array Matrix [52], and the arrays were imaged using the BeadArray Reader Scanner [54]. 96 samples (DNA or RNA) were analyzed per Sentrix Array for ∼1,500 SNPs. All RNA measurements were performed in duplicates.
To estimate the extent of allelic imbalance in heterozygote individuals at each SNP of the Illumina ASE panel, we developed algorithms using two different approaches: i) we used information from individuals of all three genotypes (AA, AB and BB), and/or ii) we used only the heterozygote individuals.
We first determined whether a given gene was expressed above a determined background in a given individual. To do so, we made use of the fact that the genotypes were known (from the DNA analysis) and developed a locus-specific expression background cut-off: homozygote individuals (i.e. AA or BB) can only express the corresponding allele, respectively A or B, at the RNA level (if at all). We thus determined a background fluorescence level (i.e. corresponding to random noise) for each allele (i.e. A and B) by measuring the emission in the corresponding dye (respectively, Cy3 and Cy5) in individuals homozygous for the other allele (respectively BB and AA). This is represented schematically on Figure S9. To avoid false positive results due to the inclusion of transcripts not expressed in the cell lines considered, we used a conservative approach and arbitrarily fixed the background emission cutoff to the maximum emission of the absent allele of all homozygotes, plus the mean emission of the absent allele divided by the number of homozygotes (to weight the uncertainty in the determination of the “maximum noise” by the numbers of individuals used to determine it). This procedure allowed us to independently estimate the background emission of each allele/dye specifically for each SNP, which is particularly important because the fluorescence emission can differ drastically between the dyes and among loci (data not shown). We then proceeded to the detection call using the background cut-offs: individuals with genotypes AA were considered to express a given transcript if the emission was larger than twice the cutoff background emission of A, individuals with genotypes AB if the fluorescence was larger than the sum of the background emission of A and the background emission of B, and individuals with genotypes BB if the emission is larger than twice the background emission of B. Since the inclusion in the analyses of transcripts expressed at low level (or not expressed at all) is very problematic, we excluded from our analyses all loci for which less than 75% of the individuals had discordant replicate expression (i.e., one replicate above expression background, the other under the cut-off value).
The first method used to determine whether some heterozygote individuals expressed significantly differently the two alleles is locus-specific but requires having at least one individual expressed from each homozygote genotype (AA and BB). In this case, we determined the median log ratio of the two dyes for each homozygote clusters at the DNA and RNA level () as well as the median absolute deviations (MAD). We used medians and MADs, instead of means and standard deviations, to down weight the influence of possible outliers. We then determined a range of “expected” (i.e. non significant) variation of allelic expression for the heterozygote individuals. We calculated the equation of the lines joining the median values plus/minus two MAD of AA and BB and estimated the range, for the log ratio of the dyes at the RNA level, between the lines at the value corresponding to the median of DNA in heterozygote individuals (Figure S10). If the observed log ratio of dyes for a given heterozygote individual fell outside the expected range of variation in absence of AI (Figure S8), we scored each heterozygote individual separately to obtain a quantitative estimation of allelic imbalance using the ratio:This simple estimate indicates both the magnitude of the allelic imbalance (i.e. the fold difference) and its direction (i.e. which allele is more expressed than its counterpart).
In order to assess allelic imbalance for SNPs with low minor allele frequencies (for which homozygote individuals with the minor allele may not be present in a small sample size panel), we developed a second method based solely on the heterozygote individuals. If a given transcript is affected by allelic imbalance we expect that either the variance of the log ratio of dyes for heterozygote RNAs to be greatly increased relative to the variance of homozygote RNAs, or, if one allele is systematically more expressed than the other, the mean value of these log ratios to be drastically shifted from its expected intermediate position (between the mean for AA and the mean for BB homozygote RNAs). For all SNPs with at least five individuals with the same genotype expressed, we estimated the standard deviation of the log ratio of dyes for DNA and RNA. The distribution of the log ratio of the standard deviations (i.e. log σDNA/σRNA) over all loci for heterozygous individuals differed from those observed using homozygous individuals and did not seem to fit a normal distribution (Figure S11). Based on the assumption that this distribution may include some loci in allelic imbalance (and thus with a higher than expected RNA variance), we fitted a mixture of two Gaussians on our dataset (i.e., one corresponding to the loci with allelic imbalance, the second including all other loci) using a Maximum Expectation algorithm implemented in R (mixdist package). For our data, the best fit was obtained with a minor distribution (including ∼3% of the loci) corresponding to the most extremely negative log ratios of variances (i.e., that the RNA standard deviation was larger than expected). For each locus, we then used the probability of belonging to the “higher-than-expected RNA variance” distribution as an indication of allelic imbalance.
We assessed the extent of allelic imbalance by quantitative sequencing following the method described in Ge et al. [33]. Briefly, we isolated RNA using TRIzol reagent following the manufacturer's instructions. We assessed RNA quality with an Agilent 2100 Bioanalyzer (Agilent, Palo Alto, USA) before synthesizing first strand cDNA using random hexamers (Invitrogen, Burlington, Canada) and Superscript II reverse transcriptase (Invitrogen, Burlington, Canada). For each locus, we designed locus-specific primers, in the exon/UTR containing the SNP analyzed, at least 50 bp away from the SNP studied. 5 ng of genomic DNA and 10 ng of total cDNA were then amplified by PCR using Hot Start Taq Polymerase (Qiagen, Mississauga, Canada) with an activation step (95°C for 15 minutes) followed by 40 cycles (95°C for 30 s, 55°C for 30 s and 72°C for 45 s) and a final extension step (72°C for 6 minutes). PCR products were purified using Exonuclease I and Shrimp Alkaline Phosphatase (USB, Cleveland, USA) and sequenced using either one of the former primers or a nested primer, on an Applied Biosystems 3730xl DNA analyzer. We used PeakPeaker v.2.0 [33] with the default settings to quantify the relative amount of the two alleles measured from the chromatogram after peak intensity normalization. To estimate the experimental variability of the entire experimental setup we used a hierarchical strategy for two genes (cf. Figure S12): for two/three individual cell lines, we extracted independently RNAs three times and performed, on each extract, three independent RT-PCRs. All cDNA obtained were then split into three aliquots, each amplified independently by locus-specific PCR. These PCR products were finally sequenced each three times (i.e. three independent sequencing reactions). To estimate the variability at each experimental stage we calculated the mean standard variation normalized to the mean using the independent triplicates. To calculate the variance in the higher hierarchical levels (PCR, RT-PCR), we averaged the values from the lower level (e.g., to estimate the variability at the PCR level, we compared the means of the three sequencing values performed on each of the three PCRs: [s1,s2,s3] vs [s4,s5,s6] vs [s7,s8,s9]). The results are presented in Text S1.
We attempted to map allelic imbalance to regulatory haplotypes for all genes with significant differences in allelic expression that fulfilled these criteria: i) they are mapped on the build 34 of the human genome, ii) the SNP used in the Illumina ASE assay has also been genotyped by the HapMap [55] and iii) there are more than four HapMap individuals heterozygous at the marker SNP. For each gene, we retrieved the haplotype information from the phased chromosomes of each of the 57 HapMap CEPH individuals for 100,000 bp upstream and downstream of the SNP used to assess allelic imbalance. When a transcript contains more than one SNP or if two SNPs used to assess allelic imbalance at two transcripts are separated by less than 200,000 bp, the region retrieved spans from the most upstream marker plus 100,000 bp to the most downstream marker minus 100,000 bp. For each individual LCL, the over expressed and under expressed haplotype/chromosome were identified and each SNP was tested for segregation of the alleles in under- and over-expressed chromosomes using a Fischer's exact test. Between 47 and 592 SNPs were tested for each gene (mean = 229) and the associations remaining significant after Bonferroni correction for multiple testing are shown in green in Table 2.
Illumina total gene expression data were obtained from the Wellcome Trust Sanger Institute for the 60 unrelated CEPH individuals genotyped by the HapMap project and included in our assay. We also determined the total expression for 10 genes using Real-Time PCR and SYBR Green labeling on an ABI 7900HT (Applied Biosystems, Foster City, CA) instrument. 8–10 ng of first strand cDNA were amplified using 0.32 µM of gene specific primers and Power SYBR Green PCR master mix (Applied Biosystems) according to the manufacturer's instructions. The amplifications started by 95°C for 10 min followed by 40 cycles at 95°C for 20 s, 58°C for 30 s and 72°C for 45 s. We performed the Real-Time PCR assays for the 60 individuals LCLs genotyped by the HapMap projects and analyzed 6 replicates per each sample. A standard curve was established using a dilution series of total cDNA of known concentration. The Ct for each replicate was transformed to a relative concentration using the estimated standard curve function (SDS 2.1, Applied Biosystems) and normalized based on 18S rRNA Taqman (Applied Biosystems) expression data obtained for each sample to account for well to well variability.
All analysis scripts are available upon request. PeakPicker v.2.0 is available at http://www.genomequebec.mcgill.ca/EST-HapMap/. |
10.1371/journal.pbio.0050035 | Chemically Diverse Toxicants Converge on Fyn and c-Cbl to Disrupt Precursor Cell Function | Identification of common mechanistic principles that shed light on the action of the many chemically diverse toxicants to which we are exposed is of central importance in understanding how toxicants disrupt normal cellular function and in developing more effective means of protecting against such effects. Of particular importance is identifying mechanisms operative at environmentally relevant toxicant exposure levels. Chemically diverse toxicants exhibit striking convergence, at environmentally relevant exposure levels, on pathway-specific disruption of receptor tyrosine kinase (RTK) signaling required for cell division in central nervous system (CNS) progenitor cells. Relatively small toxicant-induced increases in oxidative status are associated with Fyn kinase activation, leading to secondary activation of the c-Cbl ubiquitin ligase. Fyn/c-Cbl pathway activation by these pro-oxidative changes causes specific reductions, in vitro and in vivo, in levels of the c-Cbl target platelet-derived growth factor receptor-α and other c-Cbl targets, but not of the TrkC RTK (which is not a c-Cbl target). Sequential Fyn and c-Cbl activation, with consequent pathway-specific suppression of RTK signaling, is induced by levels of methylmercury and lead that affect large segments of the population, as well as by paraquat, an organic herbicide. Our results identify a novel regulatory pathway of oxidant-mediated Fyn/c-Cbl activation as a shared mechanism of action of chemically diverse toxicants at environmentally relevant levels, and as a means by which increased oxidative status may disrupt mitogenic signaling. These results provide one of a small number of general mechanistic principles in toxicology, and the only such principle integrating toxicology, precursor cell biology, redox biology, and signaling pathway analysis in a predictive framework of broad potential relevance to the understanding of pro-oxidant–mediated disruption of normal development.
| Discovering general principles underlying the effects of toxicant exposure on biological systems is one of the central challenges of toxicological research. We have discovered a previously unrecognized regulatory pathway on which chemically diverse toxicants converge, at environmentally relevant exposure levels, to disrupt the function of progenitor cells of the developing central nervous system. We found that the ability of low levels of methylmercury, lead, and paraquat to make progenitor cells more oxidized causes activation of an enzyme called Fyn kinase. Activated Fyn then activates another enzyme (c-Cbl) that modifies specific proteins—receptors that are required for cell division and survival—to initiate the proteins' degradation. By enhancing degradation of these receptors, their downstream signaling functions are repressed. Analysis of developmental exposure to methylmercury provided evidence that this same pathway is activated in vivo by environmentally relevant toxicant levels. The remarkable sensitivity of progenitor cells to low levels of toxicant exposure, and the discovery of the redox/Fyn/c-Cbl pathway as a mechanism by which small increases in oxidative status can markedly alter cell function, provide a novel and specific means by which exposure to chemically diverse toxicants might perturb normal development. In addition, the principles revealed in our studies appear likely to have broad applicability in understanding the regulation of cell function by alterations in redox balance, regardless of how they might be generated.
| Determining whether chemically diverse substances induce similar adverse effects at the cellular and molecular level is one of the central challenges of toxicological research. If the structural diversity of different toxicants, and of potential toxicants, means that each works through distinctive mechanisms then this creates a potentially unsolvable challenge in developing means of screening the many tens of thousands of different chemicals for which little or no toxicological information exists. In contrast, the identification of general principles that transcend the specific chemistries of individual substances has the potential of providing broadly relevant insights into the means by which toxicants disrupt normal development. If such principles were found to apply to the analysis of toxicant levels frequently encountered in the environment, this would be of even greater potential importance in providing efficient means of analyzing this diverse array of chemicals.
Of all of the effects associated with toxicant exposure, one of the few that appears to be common to multiple chemically diverse substances is the ability of these agents to cause cells to become more oxidized. The range of toxicants reported to alter oxidative status is very broad, and includes metal toxicants such as methylmercury (MeHg; e.g., [1–6], lead [Pb] [6–9], and organotin compounds [1,2,5,10,11]), cadmium [12,13], and arsenic [12,14]. Ethanol exposure also is associated with oxidative stress [15], as is exposure to a diverse assortment of agricultural chemicals [16], including herbicides (e.g., paraquat [17,18]), pyrethroids [19–21], and organophosphate and carbamate inhibitors of cholinesterase [22–26]). Thus, the ability to cause cells to become more oxidized is shared by many toxicants, regardless of their chemical structure.
The observations that chemically diverse toxicants share the property of making cells more oxidized is of particular interest in light of the increasing evidence that oxidative regulation is a central modulator of normal physiological function. Although increases in oxidative status in a cell have been most extensively studied in the context of their adverse effects (in particular, the induction of cell death or of cell senescence), multiple studies have demonstrated that changes in redox state as small as 15%–20% may be critical in regulating such normal cellular processes as signal transduction, division, differentiation, and transcription (reviewed in, e.g., [27–31]. Although the mechanistic basis for such regulation is frequently unclear, the importance of redox status in modulating cell function makes convergence of different toxicants on this physiological parameter a matter of considerable potential interest.
Despite the observations that many toxicants share the property of making cells more oxidized, multiple questions exist regarding the relevance of such observations for the understanding of toxicant function.
First, there is considerable uncertainty about the relative importance of effects on redox state in the analysis of individual toxicants, and it is generally believed that the major effects of toxicants on cellular function are distinct from any effects on oxidative status. For example, in the context of agents analyzed in the present studies, MeHg-mediated effects on cellular function generally are thought to be mediated through binding to cysteine residues, thus disrupting function of microtubules and other proteins, but may also involve disruption of Ca2+ homeostasis (e.g., [32–34]). In contrast, Pb does not bind to cysteine residues and instead is thought to exert its functions through altering normal calcium metabolism by mimicking calcium action and/or by disrupting calcium homeostasis (e.g., [35,36]). This would lead to alterations in function of multiple proteins, of which the most extensively studied have been members of the protein kinase C (PKC) family of enzymes (e.g., [37,38]).
A further concern is the general lack of knowledge about whether, or where, oxidation induced by different means would mechanistically converge. For example, MeHg has been suggested to cause oxidative stress by a variety of mechanisms, including by binding to thiols, by causing a depletion in glutathione levels, or by impairing mitochondrial function [39,40], whereas Pb is thought to disrupt mitochondrial function through its effects on calcium metabolism (e.g., [35,36,41–45]) The organic herbicide paraquat (the third agent examined in the present studies) is another example of a toxicant with pro-oxidant activities, but in this case, resulting from initiation of a cyclic oxidation/reduction process in which paraquat first undergoes one electron reduction by NADPH to form free radicals that donate their electron to O2, producing a superoxide radical; upon exhaustion of NADPH, superoxide reacts with itself and produces hydroxyl free radicals (e.g., [17,18]). Whether these different means of altering oxidative state would have different mechanistic consequences is unknown.
A further concern regarding the hypothesis that changes in redox state represent an important convergence point of toxicant action is whether oxidative changes are even associated with toxicant exposure at levels frequently encountered in the environment. For example, although several studies have documented the ability of MeHg to cause cells to become more oxidized, effective exposure levels employed in these studies have generally ranged from 1–20 μM [2–5], which is 30–600 times the upper range of average mercury concentrations found in the bloodstream of as many as 600,000 newborn infants in the United States alone [46]. Similar concerns apply to the analysis of multiple toxicants, for which pro-oxidant effects have largely been studied at exposure levels much higher than those with broad environmental relevance.
In addition, a more general concern regarding the search for general principles of toxicant action is whether such convergence, if it exists, would occur only at exposure levels that induce cell death or whether common mechanisms might be relevant to the understanding of more subtle effects of toxicant exposure, particularly during critical developmental periods. Because development is a cumulative process, the effects of small changes in, e.g., progenitor cell division and/or differentiation, that are maintained over multiple cellular generations could have substantial effects on the organism. Such changes are poorly understood, however, at both cellular and molecular levels.
Our present studies have led to the discovery of a previously unrecognized regulatory pathway on which environmentally relevant levels of chemically diverse toxicants converge to compromise division of a progenitor cell isolated from the developing central nervous system (CNS). We found that exposure of cells to low levels of MeHg, Pb, or paraquat is sufficient to make cells more oxidized and to activate Fyn kinase, a Src family member known to be activated by increased oxidative status. This first step activates a pathway wherein Fyn activates c-Cbl, a ubiquitin ligase that plays a critical role in modulating degradation of a specific subset of receptor tyrosine kinases (RTKs). c-Cbl activation in turn leads to reductions in levels of target RTKs, thus suppressing division of glial progenitor cells. The effects of all three toxicants are blocked by co-exposure to N-acetyl-L-cysteine, which is widely used to protect against oxidative stress. We also provide evidence that our in vitro analyses successfully predict previously unrecognized effects of developmental MeHg exposure at levels 90% below those previously considered to represent low-dose exposure levels.
The progenitor cells that give rise to the myelin-forming oligodendrocytes of the CNS offer multiple unique advantages for the study of toxicant action, particularly in the context of analysis of toxicant effects mediated by changes in intracellular redox state. These progenitors (which are referred to as both oligodendrocyte-type-2 astrocyte [O-2A] progenitor cells ([47] and oligodendrocyte precursor cells, here abbreviated as O-2A/OPCs) are one of the most extensively studied of progenitor cell populations (reviewed in, e.g., [48–52]. They also are among a small number of primary cell types that can be analyzed as purified populations, and at the clonal level, and for which there is both extensive information on the regulation of their development and also evidence of their importance as targets of multiple toxicants (including such chemically diverse substances as Pb [38,53], ethanol [e.g., [54–57]], and triethyltin [10,58]).
Another important feature of O-2A/OPCs, in regard to the present studies, is that their responsiveness to small (∼15%–20%) changes in the intracellular redox state provides a central integrating mechanism for the control of their division and differentiation [59]. O-2A/OPCs purified from developing animals on the basis of the cell's intracellular redox state exhibit strikingly different propensities to divide or differentiate. Cells that are more reduced at the time of their isolation undergo extended division when grown in the presence of platelet-derived growth factor (PDGF, the major mitogen for O-2A/OPCs [60–62]), whereas those that are more oxidized are more prone to undergo differentiation [59]. Pharmacological agents that make cells slightly more reduced enhance self-renewal of dividing progenitors, whereas pharmacological agents that make cells more oxidized, by as little as 15%–20%, suppress division and induce oligodendrocyte generation. Moreover, cell-extrinsic signaling molecules (e.g., neurotrophin-3 [NT-3] and fibroblast growth factor-2 [FGF-2]) that enhance the self-renewal of progenitors dividing in response to PDGF cause cells to become more reduced. In contrast, signaling molecules that induce differentiation to oligodendrocytes (i.e., thyroid hormone [TH] [63,64]) or astrocytes (i.e., bone morphogenetic protein-4 [BMP-4] [65,66]) cause cells to become more oxidized [59]. The ability of these signaling molecules to alter redox state is essential to their mechanisms of action, because pharmacological inhibition of the redox changes they induce blocks their effects on either division or differentiation of O-2A/OPCs. Thus, multiple lines of evidence have demonstrated that responsiveness to small changes in redox status represents a central physiological control point in these progenitor cells (as summarized in Figure 1).
We initiated our studies of toxicant effects on O-2A/OPCs with an examination of MeHg, which has been previously studied for its effects on neuronal migration, differentiation, and survival, and on astrocyte function (e.g., [67–74]). Little is known about the effects of MeHg on the oligodendrocyte lineage, despite the fact that there are several reports over the past two decades documenting decreases in conduction velocity in the auditory brainstem response (ABR) of MeHg-exposed children [75–78] and rats [79]. Such a physiological alteration has long been considered to be indicative of myelination abnormalities in children whose development has been compromised by iron deficiency (see, e.g., [80,81]).
We found that exposure of O-2A/OPCs (growing in chemically defined medium supplemented with PDGF) to environmentally relevant levels of MeHg makes these cells approximately 20% more oxidized (Figure 2A), a degree of change similar to that previously associated with reductions in progenitor cell division [59]. Exposure to MeHg inhibited progenitor cell division as determined both by analysis of bromodeoxyuridine (BrdU) incorporation (Figure 2B) and by analysis of cell division in individual clones of O-2A/OPCs (Figure 2C–2E). These oxidizing effects of MeHg were seen at exposure levels as low as 20 nM, less than the 5.8 μg/l or more (i.e., parts per billion [ppb]) of MeHg found in cord blood specimens of as many as 600,000 infants in the US each year [46] and 0.3% or less of the exposure levels previously found to induce oxidative changes in astrocytes [4]. Exposure to 20 nM MeHg was sufficient to cause an approximately 25% drop in the percentage of O-2A/OPCs incorporating BrdU in response to stimulation with PDGF. When examined at the clonal level, MeHg exposure was associated with a reduction in the number of large clones and an increase in the number of small clones, as seen for other pro-oxidant stimuli [59]. Increasing MeHg exposure levels above 50 nM was associated with significant lethality, but little or no cell death was observed at the lower concentrations used in the present studies (unpublished data). Thus, division of O-2A/OPCs exhibits a striking sensitivity to low concentrations of MeHg.
One possible explanation for the reduced division associated with MeHg exposure would be disruption of PDGF-mediated signaling, and molecular analysis revealed that exposure of O-2A/OPCs to 30 nM MeHg for 24 h suppressed PDGF-induced signaling pathway activation at multiple points from the nucleus back to the receptor. One pathway stimulated by PDGF binding to the PDGF receptor-α (PDGFRα) leads to sequential activation of Raf-1, Raf-kinase, and extracellular signal-regulated kinase 1 and 2 (ERK1/2), which further leads to activation of the Elk-1 transcription factor and up-regulation of immediate early-response gene expression, at least in part through activation of the serum response element (SRE) promoter sequence [82,83]. MeHg exposure was associated with reduced expression of an SRE-luciferase reporter gene (Figure 3A), and reduced ERK1/2 phosphorylation (Figure 3B). PDGFRα activation also stimulates activity of PI-3 kinase, leading to activation of Akt and induction of NF-κB–mediated transcription (e.g., [82,84,85]), both of which also were inhibited by MeHg exposure. Expression of an NF-κB-luciferase reporter gene was decreased (Figure 3C), as was phosphorylation of Akt (Figure 3D). Phosphorylation of PDGFRα, indicating receptor activation, was also reduced in cells exposed to MeHg (Figure 3E). Because O-2A/OPCs growing in these cultures are absolutely dependent upon PDGF for continued division (e.g., [60,61,86]), the suppression of PDGF signaling would necessarily cause a reduction in cell division.
We next found that the effects of MeHg were pathway specific and were associated with reductions in total levels of PDGFRα. O-2A/OPCs exposed to 30 nM MeHg exhibited no reduction in ERK1/2 phosphorylation induced by exposure to NT-3 (Figure 4A), and no reduction in NT-3–induced expression from an SRE-luciferase reporter construct (unpublished data). This result suggested that the site of action of MeHg was upstream of ERK1/2 regulation, prompting us to look directly at the PDGFRα. We found that the reduction in phosphorylated PDGFRα (Figure 3E) was paralleled by a reduction in levels of the PDGFRα itself (Figure 4B). In contrast, no reduction in levels of TrkC (the receptor for NT-3 [87]) was caused by exposure to MeHg (Figure 4C).
One possible explanation for the ability of MeHg to cause a reduction in PDGF-mediated signaling and in total levels of PDGFRα, without affecting NT-3–mediated signaling or TrkC levels, would be that exposure to this toxicant leads to activation of c-Cbl, an E3 ubiquitin ligase that ubiquitylates the activated PDGFRα [88,89], thus leading to its internalization and potential lysosomal degradation [90–92]. Such a possibility is particularly intriguing in light of multiple reports that c-Cbl can be activated by Fyn kinase (e.g., [93–96]), a Src family kinase that can be activated by oxidative stress [97–100]. O-2A/OPCs are known to express Fyn, which has been studied in these cells for its effects on regulation of RhoA activity and control of cytoskeletal organization [101,102]. Because TrkC does not appear to be regulated by c-Cbl, redox-modulated activation of Fyn, leading to c-Cbl activation and enhanced PDGFRα degradation, would provide a potential mechanistic explanation integrating the observations reported thus far.
A variety of data support the hypothesis that MeHg exposure activates Fyn, leading to activation of c-Cbl, followed by degradation-mediated reductions in levels of activated PDGFRα. Exposure of O-2A/OPCs to 30 nM MeHg stimulated Fyn activation and c-Cbl phosphorylation (Figure 5A and 5B). Activation of Fyn and c-Cbl was blocked by the Src family kinase inhibitors PP1 (Figure 5A and 5B) and PP2 (unpublished data). We next found that exposure to MeHg enhanced ubiquitylation of PDGFRα (a predicted consequence of c-Cbl activation), an increase readily observed even in the presence of markedly reduced levels of the receptor itself (Figure 5C). Co-exposure to ammonium chloride (NH4Cl, a lysosomotropic weak base that increases lysosomal pH and disrupts lysosomal protein degradation [103–105]) prevented receptor degradation, and was associated with increased levels of ubiquitylated receptor in treated O-2A/OPCs. The increase in levels of ubiquitylated receptor was as predicted by the lack of effect of NH4Cl on either Fyn activation or c-Cbl phosphorylation (Figure 5A and 5B). Treatment with PP1, which inhibits Fyn activity (Figure 5A), was also associated with a marked reduction in the amount of ubiquitylated PDGFRα, particularly in comparison with levels of total receptor (compare upper and lower lanes in Figure 5C). As further confirmation that reductions in levels of PDGFRα were due to protein degradation, exposure to MeHg did not have any significant effects on levels of PDGFRα mRNA, as determined by quantitative PCR analysis (Figure S1A). In the presence of cycloheximide, an inhibitor of protein synthesis, MeHg further accelerated receptor loss as compared with that occurring solely due to failure to synthesize new protein (Figure S1B). Collectively, these results indicate that MeHg enhances active degradation of PDGFRα, as contrasted with reducing receptor levels as an indirect consequence of altering transcriptional or translational regulation of receptor levels.
Molecular confirmation of the role of Fyn and c-Cbl in the effects of MeHg on levels of PDGFRα was obtained by expression of dominant negative c-Cbl, or small inhibitory RNA (RNAi) for Fyn or Cbl, in MeHg-exposed O-2A/OPCs. Expression of the dominant-negative (DN) 70z mutant of c-Cbl [106–108] in O-2A/OPCs prevented MeHg-induced reductions in levels of PDGFRα (Figure 6A). Reduction in levels of Fyn protein by introduction of Fyn-specific small interfering RNA (siRNA) constructs (Figure 6B) also protected against MeHg-induced reductions in levels of PDGFRα (Figure 6C), as predicted by the hypothesis that MeHg-induced activation of Fyn mechanistically precedes reductions in receptor levels. Similar results were obtained using RNAi constructs for c-Cbl, but are presented later in the paper, in the context of analysis of other toxicants.
Suppression of Fyn or c-Cbl activity, or overexpression of PDGFRα itself, also protected against the functional effects of MeHg exposure (Figure 7). Pharmacological inhibition of Fyn activity with PP1 enabled analysis of O-2A/OPC division at the clonal level, and demonstrated that PP1 blocked MeHg-induced suppression of cell division (Figure 7A). O-2A/OPCs expressing DN-70Z-c-Cbl and exposed to MeHg were also protected from effects of MeHg on cell division, as analyzed by BrdU incorporation (Figure 7B). Co-treatment of MeHg-exposed O-2A/OPCs with PP1 or NH4Cl also blocked MeHg-associated suppression of ERK1/2 phosphorylation (and MeHg-induced reductions in levels of PDGFRα, indicating that ERK1/2 suppression was a secondary consequence of the effects of Fyn and c-Cbl activation (Figure 7C). Overexpression of PDGFRα in MeHg-exposed O-2A/OPCs also protected cells from MeHg-associated reductions in ERK1/2 phosphorylation (Figure 7D).
To determine whether effects of MeHg revealed a general mechanism by which chemically diverse toxicants with pro-oxidant activity could alter cellular function in similar ways, we next examined the effects of exposure of dividing O-2A/OPCs to Pb (a heavy metal toxicant) and paraquat (an organic herbicide). As discussed in the Introduction, these toxicants both make cells more oxidized, but through mechanisms that differ between them and also from effects of MeHg.
Despite their chemical differences from MeHg, and from each other, Pb and paraquat had apparently identical effects as MeHg on ERK1/2 phosphorylation, activation of Fyn and c-Cbl, and reductions in levels of phosphorylated PDGFRα and on total levels of PDGFRα (Figure 8). O-2A/OPCs were exposed to 1 μM Pb (equivalent to the level of 20 μg/dl that is known to be associated with cognitive impairment, and a level of Pb previously found to inhibit O2A/OPC division without causing cell death [38,53,109]) or to 5 μM paraquat (an exposure level selected as being in the lowest 0.1% of the range of paraquat concentrations studied by others in vitro, which range from 8 μM–300 mM (e.g., [110–114]). Pb and paraquat exposure at these levels did not cause cell death, but did make O-2A/OPCs approximately 20% more oxidized, as determined by analysis of cells with the redox-indicator dyes dihydro-chloromethyl-rosamine or dihydro-calcein-AM (unpublished data). Both Pb and paraquat exposure were associated with activation of Fyn (Figure 8A), increased phosphorylation of c-Cbl (Figure 8B), reduced levels of ERK1/2 phosphorylation, and reduced levels of phosphorylated and total PDGFRα (Figure 8C). As for MeHg, the effects of Pb and paraquat on PDGFRα levels were prevented by expression of RNAi for c-Cbl (Figure 8D), DN(70Z) c-Cbl, or RNAi for Fyn (unpublished data).
It has previously been suggested that the effects of Pb on O-2A/OPCs are mediated through activation of PKC [38], a pathway that has not been implicated in the activity of MeHg or paraquat. To determine whether PKC inhibition could distinguish between effects of Pb versus MeHg or paraquat, and to determine if PKC activation was relevant to the effects of toxicants on Fyn or c-Cbl activation or reductions in PDGFRα levels, we next examined the effects of co-exposure of O-2A/OPCs to bisindolylmaleimide I (BIM-1, a broad-spectrum PKC inhibitor previously used in the analysis of the role of PKC activation in the effects of Pb on O-2A/OPCs [38]). As shown in Figure S2, we found that co-exposure of O-2A/OPCs to BIM-1 with Pb, MeHg, or paraquat did not prevent toxicant-mediated activation of Fyn (Figure S2A) or c-Cbl (Figure S2B). BIM-1 co-exposure also did not protect against MeHg-, Pb- or paraquat-induced reductions in levels of PDGFRα (Figure S2C).
If it is correct that Fyn activation, with its consequences, is regulated by the ability of toxicants to make cells more oxidized, then antagonizing such redox changes should prevent Fyn activation. Previous studies have shown that an effective means of preventing the increase in oxidative status and the suppression of cell division caused by exposure of O-2A/OPCs to TH is to treat cells with N-acetyl-L-cysteine (NAC), a cysteine pro-drug that is readily taken up by cells and converted to cysteine [59]. Cysteine is the rate-limiting precursor for synthesis of glutathione, one of the major regulators of intracellular redox status (e.g., [115,116]. NAC also possesses anti-oxidant activity, has long been used as a protector against many types of oxidative stress (e.g., [9,117,118]), and has been shown to confer protection against a wide range of toxicants, including MeHg (e.g., [119–121]), Pb (e.g., [9,122,123]), and paraquat (e.g., [17,124]), as well as such other substances as aluminum [125], cadmium [126], arsenic [127], and cocaine [128].
As predicted by the hypothesis that the pro-oxidant activities of chemically diverse toxicants are causal in Fyn activation, NAC was equally effective at preventing Fyn activation—and its consequences—induced by exposure to MeHg, Pb, or paraquat (Figures 2–5, 7, and 8). For cells grown at the clonal level, NAC blocked the suppressive effects of MeHg on cell division (Figure 2). NAC also blocked all effects of MeHg on PDGF-mediated signaling, and rescued normal level of activity of SRE and NF-kB promoter-reporter constructs and levels of phosphorylation of ERK1/2, Akt, and PDGFRα (Figure 3). Consistent with the hypothesis that Fyn is activated when cells become more oxidized [97–100], NAC also blocked MeHg-induced activation of Fyn and phosphorylation of c-Cbl (Figure 5), and prevented MeHg-induced reductions in levels of PDGFRα (Figure 4). Critically, for the hypothesis that Pb and paraquat effects also were mediated by changes in redox state, NAC also blocked the effects of Pb and paraquat on Fyn activation and c-Cbl phosphorylation, and protected against effects of these toxicants on ERK1/2 phosphorylation and levels of PDGFR (Figure 8). Levels of PDGFRα were also protected by exposure of O-2A/OPCs to procysteine (Figure 8), a thiazolidine-derivative cysteine pro-drug that differs from NAC in having no intrinsic anti-oxidant activity [129]. Although it is conceivable that the ability of cysteine pro-drugs to protect against the effects of MeHg, Pb, and paraquat is due to enhanced toxicant clearance associated with elevated levels of glutathione, analysis of Pb uptake with Leadmium Green AM (a fluorescent indicator of Pb levels) showed no significant difference in Pb levels between cells exposed to Pb as compared with cells exposed to Pb and NAC (Figure S3).
The ability of NAC to block toxicant-induced activation of Fyn raises the question of whether this is due to a true prevention of the effects of toxicant exposure on activation of this kinase or, alternatively, is due to an ability of NAC to independently suppress Fyn activity to such an extent that the apparent block of toxicant effects instead represents the summation of two opposing influences of equivalent magnitude. To evaluate these two possibilities, O-2A/OPCs were exposed to 1 mM NAC in the absence of toxicants, and Fyn and c-Cbl activation were evaluated as in Figure 5. We found that NAC exposure had only a slight, and nonsignificant, effect on the levels of basal Fyn activity in O-2A/OPCs (Figure 9A). In agreement with this outcome, NAC exposure did not have any marked effect on levels of c-Cbl phosphorylation (Figure 9B). Thus, it appears that NAC-mediated counteraction of the effects of toxicants on Fyn activation is far greater in its magnitude than its direct effects on basal levels of Fyn activity.
If the hypothesis is correct that exposure of O-2A/OPCs to toxicants causes activation of the Fyn/c-Cbl pathway, then other c-Cbl targets should be affected similarly to the PDGFRα. One member of the c-Cbl interactome [92] known to be expressed by O-2A/OPCs is c-Met [130], the receptor for hepatocyte growth factor (HGF; [131,132]). Oligodendrocytes also have recently been reported to be responsive to epidermal growth factor (EGF) application with morphological changes [133], and microarray analysis confirms that the EGF receptor (EGFR) is expressed by O-2A/OPCs (C. Pröschel and M. Noble, unpublished results). The EGFR is perhaps the most extensively studied RTK target of c-Cbl [90,96,107,134–137], but c-Met regulation by c-Cbl appears to follow similar principles [106,138].
As shown in Figure 10, exposure of O-2A/OPCs to MeHg was associated with reductions in levels of c-Met (Figure 10A) and EGFR (Figure 10B). As predicted by the hypothesis that Pb and paraquat converge with MeHg on activation of the Fyn/c-Cbl pathway, levels of C-Met and EGFR were also reduced in O-2A/OPCs exposed to these additional toxicants. Consistent with the hypothesis that such changes were associated with the ability of toxicants to make cells more oxidized, NAC protected both c-Met and EGFR levels from reductions associated with exposure to MeHg, Pb, or paraquat.
Further support for the Fyn/c-Cbl hypothesis of toxicant convergence was provided by observations that neither Pb or paraquat caused a reduction in levels of TrkC (Figure 10C), just as observed for MeHg (Figure 4C).
Although the central goal of the present studies was the identification of mechanistic pathways on which chemically diverse toxicants converge, it is important to also consider whether any aspects of our in vitro findings are predictive of in vivo outcomes. Although detailed in vivo investigations will be the subject of future studies, we have tested three of the key findings of our present work for which previous studies are not predictive of likely experimental outcomes.
The three questions we examined in vivo were whether toxicant exposure is associated with specific reductions in RTKs that are c-Cbl targets, whether this occurs at levels of toxicant exposure approximating the effects of environmental exposure, and whether such exposure can be shown to cause subtle changes in O-2A/OPC function. These experiments were conducted entirely with MeHg for several reasons. First, there is already extensive evidence that Pb exposure in vivo has adverse effects on myelination and on O-2A/OPCs (e.g., [38,43,53,139–142]). In contrast, evidence that MeHg exposure may have any effects on myelination thus far comes only from observations of increased latencies in ABRs [75–79], with no studies examining effects of this toxicant on the function of cells important for myelination (i.e., oligodendrocytes or their ancestral O-2A/OPCs). Third, previous studies on mice have not been conducted using levels of exposure of broad environmental relevance. Instead, such studies have defined a low exposure range as being exposure of animals to MeHg in their drinking water at a concentration of one or more parts per million (e.g., [143–147]), an exposure level considerably higher than what our studies would predict as being necessary to affect progenitor cells of the developing CNS. Thus, the question of whether MeHg exposure levels of broader environmental relevance would have any effects at all in vivo appears to be largely unaddressed.
To test the hypothesis that environmentally relevant levels of MeHg exposure can perturb the developing CNS in subtle ways, we exposed SJL mice to 100 or 250 ppb MeHg in their drinking water throughout gestation, and maintained this exposure until sacrifice of pups at 7 and 21 d after birth. As discussed in Materials and Methods, these exposure levels enabled us to approximate the predicted mercury levels in the CNS of 300,000–600,000 infants in the US. The exposure levels examined in our studies are 75%–90% below what has otherwise been considered to be low-dose exposure in mice.
We found that developmental exposure of mice to MeHg at either 100 ppb or 250 ppb in the maternal drinking water was associated with clear and significant reductions in levels of PDGFRα and EGFR, but not of TrkC (Figure 11). Treatment of SJL mice with 100 or 250 ppb MeHg in the drinking water during gestation and suckling was associated with reductions in levels of PDGFRα and EGFR in the cerebellum, hippocampus, and corpus callosum when brain tissue was sampled at 7 and 21 d after birth. In contrast, levels of the NT-3 receptor TrkC were not reduced in these animals, as predicted by our in vitro analyses. It was particularly striking that exposure even to 100 ppb MeHg in the drinking water was enough to have significant effects on levels of PDGFRα and EGFR. These changes, and the lack of effect of MeHg exposure on TrkC levels, are as predicted from our in vitro analyses.
Analysis of BrdU incorporation revealed that these low levels of MeHg exposure also were associated with statistically significant reductions in the division of O-2A/OPCs in vivo. In these experiments, postnatal day 14 (P14) animals were treated as for analysis of receptor levels except that BrdU was administered 2 h before sacrifice. Sections then were analyzed with anti-BrdU antibodies to identify cells engaged in DNA synthesis and with antibodies to olig2 to identify O-2A/OPCs (as in [148]). Olig2 is a transcriptional regulator expressed in oligodendrocytes and their ancestral precursor cells (e.g., [50,149–152]. In white matter tracts of the CNS, BrdU+ cells that express Olig2 are considered to be O-2A/OPCs [153,154]). In our studies, greater than 90% of all BrdU+ cells in the corpus callosum were also Olig2+. When we analyzed the number of Olig2+/BrdU+ cells found in the corpus callosum of control and experimental animals (see Materials and Methods for details of analysis), we found a 20% reduction in the number both of total BrdU+ cells and of Olig2+/BrdU+ cells (Figure 11), an outcome in agreement with the results of our in vitro studies (Figure 2B).
Our studies demonstrate that chemically diverse toxicants converge on activation of a previously unrecognized pathway of cellular regulation that leads from increases in oxidative status to reductions in levels of specific RTKs. Analysis of effects of MeHg on O-2A/OPCs dividing in response to PDGF first demonstrated suppression of PDGF-induced signaling, but no reduction in NT-3–induced phosphorylation of ERK1/2. Further analysis demonstrated that MeHg exposure enhanced degradation of PDGFRα as a consequence of the sequential activation of Fyn and c-Cbl. As predicted by the hypothesis that MeHg exposure activates the redox/Fyn/c-Cbl pathway, exposure to this toxicant was also associated with reductions in levels of EGFR and c-Met (which are c-Cbl targets), but not in levels of TrkC (which is not a c-Cbl target). The redox/Fyn/c-Cbl pathway was also activated by Pb and paraquat, leading to negative modulation of RTK-mediated signaling by regulating receptor degradation and causing reductions in levels of PDGFRα, EGFR, and c-Met, but not of TrkC. Developmental exposure to MeHg was also associated with reduced levels of PDGFRα and EGFR, but not of TrkC, consistent with the hypothesis that this same regulatory pathway is activated in association with in vivo toxicant exposure.
The results of our studies are novel in a number of ways, beginning with the identification of a previously unrecognized regulatory pathway activated by chemically diverse toxicants. Although the importance of identifying general principles that apply to chemically diverse toxicants is a widely recognized goal of toxicology research, relatively few such principles have been identified. For example, although toxicants may be classified as hormonal mimetics, mutagens, carcinogens, neurotoxins, etc., relatively few mechanistic pathways have been identified on which chemically diverse substances converge.
Our present studies have identified Fyn activation as a common cellular target for the action of chemically diverse toxicants with pro-oxidant activity. Whether oxidative changes are by themselves sufficient to induce sequential activation of Fyn and c-Cbl will be a subject of continued analysis, but existing data make it difficult to imagine a compelling alternative hypothesis to explain our results. Fyn is well established as being activated when cells become more oxidized [97–100], and there is no evidence for any other unifying feature of MeHg, Pb, and paraquat that would cause Fyn activation. Activation of Fyn, and the effects of activation of the Fyn/c-Cbl pathway, were blocked by NAC (which antagonizes oxidative changes in O-2A/OPCs [59]) as effectively as by expression of Fyn-specific RNAi constructs or by pharmacological inhibition of Fyn activity. NAC protects against physiological stress in two ways, both as an anti-oxidant itself and by providing increased levels of cysteine, the rate-limiting precursor in glutathione biosynthesis (e.g., [115,116]). The ability of ProCys (which has no intrinsic anti-oxidant properties [129]) to confer similar protection as NAC suggests that it is through their enhancement of glutathione production that these two cysteine pro-drugs exert their protective effects. The relatively small effect of NAC exposure by itself on basal Fyn activity in the experimental conditions used indicates that, at least in these experiments, NAC's protective effect was more likely to be due to protection against increases in oxidative status than due to a direct suppression of Fyn activity to an extent that would neutralize the activating effects of toxicant exposure. Although increased glutathione levels theoretically could also protect against the effects of toxicants by enabling enhanced cellular export of physiological stressors (reviewed in, e.g., [155,156]), analysis with Leadmium Green AM (which can detect intracellular Pb in the nM range) revealed no apparent effect of NAC treatment on cellular levels of Pb (Figure S3). Further support for the hypothesis that transport of xenobiotics is not a likely explanation for the protective effects of NAC is also provided by ongoing studies demonstrating that TH and BMP-4 (both of which cause O-2A/OPCs to become more oxidized [59]) also cause activation of Fyn and c-Cbl, with associated reductions in PDGFRα levels (Z. Li and M. Noble, unpublished data). NAC blocks the effects of TH and BMP on differentiation, and also prevents TH- and BMP-mediated activation of Fyn and c-Cbl ([59]; Z. Li and M. Noble, unpublished data). Changes in intracellular redox state, and the predicted ability to protect with NAC, are the common features linking the activation of Fyn with MeHg, Pb, paraquat, TH, and BMP.
Although Fyn has multiple targets, it seems most likely that activation of c-Cbl provides the explanation for the effects of MeHg on PDGF-mediated signaling. Suppression of c-Cbl activity by expression of DN(70Z) c-Cbl or RNAi protected against the effects of MeHg on cell division and reductions in levels of PDGFRα. Moreover, the induction of PDGFRα ubiquitylation by MeHg, the lack of effects of MeHg on PDGFRα mRNA levels, the rescue of receptor levels by disrupting lysosomal function, and other observations all strongly indicate the importance of c-Cbl regulation in understanding the effects of toxicant exposure. The importance of Fyn in activation of c-Cbl is supported by the ability of expression of Fyn-specific RNAi, or pharmacological inhibition of Fyn activity, to protect against the effects of toxicant exposure. Because Fyn activation in O-2A/OPCs also leads to activation of Rho-GTPase, leading to inhibition of Rho kinase activity [102,157], we also examined the effects of treatment of cells with the Rho kinase inhibitor Y23762 (Figure S4). Although this agent inhibited Rho kinase activity in O-2A/OPCs, it neither protected against nor exacerbated the effects of MeHg on progenitor cell division (as determined by BrdU incorporation). Thus, although it will be of interest to examine the effects of toxicant exposure on other Fyn targets, it currently seems that Fyn-mediated activation of c-Cbl is central to understanding the effects of toxicants on O-2A/OPCs.
The discovery of sequential activation of Fyn and c-Cbl by pro-oxidants provides a new means of integrating the effects of changes in intracellular redox state with the control of the cell cycle. Although the ability of Fyn to be activated by increases in oxidative status [97–100], the functional interaction of Fyn with c-Cbl (e.g., [93–96]), and the regulation of degradation of specific RTKs by c-Cbl (e.g., see [88–90,96,106,107,134–138]) have all been subjects of study by multiple laboratories, our studies appear to provide the first integration of all of these components into a regulatory pathway of obvious relevance to the regulation of cell function by redox status. This regulatory pathway, summarized in Figure 12, offers a number of clear predictions, some of which have been tested in our present studies. Several studies on different cell types have confirmed our own finding [59] that making dividing cells more oxidized can suppress division and induce differentiation [158–160], and it will be of interest to determine the contribution of the redox/Fyn/c-Cbl pathway in these other cell systems, as well as in modulating other changes in cellular function that have been attributed to increased oxidative status (e.g., [73,161–166]).
It is particularly striking that the changes we observed were seen at environmentally relevant exposure levels for both MeHg and Pb. As many as 600,000 newborn infants in the US each year have cord blood mercury levels greater than 5.8 ppb [46] (i.e., ∼30 nM). It is reported that the blood:brain ratio for humans may be as high as 1:5 to 1:6.7 [167,168], therefore, in vivo levels in brain may be still higher than those we have studied. It is also noteworthy that levels of MeHg exposure at which selective reduction in PDGFRα expression was readily observed in vivo were 90% or more lower than exposure levels generally considered to constitute low-to-moderate exposure (e.g., [143–147]). Blood Pb levels may be of concern at levels as low as 10 μg/dl (e.g., [169–176]), which is equivalent to 0.48 μM, but which may be increased to micromolar in the brain by mechanisms relevant to Ca2+ transport [171]). Even given equivalence in blood:brain Pb levels, a concentration of 1 μM is equivalent to the approximately 20 μg/dl blood Pb levels known to be associated with cognitive impairment (e.g., [169–177], an exposure level of particular concern in countries where leaded gasoline is still used and in which mean blood lead levels in schoolchildren may be as high as 15 μg/dl [178].
The study of environmentally relevant levels of toxicant exposure is a great challenge, both in vitro and in vivo, and it may be that analysis of stem and progenitor cell populations will be critical in furthering such analysis. In vitro, O-2A/OPCs appear to offer a particularly useful target cell for such studies, in part due to their sensitivity to environmentally relevant exposure levels of toxicants, but also due to the ability to use clonal analysis in quantitative studies on the cumulative effects of small changes in the balance between division and differentiation [179–181]. Such studies have shown that even such potent physiological regulators as TH may only increase the probability of oligodendrocyte differentiation at each progenitor cell cycle from approximately 0.5 to 0.65 [179]. Thus, although their cumulative effects over time may be readily observable, analysis of subtle effects in acute assays may fail to identify important alterations in progenitor cell function. In addition, it will be important to extend analysis on differentiation to other precursor cell populations, as indicated by recent observations that neuronal differentiation of neuroepithelial stem cells may be compromised by MeHg exposure levels as low as 2.5 nM [182]. In vivo, the 20% reduction in number of dividing O-2A/OPCs observed in animals exposed to 100 ppb MeHg during development was of particular interest, as such relatively subtle changes might be predicted to reduce myelination in ways that require equally subtle analysis to detect functional outcomes. Analysis of conduction velocity in the auditory system may offer one such analytical tool, and the sensitivity of O-2A/OPCs to toxicant exposure may provide an explanation for the consistency with which increases in ABR latency suggestive of myelination abnormalities are associated with exposure to a variety of toxicants and physiological stressors, including MeHg [75–79], Pb [183,184]), cocaine [185,186], and carbamazepine [187].
The general importance of the signaling pathways regulated by Fyn and c-Cbl suggests that the ability of chemically diverse toxicants to converge on this pathway may be of broad relevance to the understanding of toxicant action. Such c-Cbl targets as PDGFRα, EGFR, and c-Met play critical roles in processes as diverse as cell proliferation, survival, and differentiation, cortical neurogenesis, maintenance of the subventricular zone, astrocyte development, development of cortical pyramidal dendrites, motoneuron survival and pathfinding, sympathetic neuroblast survival, and hippocampal neuron neurite outgrowth, as well as having extensive effects on development of kidney, lung, breast, and other tissues (e.g., [60,61,130,188–195]). Indeed, the range of targets of c-Cbl [92,135] offers a rich fabric of potentially critical regulatory molecules that would be affected by changes in activity of this protein, with the importance of particular proteins being dependent on the cell type and developmental stage under consideration. In addition, Fyn regulation of the Rho/ROCK signaling pathway could be of relevance in understanding toxicant-mediated alterations on such cytoskeletal functions as cell migration, neurite outgrowth, and development of dendritic morphology (e.g., [196–198]). Our studies predict that any toxicant that makes cells and/or tissues more oxidized would activate Fyn, a list that includes substances as chemically diverse as MeHg (e.g., [1–6], Pb [6–9], and organotin compounds [1,2,5,10,11]), cadmium [12,13], arsenic [12,14], ethanol [15,16], and various herbicides (e.g., paraquat [17,18], pyrethroids [19–21], and organophosphate and carbamate inhibitors of cholinesterase [22–26]).
In summary, our studies provide a new general principle and evidence of a new regulatory pathway that may be relevant to the understanding of the action of a large number of chemically diverse toxicants and other modulators of oxidative status. Because the outcomes we have identified occur at quite low toxicant exposure levels, they may provide a particularly useful unifying principle for the analysis of toxicant effects. Our present studies, combined with our previous analysis of the central importance of intracellular redox state in modulating progenitor cell function [59], lead to the prediction that any toxicant with pro-oxidant activity will exhibit these effects. Although toxicants of differing chemical structures will also have additional activities, the convergence of small increases in oxidative status on regulation of the redox/Fyn/c-Cbl pathway provides a specific means by which exposure to low levels of a wide range of chemically diverse toxicants might have similar classes of effects on development. Our findings also provide a strategy for rapid identification of such effects by any of the estimated 80,000 to 150,000 chemicals for which toxicological information is limited or nonexistent, thus enabling a preliminary identification of compounds that would need to be examined in vivo. The sensitivity of O-2A/OPCs to environmentally relevant levels of MeHg and Pb provides a great advantage over established cell lines and other such neural cells as astrocytes, for which these low exposure levels may have little effect, and the importance of understanding the effects of toxicants on progenitor cell function provides a direct link between our studies and the broad field of developmental toxicology. In addition, the ability of NAC to protect progenitor cells against the adverse effects of chemically diverse toxicants raises the possibility that this benign therapeutic agent may be of benefit in protecting children known to be at increased risk from the effects of toxicant exposure during critical developmental periods. Finally, the principles indicated by our findings appear likely to have broad applicability in understanding the regulation of cell function by alterations in redox balance, regardless of how they might be generated.
O-2A/OPCs were purified from corpus callosum of P7 CD rats as described previously to remove type 1 astrocytes, leptomeningeal cells, and oligodendrocytes [59,64,199]. Cells were then grown in DMEM/F12 supplemented with 1-μg/ml bovine pancreas insulin (Sigma, St. Louis, Missouri, United States), 100-μg/ml human transferrin (Sigma), 2 mM glutamine, 25-μg/ml gentamicin, 0.0286% (v/v) BSA pathocyte (ICN Biochemicals, Costa Mesa, California, United States), 0.2 μM progesterone (Sigma), 0.10 μM putrescine (Sigma), bFGF-2 (10 ng/ml; PEPRO Technologies, London, United Kingdom), and PDGF-AA (10 ng/ml; PEPRO) onto poly-l-lysine (Sigma) coated flasks or dishes. Under these conditions, O-2A/OPCs derived from the corpus callosum of P7 rats are predominantly in cell division and do not generate large numbers of oligodendrocytes during the time periods utilized in this analysis.
To generate sufficient numbers of cells for biochemical analysis, cells were expanded through one to two passages in PDGF + FGF-2 before replating in the presence of PDGF alone. When cells achieved approximately 50% confluence, MeHg, Pb, or paraquat was added to their medium at concentrations indicated in the text. Doses for the toxicants were chosen on the basis of dose-response curves to identify sublethal exposure levels (unpublished data), as a reflection of blood and brain toxicant levels of these compounds and, where applicable, on the basis of previous reports. All toxicant concentrations examined were confirmed to cause death of less than 5% of cells over the time course of the experiment.
For analysis of the effects of potential inhibitors of toxicant action, cells were exposed to the blocking compound of interest 1 h before addition of toxicant. The concentrations of inhibitors used are listed as following: 0.5 μM BIM-1 (PKC inhibitor), 0.5 μM PP1/PP2 (Src family kinase inhibitors), and 10 mM NH4Cl (lysosome inhibitor); and the concentrations of toxicants used are listed as following, except when mentioned specifically: MeHg (20 nM), Pb (1 μM), and paraquat (5 μM).
To examine the degradation of PDGFRα, O-2A/OPCs were treated with MeHg (20 nM) for different durations with or without cycloheximide (CHX; 1 μg/ml) added 1 h before MeHg. The cells were then collected and lysed for Western blotting. For example, in the multi-toxicant analysis of Figures 8–10, for analysis of PDGFRα, O-2A/OPCs were exposed for 24 h to MeHg (20 nM), Pb (1 μM), and paraquat (5 μM) for 24 h in the presence of 0.5 μM bisindolylmaleimide 1 (BIM-1), 0.5 μM of PP1, 1 mM NAC, or 1 mM procysteine, which had been added 1 h prior to toxicant addition. Cells were lysed for Western blot analysis using anti-PDGFRα(pY742) antibody. The membranes were de-probed and then re-probed with antibody against total PDGFRα and anti–β-tubulin antibody. For analysis of Fyn activity and c-Cbl phosphorylation, progenitors were exposed to MeHg (20 nM), Pb (1 μM), and paraquat (5 μM) for 3–4 h in the presence of 0.5 μM BIM1, 0.5 μM PP1,or 1 mM NAC (each of which was added 1 h before addition of toxicant).
Cells were deprived of PDGF-AA for 5 h before re-exposure to PDGF-AA (10 ng/ml) for 1 h for Western blot or 6 h for luciferase assays of pathway activation. Transient transfection was performed using Fugene6 (Roche, Basel, Switzerland) transfection solution according to the manufacturer's protocol. For the luciferase assay, cells seeded in 12-well plates were transfected with a reporter plasmid SRE-Luc(firefly) or NFκB-Luc(firefly) (BD-Clontech, Palo Alto, California, United States) and an internal control plasmid pRLSV40-LUC. Analyses of luciferase activity were performed according to the protocol of the Dual Luciferase Assay System (Promega, Madison, Wisconsin, United States), which uses an internal control of Renilla luciferase for quantification, and relative light units were measured using a luminometer.
Anti-phosphorylated ERK monoclonal, anti-ERK monoclonal, anti-TrkC polyclonal, anti-Fyn polyclonal, anti-EGFR polyclonal, anti–c-Met polyclonal, anti–phospho-tyrosine monoclonal, and anti-PDGFRα polyclonal antibodies were obtained from Santa Cruz Biotechnology (Santa Cruz, California, United States). Anti–c-Cbl monoclonal antibody was obtained from BD PharMingen (San Diego, California, United States). Anti-phosphorylated Akt monoclonal and anti-Akt polyclonal antibodies were obtained from Cell Signaling Technology (Beverly, Massachusetts, United States). Anti-phosphorylated PDGFRα polyclonal antibody was obtained from Biosource (Carlsbad, California, United States). The cell culture samples were collected and lysed in RAPI buffer, whereas dissected tissue samples were sonicated in RAPI buffer. Samples were resolved on SDS-PAGE gels and transferred to PVDF membranes (PerkinElmer Life Science, Wellesley, Massachusetts, United States). After being blocked in 5% skim milk in PBS containing 0.1% Tween 20, membranes were incubated with a primary antibody, followed by incubation with an HRP-conjugated secondary antibody (Santa Cruz Biotechnology). Membranes were visualized using Western Blotting Luminol Reagent (Santa Cruz Biotechnology). All analyses of signaling pathway components were conducted in the presence of ligand for the receptor pathway under analysis (either PDGF-AA for PDGFRα, NT-3 or TrkC, HGF for c-Met, or EGF for EGFR).
Cell proliferation was assessed by bromodeoxyuridine (BrdU) incorporation and by using the mouse anti-BrdU mAb IgG1 (1:100; Sigma) to label dividing cells. Stained cells on coverslips were rinsed two times in 1× PBS, counterstained with 4′6-diamidino-2-phenylindole (DAPI; Molecular Probes, Eugene, Oregon, United States) and mounted on glass slides with Fluoromount (Molecular Probes). Staining against surface proteins was performed on cultures of living cells or on cells fixed with 2% paraformaldehyde. Staining with intracellular antibodies was performed by permeabilizing cells with ice-cold methanol for 4 min or by using 0.5% Triton for 15 min on 2% paraformaldehyde–fixed cells. Antibody binding was detected with appropriate fluorescent dye–conjugated secondary antibodies at 10 μg/ml (Southern Biotech, Birmingham, Alabama, United States) or Alexa Fluor–coupled antibodies at a concentration of 1 μg/ml (Molecular Probes), applied for 20 min. Anti-BrdU monoclonal antibody was obtained from Sigma.
Cells were plated in 96-well microplates and grown to about 60% confluence. Prior to treatment, cells were washed twice with Hank's buffered saline solution (HBSS), loaded with 20 μM H2DCFDA (in HBSS 100 μl/well), and incubated at 37 °C for 30 min. Cells were then washed once with HBSS and growth medium to remove free probe. Then, fresh growth medium was added and a baseline fluorescence reading was taken prior to treatment. For NAC pre-treatment, NAC was added into media 1 h before further addition of MeHg, and both compounds remained in the medium during the incubation period with H2DCFDA. Fluorescence was measured in a Wallac 1420 Victor2 multilabel counter (PerkinElmer) using excitation and emission wavelengths of 485 nm and 535 nm, respectively, at different time courses as indicated in the figures. Results are presented as the value change from baseline by the formula (Ftexp − Ftbase)/Ftbase normalized with the control group, where Ftexp = fluorescence at any given time during the experiment in a give well and Ftbase = baseline fluorescence of the same well.
We further determined whether pre-treatment with NAC altered levels of intracellular Pb by analysis with the Leadmium Green AM dye (Molecular Probes), according to the manufacturer's instructions. In five separate experiments, we found no significant difference between O-2A/OPCs treated with 1 μM Pb versus [Pb + NAC] (unpaired t-test), and the values for both Pb-treated samples were several-fold higher than control values. All of these data strongly support the hypothesis that the major effect of NAC is to antagonize cellular oxidation.
For the co-immunoprecipitation assay, anti-c-Cbl monoclonal antibody (BD PharMingen) or anti-PDGFRα polyclonal antibody (Santa Cruz Biotechnology) was added to the pre-cleared cell lysates (250 μg of total protein), and the mixtures were gently rocked for 2 h at 4 °C. A total of 30 μl of protein A/G agarose was then added to the mixture followed by rotating at 4 °C overnight. The protein A/G agarose was then spun down and washed thoroughly three times. The precipitates were resolved on an 8% SDS-PAGE gel and then were subjected to Western blot analysis using an anti-p-Tyr (for c-Cbl phosphorylation assay) or ubiquitin (for PDGFR ubiquitination assay) antibody (Santa Cruz Biotechnology).
Fyn kinase activity was quantified using the Universal Tyrosine Kinase Assay Kit (Takara, Madison, Wisconsin, United States). O-2A/OPCs exposed to different treatments were solubilized with an equal volume of the extraction buffer provided with the kit for 15 min, and the resulting lysates were centrifuged at 13,000 × g for 15 min at 4 °C; 250 μg of total cell lysates were immunoprecipitated with anti-Fyn antibody (Santa Cruz Biotechnology). Following immunoprecipitation, Fyn immune complexes were washed four times with extraction buffer, and then Fyn kinase activities of each sample were assayed using the kit according to the manufacturer's instructions.
Rho kinase activity was quantified using the CycLex Rho-Kinase Assay kit (MBL International, Woburn, Massachusetts, United States) as described. Cells were lysed and about 500 μg of total cell lysates were immunoprecipitated with anti-ROCK1 antibody (Sigma), and the precipitates were re-suspended with kinase reaction buffer provided in the kit. Rho kinase activities of each sample were assayed using the kit according to the manufacturer's instructions.
siRNA target sites were selected by scanning the cDNA sequence for AA dinucleotides via siRNA target finder (Ambion, Austin, Texas, United States). Those 19-nucleotide segments that start with G immediately downstream of AA were recorded and then analyzed by BLAST search to eliminate any sequences with significant similarity to other genes. The siRNA inserts, containing selected 19-nucleotide coding sequences followed by a 9-nucleotide spacer and an inverted repeat of the coding sequences plus 6 Ts, were made to double-stranded DNAs with ApaI and EcoRI sites by primer extension, and then subcloned into plasmid pMSCV/U6 at the ApaI/EcoRI site. The corresponding oligonucleotides for the Fyn and c-Cbl RNAi's are listed in Table 1. Several nonfunctional siRNAs, which contain the scrambled nucleotide substitutions at the 19-nucleotide targeting sequence of the corresponding RNAi sequence, were constructed as negative controls. All of these plasmids were confirmed by complete sequencing.
pJEN/neo-HA-70z-c-Cbl plasmids were generously provided by Dr. Wallace Langdon. The pBabe(puro)-HA-70z-c-Cbl plasmids were constructed by transferring the BamH1-digested HA-70z-c-Cbl from pJEN/neo-HA-70z-c-Cbl into the BamH1 digested pBabe(puro) vector. The pBabe(puro)-HA-70z-c-Cbl, pMSCV/U6-Fyn-RNAi, pMCV/U6-c-Cbl-RNAi, and the corresponding scrambled RNAi plasmids and the empty plasmids were transfected into Pheonix Ampho cells by Fugene6 (Roche) transfection solution according to the manufacturer's protocol. Twenty-four hours after transfection, medium was changed to DMEM/F12(SATO, but with no TH) supplemented with 10-ng/ml PDGF-AA and bFGF. Virus supernatant was collected 48 h post-transfection, filtered through 0.45-μm filter to remove non-adherent cells and cellular debris, frozen in small aliquots on dry ice, and stored at −80 °C. Twenty-four hours prior to infection, O-2AOPCs were seeded. The following day, the culture medium was aspirated and replaced with virus supernatant diluted 1:1 in the O-2A growth media. Medium was then changed into O-2A/OPC growth medium after 8 h or overnight. Twenty-four hours after infection, the cells were collected by trypsinization and reseeded in the selective medium (growth medium + 200-ng/ml puromycin). By the next day, all noninfected cells were floating and presumably dead or dying. The infected cells were allowed to proliferate for 2 d, and then collected and re-seeded for the following experiments.
Total RNA was isolated using TRIZOL reagent (Invitrogen, Carlsbad, California, United States) according to the manufacturer's protocol. A total of 1 μg of RNA was subjected to reverse transcription using Superscript II (Invitrogen). The reactions were incubated at 42 °C for 50 min. The FAM-labeled probe mixes for rat PDGFRα and Fyn, and the VIC-labeled GAPDH probe mix were purchased from Applied Biosystems (Foster City, California, United States). For multiplex real-time PCR, reactions each containing 5 μl of 10-fold–diluted reverse transcription product, 1 μl of interest gene probe mix, 1 μl of GAPDH probe mix, and 10 μl of TaqMan Universal PCR Master Mix were performed on an iCycler iQ multicolor real-time PCR system (Bio-Rad, Hercules, California, United States) and cycling condition was 50 °C for 2 min and 95 °C for 10 min, followed by 40 cycles of 95 °C for 15 sec and 60 °C for 1 min. Each sample was run in triplicate. Data were analyzed by iCycler iQ software (Bio-Rad).
O-2A/OPCs purified from P7 rat optic nerve were plated in poly-L-lysine–coated 25-cm2 flasks at clonal density with DMEM medium in the presence of 10-ng/ml PDGF as previously described [59,64,199]. After 24-h recovery, cells were treated with different toxicants, each for 3 d, until visual inspection and immunostaining was performed. NAC was added 1 h before exposure to other toxicants for NAC pretreatment, and NAC co-exists throughout the culture period. The numbers of O-2A/OPCs and oligodendrocytes in each clone were determined by counting under fluorescent microscope. The three-dimensional graph shows the number of clones containing O-2A/OPC cells and oligodendrocytes. Experiments were performed in triplicate in at least two independent experiments.
Six-week-old female SJL mice were treated with MeHg in their drinking water at a concentration of 100 or 250 ppb for 30–60 d prior to mating, and then throughout pregnancy and gestation. This is a level of treatment that is 75%–90% below levels generally considered to be low to moderate and is below levels that have been associated with gross defects in adult or developing animals (e.g., [143–147]).
The exposure levels used in our studies were first determined as candidate exposures from the results of two different previous studies on the relationship between MeHg exposure and levels of toxicant in the brain. Studies by Weiss and colleagues [143] demonstrated that mice exposed to MeHg in their drinking water for up to 14 mo have brain mercury levels roughly equivalent to that in the water. In these studies, mice exposed to MeHg in their drinking water from conception at a concentration of one part per million (ppm) had brain levels of MeHg of 1.20 mg/kg (i.e., ppm) at 14 mo of age, whereas those exposed to MeHg at a concentration of 3 ppm had brain levels of 3.66 mg/kg at this age. It has also been shown, however, that mercury levels in the brain of pre-weanling animals exposed to MeHg via the mother's drinking water throughout gestation and suckling drop rapidly to one-fifth of the levels found at birth, presumably due to reduced MeHg transfer in milk [200]. As an estimated 300,000 to 600,000 infants in the US have blood cord mercury levels of 5.8 μg/l or more [46], and because the human brain concentrates MeHg 5- to 6.7-fold over the concentration occurring in the bloodstream, our goal was to achieve postnatal brain mercury levels of 30 ppb (i.e., ng/g) or less.
In practice, we found that exposure of female mice to MeHg in their drinking water at a concentration of 250 ppb prior to conception, and maintenance of this exposure during suckling, was associated with brain mercury levels in the offspring (examined at P14) of 50 ng/g, a fall that was precisely in agreement with predictions based on prior studies on the fall of mercury levels occurring during this period in suckling mice [200]. In offspring of dams exposed to MeHg at a concentration of 100 ppb in the drinking water, brain mercury was below the levels of detection of the Mercury Analytical Laboratory of the University of Rochester Medical Center. The exposure levels of 100 and 250 ppb are 75%–90% below what has otherwise been considered to be low-dose exposure in mice.
At the time of sacrifice, mice were anesthetized using Avertin (tribromoethanol, 250 mg/kg, 1.2% solution; Sigma) and were perfused transcardially with 4% paraformaldehyde in phosphate buffer (pH 7.4) following the removal of the blood by saline solution washing. The brains were removed and stored in 4% paraformaldehyde for 1 d, and then changed to 25% sucrose in 0.1 M phosphate buffer. Brains were cut coronally as 40-μm sections with a sliding microtome (SM/2000R; Leica, Heidelberg, Germany) and stored at −20 °C in cryoprotectant solution (glycerol, ethylene glycol, and 0.1 M phosphate buffer[ pH 7.4], 3:3:4 by volume). All animal experiments were conducted in accordance with National Institutes of Health guidelines for the humane use of animals.
To analyze DNA synthesis in vivo, mice were injected with a single dose of 5-BrdU (50 mg/kg body weight), dissolved in 0.9% NaCl, filtered (0.2 μm), and applied intraperitoneally 2 h prior to perfusion. After removal and sectioning of brains, 40-μm free-floating sections were incubated for 2 h in 50% formamide/2× SSC (0.3 M NaCl and 0.03 M sodium citrate) at 65 °C, rinsed twice for 5 min each in 2× SSC, incubated for 30 min in 2N HCl at 37 °C, and rinsed for 10 min in 0.1 M boric acid (pH 8.5) at room temperature. Several rinses in TBS were followed by incubation in TBS/0.1% Triton X-100/3% donkey serum (TBS-plus) for 30 min. Sections were then incubated with monoclonal rat anti-BrdU antibody (1:2,500; Harlan Sera-Lab, Loughborough, United Kingdom) and polyclonal rabbit anti-Olig2 (a generous gift from Dr. David H. Rowitch) in TBS-plus for 48 h at 4 °C. Sections were rinsed several times in TBS-plus and incubated for 1 h with donkey anti-rat FITC and donkey anti-rabbit TRITC (Jackson ImmunoResearch Laboratories, West Grove, Pennsylvania, United States). After several washes in TBS, sections were mounted on gelatin-coated glass slides using Fluoromount-G mounting solution (Southern Biotech).
Quantification of BrdU+ cells was accomplished with unbiased counting methods by confocal microscopy. BrdU immunoreactive nuclei were counted in one focal plane to avoid oversampling. In corpus callosum, BrdU+ cells were counted in every sixth section (40 μm) from a coronal series between interaural AP + 5.2 mm and AP + 3.0 mm in the entire extension of the rostral and medial part of the corpus callosum. Quantitative data are presented as mean percentage normalized to control animals. Error bars represent ± the standard error of the mean.
Digital images were captured using a confocal laser scanning microscope (Leica TCS SP2). Photomicrographs were processed on a Macintosh G4 and assembled with Adobe Photoshop 7.0 (Adobe Systems, Mountain View, California, United States). Unpaired, two-tailed Student t-test was used for statistical analysis.
|
10.1371/journal.pgen.1003115 | SLI-1 Cbl Inhibits the Engulfment of Apoptotic Cells in C. elegans through a Ligase-Independent Function | The engulfment of apoptotic cells is required for normal metazoan development and tissue remodeling. In Caenorhabditis elegans, two parallel and partially redundant conserved pathways act in cell-corpse engulfment. One pathway, which includes the small GTPase CED-10 Rac and the cytoskeletal regulator ABI-1, acts to rearrange the cytoskeleton of the engulfing cell. The CED-10 Rac pathway is also required for proper migration of the distal tip cells (DTCs) during the development of the C. elegans gonad. The second pathway includes the receptor tyrosine kinase CED-1 and might recruit membranes to extend the surface of the engulfing cell. Cbl, the mammalian homolog of the C. elegans E3 ubiquitin ligase and adaptor protein SLI-1, interacts with Rac and Abi2 and modulates the actin cytoskeleton, suggesting it might act in engulfment. Our genetic studies indicate that SLI-1 inhibits apoptotic cell engulfment and DTC migration independently of the CED-10 Rac and CED-1 pathways. We found that the RING finger domain of SLI-1 is not essential to rescue the effects of SLI-1 deletion on cell migration, suggesting that its role in this process is ubiquitin ligase-independent. We propose that SLI-1 opposes the engulfment of apoptotic cells via a previously unidentified pathway.
| Cell death is a normal part of organismal development. When cells die, other cells engulf them. In the roundworm C. elegans, engulfment is facilitated by one pathway that rearranges the actin cytoskeleton and another that recruits membrane. Together they cause the formation of cellular extensions that surround the dead cell. Notably, little is known about how engulfment is inhibited. The cytoskeletal regulatory pathway, which also promotes cell migration, includes CED-10 and ABI-1, homologs of the actin regulators Rac and the Abi proteins, respectively. In mammals, the c-Cbl proto-oncogene interacts with Rac and Abi2 and has been shown to regulate the actin cytoskeleton, so we tested whether the C. elegans homolog of Cbl, SLI-1, regulates engulfment and cell migration. We found that SLI-1 inhibits both processes. Our analysis further showed that SLI-1 does not function by inhibiting other known engulfment proteins. Cbl proteins have ubiquitin ligase domains through which they target proteins for destruction or sequestration. Most of the known functions of Cbl proteins require that domain, but we found that SLI-1 did not require it to block engulfment and cell migration. We propose that SLI-1 inhibits engulfment and cell migration through a previously unidentified pathway.
| The engulfment of apoptotic cells requires at least two processes to occur in the engulfing cell at the interface with the dying cell. Actin cytoskeletal elements need to be reorganized and membrane needs to be recruited. Together, these two processes result in the engulfing cell surrounding the dying cell. Two conserved molecular pathways were originally identified in Caenorhabditis elegans that are required for apoptotic cell engulfment and regulate these two processes.
In the pathway for membrane recruitment, which we refer to as the CED-1 pathway, four proteins have been identified, CED-7, CED-1, CED-6 and DYN-1 (Figure 1) [1]. These proteins activate DYN-1, a C. elegans dynamin homolog [2], which might recruit membrane for engulfment; in mammalian cells dynamin promotes extension of lamellipodial membrane protrusions [3].
The pathway for cytoskeletal rearrangement requires the small GTPase CED-10 Rac, the adapter protein CED-2 and the heterodimeric guanine nucleotide exchange factor CED-5/CED-12. CED-2 is thought to activate CED-5/CED-12, which, in turn, activates CED-10 Rac. Rac proteins are members of the Rho family of small GTPases that regulate the cytoskeleton and function in intracellular signaling [4]. CED-10 Rac activation causes actin cytoskeletal rearrangement and promotes engulfment [5], [6].
In addition to the two core engulfment pathways, more recent studies have identified a number of factors that regulate engulfment through these pathways. In C. elegans, MIG-2, the mammalian homolog of RhoG, another Rho family GTPase activates CED-5/CED-12 in parallel to CED-2 [7], [8]. The phosphatidylserine receptor PSR-1, the integrins INA-1 and PAT-3 and a WNT signaling pathway all appear to act upstream of CED-2 [9], [10]. In Drosophila, the Src protein Src42 and the non-receptor tyrosine kinase Shark act through the CED-1 Draper pathway [11]. Furthermore, Calcium release from the endoplasmic reticulum by a junctophilin-containing complex is also required for CED-1 Draper activity [12], [13].
Recently, we reported that the cytoskeletal regulatory protein ABI-1 is also an engulfment protein [14]. The mammalian homolog of ABI-1, Abi2, is found in a number of protein complexes, all of which regulate the actin cytoskeleton. One particular complex, the Wave Regulatory Complex (WRC) causes the formation of actin structures in response to activation by Rac [15], [16]. The WRC is composed of five proteins in C. elegans: WVE-1, GEX-2, GEX-3, ABI-1 and NUO-3. Soto et al. (2002) [17] and Patel et al. (2008) [18] presented evidence that suggested that GEX-2 and WVE-1, respectively, promote engulfment. Our genetic analysis, however, demonstrated that the CED-10 Rac pathway and ABI-1 act at least partially independently of each other. Our current model, based on all of these data is that the CED-10 Rac pathway activates the WRC but that there are other as yet unidentified molecular pathways that activate the WRC in parallel.
Far less studied are proteins that inhibit these two pathways. We showed that the tyrosine kinase and cytoskeletal regulator ABL-1 inhibits engulfment through ABI-1 in parallel to the CED-10 Rac pathway [14]. A small number of other proteins have been shown to inhibit apoptotic cell engulfment (compared to 25 proteins that promote engulfment). In mammalian cell culture, the small GTPase RhoA and its effector Rho-kinase have been shown to inhibit engulfment of apoptotic cells [19], consistent with the fact that RhoA and Rac oppose each other in many cellular processes. How Rho-kinase inhibits engulfment has not been demonstrated. In C. elegans, the Rac GTPase activating protein SRGP-1 inhibits engulfment by inactivating CED-10 [20]. The myotubularin lipid phosphatase MTM-1 and a CED-10 binding protein, SWAN-1, have also been shown to inhibit engulfment in C. elegans [21]–[23]. They are both proposed to act through the CED-10 Rac pathway. Recently, PGRN-1, a C. elegans progranulin has been shown to act in engulfment [24]. Notably, it is unclear how any of these proteins are regulated for their engulfment-inhibitory functions.
Cbl family proteins are E3 ubiquitin ligase and adaptor proteins with multiple cellular functions [25]. Cbl proteins consist of an N-terminal tyrosine kinase binding (TKB) domain followed by a conserved linker, then a RING finger domain and a C-terminal proline rich domain. The TKB domain is comprised of three subdomains: a 4-helix bundle, an EF hand and a modified SH2 domain. The crystal structure of the TKB domain has revealed that the three subdomains act together to bind to phosphotyrosines [26] and orient substrate proteins (usually tyrosine kinases) to allow the RING finger to promote their ubiquitination, targeting them for destruction or sequestration. Thus a major function of Cbl proteins is to downregulate signaling pathways in response to interactions with tyrosine phosphorylated signaling proteins [27]. Recent data show that Abi proteins are activated by epidermal growth factor (EGF) signaling and then in turn activate c-Cbl to polyubiquitinate the EGF receptor in a negative feedback regulatory loop [28]. In C. elegans, the Cbl homolog SLI-1 downregulates EGF signaling by causing ubiquitination of the LET-23 EGFR [29], [30], which decreases signaling from the downstream Ras homolog LET-60. Cbl has also been shown to interact with Rac, the CED-2-related protein Crk and Abl kinase [31]–[33]. We hypothesized that SLI-1 might act in engulfment pathways. In addition, we asked whether it did so by interacting with the C. elegans homologs of the above proteins.
We now present evidence that SLI-1 inhibits apoptotic cell engulfment. Surprisingly, we find that SLI-1 does so in parallel to the two core engulfment pathways and ABL-1 and independent of LET-60 Ras signaling. Lastly, we demonstrate that the ubiquitin ligase domain is partially dispensable for this process demonstrating that its tyrosine kinase-ubiquitinating function is unrelated to its mechanism of action in engulfment.
In animals with defects in apoptotic cell engulfment, the number of unengulfed corpses in the heads of first larval stage (L1) animals increases with the strength of the engulfment defect and defines a quantitative assay of engulfment defects [34]. L1 wild-type (N2) animals have no unengulfed corpses in their heads. Neither do animals with sli-1 mutations alone. We used two alleles of sli-1 in this study, sy143 and n3538 [35], [36]. sy143 is a C to T transition that changes Gln152 to an amber stop codon; n3538 is a C to T transition that changes Ser305 to Leu. To assess whether sli-1 modulates apoptotic cell engulfment, we tested whether sli-1 mutations suppressed or enhanced the engulfment defects of engulfment pathway genes.
The heads of animals containing a mutation in sli-1 and null mutations in ced-1 or ced-7 or a strong mutation in ced-6 (alleles e1735, n1996 and n2095, respectively) had fewer unengulfed corpses than those with each of the engulfment mutants alone (Table 1). We did not test dyn-1 mutants because they die during embryogenesis. Thus, SLI-1 appears to inhibit the engulfment of apoptotic cells. Alternative explanations for the effect of SLI-1 on these engulfment defects are presented in the next section of the paper. The fact that loss of sli-1 function suppresses the engulfment defects caused by null ced-1, ced-6 and ced-7 mutation demonstrates that SLI-1 acts in parallel to or downstream of the CED-1 pathway.
Loss of sli-1 function did not suppress the engulfment defects of null mutations in the CED-10 Rac pathway. Specifically, the engulfment defects of ced-2(n5101), ced-5(n1812) and ced-12(n3261) null mutants were not significantly modified by the presence of sli-1(sy143) or sli-1(n3538) mutations (Table 1). ced-10 null mutants die during embryogenesis but we tested the effect of sli-1 mutations on a partial loss-of-function allele, ced-10(n1993). ced-10(n1993) was suppressed by sli-1(lf) (for sy143, a decrease from 20.0 to 14.1 unengulfed corpses, p<0.0001; for n3538, a decrease to 13.8, p<0.0001). Suppression of a ced-10 partial loss-of-function defect by sli-1 mutations is consistent with a general inhibition of engulfment by sli-1 but suppression of a partial loss-of-function mutation cannot be used to order genes within genetic pathways. In summary, sli-1(lf) was not able to suppress the engulfment defects caused by complete loss-of-function CED-10 pathway mutants. Thus, the CED-10 Rac pathway is unlikely to act by inhibiting SLI-1; rather, SLI-1 acts either parallel to or upstream of the CED-10 Rac pathway.
sli-1 mutation might decrease the number of unengulfed cell corpses in engulfment mutants in a number of ways other than by suppressing apoptotic cell engulfment. sli-1 mutation could (1) decrease programmed cell death, resulting in fewer cell corpses as is seen in ced-3 caspase mutants [37], (2) alter the timing of corpse appearance during development like the protein CED-8 [38], resulting in fewer corpses at the time of observation, (3) alter cell-corpse morphology so that they could not be identified by DIC microscopy as corpses or, (4) cause the corpses to be unstable and lost rapidly.
To address whether sli-1 normally prevents programmed cell death, we determined whether cells that are known to die by apoptosis normally during development do so in sli-1 mutants. 16 cells undergo programmed cell death in the anterior pharynx during embryogenesis in wild-type animals [39]. The nuclei of these cells are identified easily using DIC microscopy [34]. Mutations in genes that normally cause cell death, such as ced-3 or ced-4, have up to 14 extra recognizable cell nuclei in the anterior pharynx [34], [40]. sli-1(sy143) animals had no more nuclei than wild-type animals in their anterior pharynges (Table 2, sli-1 mutation does not block cell death in the pharynx). To test for apoptosis defects more stringently, we observed whether sli-1 mutation enhanced the death defect of a partial loss-of-function ced-3 mutant (n2427) [41]. We observed no difference between ced-3(n2427) and ced-3(n2427); sli-1(sy143) animals (1.6 vs. 1.1 extra cells, Table 2, sli-1 mutation does not block cell death in the pharynx).
We used time-lapse DIC microscopy to assess whether sli-1 loss-of-function affected the timing, persistence or morphology of cell corpses. The development of wild-type and sli-1(sy143) animals was recorded for approximately 150 minutes. We found that sli-1(sy143) animals developed on average more slowly than wild-type animals. To account for the difference in the rate of development, we counted the number of cell deaths that occurred from the first cell death up to the comma stage. sli-1(sy143) worms take approximately 31 minutes longer than wild-type animals to develop to that stage at 20°C (144 minutes compared to 103 minutes). During this time, approximately 60–65 cell corpses appear in the wild-type animal. The number of cell corpses that appeared and when they appeared in wild-type and sli-1(sy143) embryos did not differ significantly (Figure 2A). However, the timing of appearance approaches statistical significance (p value = 0.053), probably related to the difference in developmental speed. The length of time that corpses persisted was similar in wild-type and sli-1(sy143) animals (Figure 2B). In addition, apoptotic cell corpses in wild-type and sli-1(sy143) animals looked similar (Figure 2C). We conclude that the morphology and time of appearance of apoptotic cell corpses is not affected by sli-1 mutation.
To study the expression pattern of SLI-1, we expressed gfp under control of the sli-1 promoter. Specifically, we fused the 5000 bp 5′ of the sli-1 ATG to gfp and injected that construct into wild type (N2) worms. Fluorescence was seen broadly throughout the embryo beginning prior to gastrulation and continuing through the L1 stage (Figure S1). Higher levels of expression were seen in cells that would form the head beginning at approximately the 1½ fold stage of embryonic development. This pattern continued through the first larval stage with L1 animals showing GFP expression at high levels in the head and at lower levels throughout the body, including in body wall muscles, hypodermis, intestine, anal depressor muscles, and several neurons. During later larval development expression is seen in the distal tip cells (DTCs) (Figure S1 panel ix). In adults, GFP was found in the head, body wall muscles, hypodermis, DTCs and some neurons. This expression pattern is consistent with our results; we observe sli-1-dependent phenotypes in the heads and DTCs (see next section in results).
We also generated a translational fusion with sli-1 containing a C-terminal gfp expressed under control of the sli-1 promoter. This transgene was injected as an extrachromosomal array into ced-10(n1993); sli-1(sy143) animals. In animals in which high levels of GFP were observed, the animals invariably died during embryogenesis with bizarre morphological defects, indicating that overexpression of sli-1 is toxic to worms (Figure 3B). However, in animals with low levels of SLI-1::GFP expression, morphological abnormalities were not seen. We found that in these low level SLI-1::GFP expressing animals, the sli-1 mutant engulfment phenotype was rescued (data not shown).
We analyzed the expression pattern of Psli-1sli-1::gfp in animals that were morphologically normal. The expression profile of this transgene was quite similar to that of the transcriptional fusion but the expression level was far lower (Figure 3A). Interestingly, SLI-1::GFP was observed surrounding cell corpses in most transgenic animals though in a small minority of cell corpses in each animal (Figure 3A, panel i). While this finding suggests that SLI-1 normally is found at the cell-cell interface at some point during the engulfment process, there are several caveats, some of which argue against and others for this interpretation. SLI-1 might not normally surround cell corpses and only does so in animals in which the SLI-1::GFP transgene is overexpressed (although we suspect that it is not overexpressed at that high a level in morphologically normal animals, as we noted above). Another fact that appears inconsistent with SLI-1 normally being present at the interface between the engulfing and engulfed cell is that most corpses seen on DIC were not surrounded by GFP haloes. However, at least two phenomena could account for the lack of more GFP haloes. First, the embryos where we could analyze unengulfed cells had comparatively low levels of SLI-1::GFP expression, which would decrease the sensitivity of the assay. Second, since SLI-1 inhibits engulfment, it might need to be removed from the cell-cell interface for engulfment to occur. Thus, SLI-1 might only surround cell corpses briefly before being relocated within the engulfing cell.
In mammals, the SLI-1 homolog Cbl is found primarily in the cytoplasm, but also at the plasma membrane and bound to the cytoskeleton [25]. In our transgenic lines in which SLI-1::GFP was overexpressed at high levels, GFP was seen preferentially at the cell periphery and less so in the cytoplasm (Figure 3B) though it is unclear if this localization is physiological given the overexpression. Furthermore, these embryos were very abnormal morphologically so conclusions regarding subcellular localization in these animals should be made very cautiously.
Because high levels of SLI-1::GFP cause embryonic lethality we were unable to test whether unengulfed cell corpses accumulated in animals with high levels of SLI-1::GFP. However we did analyze animals with low levels of SLI-1::GFP. We observed the number of unengulfed cell corpses in N2 embryos that were morphologically normal at the two-fold stage and contained the sli-1::gfp transgene and compared them to N2 animals without the transgene. Embryos that contained the transgene had 9.2 unengulfed corpses vs. 8.1 for animals without the transgene (Table S1). While this difference is statistically significant (p<0.04), it is unclear if this represents a biologically significant difference; this is not surprising since the amount of overexpression appears to be low. At hatching there was no significant difference in the number of unengulfed cell corpses between animals with or without the transgene.
Studies of C. elegans mutants partially defective in programmed cell death (such as partial loss-of-function ced-3 mutants) demonstrated that engulfment dysfunction can enhance apoptotic defects [41], [42]. These studies concluded that engulfment of dying cells promotes their apoptosis. Similar promotion of cell death by engulfment has been observed in Drosophila [43], indicating that the cell-killing effect of engulfment is evolutionarily conserved.
In partial loss-of-function ced-3 mutants, such as n2427, some of the cells fated to die will begin the dying process (based on morphological appearance) but then recover and survive. However, in animals with engulfment gene mutations as well as partial ced-3 loss-of-function mutations a much larger percentage of cells normally fated to die survive.
We compared the number of extra nuclei in the pharynges of ced-12(tp2); ced-3(n2427) and ced-12(tp2); ced-3(n2427); sli-1(sy143) animals to determine if sli-1 loss-of-function could suppress the apoptotic defect of an engulfment pathway mutation. Fewer extra nuclei were seen in animals that contained the sli-1(sy143) mutation (Table 2, sli-1 suppresses the cell-killing effect of an engulfment gene), demonstrating that SLI-1 suppresses the cell-killing promoted by engulfment genes, consistent with it engulfment suppression role.
The two distal tip cells (DTCs) migrate during development from the center of the animal outward and then back again, meeting approximately in the center of the animal. As they move, the gonads form behind them, resulting in two U-shaped gonads [44], [45]. In ced-10 Rac pathway mutants, the gonads often have extra turns or arms caused by abnormal DTC migration [46]. We tested whether sli-1 mutation suppressed the DTC migration defects of ced-10 Rac pathway mutants. Mutation of sli-1 decreased the percentage of gonadal morphology defects in all ced-10 Rac pathway mutants tested, including null ced-5 and ced-12 mutants (Figure 4). 48% of the gonads of ced-5(n1812) animals were abnormal whereas only 29% of the gonads of ced-5(n1812); sli-1(sy143) animals were abnormal (p<0.008), while the percentages of abnormal gonads in ced-12(n3261) and ced-12(n3261); sli-1(sy143) animals were 40% and 12%(p<4.9×10−6). These data demonstrate that SLI-1 inhibits DTC migration and that it does so independent of the CED-10 Rac pathway. Notably, since the CED-1 pathway has no role in DTC migration, SLI-1 appears to act in parallel to both engulfment pathways.
In summary, loss of sli-1 suppresses ced-10 Rac pathway DTC migration defects but does not suppress ced-10 Rac pathway engulfment defects. At least two models could account for these findings. SLI-1 could act through one molecular pathway to inhibit apoptotic cell engulfment (e.g. the CED-10 Rac pathway) and through another molecular pathway to inhibit DTC migration. Alternatively, SLI-1 might act in a common pathway to inhibit both engulfment and migration but the relative importance of that pathway might be much greater in DTC migration than in engulfment. This difference would account for the ability of sli-1 mutation to suppress CED-10 Rac pathway DTC migration defects but not CED-10 Rac pathway engulfment defects.
For example, the CED-10 Rac pathway and another SLI-1-inhibited pathway might both promote DTC migration and either pathway alone is sufficient for normal DTC migration. If this were the case, loss of SLI-1 function would derepress the SLI-1-regulated pathway and suppress DTC migration defects even if the CED-10 Rac pathway were completely non-functional, as we observed. Engulfment, however, might be totally dependent on the CED-10 Rac pathway. In this case, even if sli-1 loss-of-function derepressed the other parallel pathway, the defect caused by loss of the CED-10 Rac pathway might not be able to be overcome by derepression of the SLI-1-regulated pathway. We favor this model (i.e. that SLI-1 acts on the same parallel pathway in both engulfment and DTC migration) both because of its parsimony and because of data we will present later in the paper (See last paragraph of the section titled SLI-1 acts independently of ABL-1).
To determine whether SLI-1 function is required in the dying cell or the engulfing cell, we used sli-1 mutant animals containing a sli-1 transgene that was expressed under the control of heat shock promoters (protocol adapted from Wu and Horvitz (1998) [46]). Specifically, the number of cell corpses in the heads of newly hatched worms was counted within 300 minutes of heat shock. Since all apoptotic deaths in the heads occur prior to 300 minutes before hatching, sli-1 could not be expressed in the dying cells.
Expression of sli-1 in ced-10(n1993); sli-1(sy143) animals increased the number of unengulfed corpses in L1 heads from 13.4 to 23.7 (p<1×10−4) (for comparison, ced-10(n1993) animals had 20.0 corpses (Table 1)), whereas expression of a gfp-only control transgene did not increase the number of unengulfed corpses (17.3 vs. 16.4; p>0.2) (Table 3). Notably, in the gfp-expressing animals, GFP was not seen in the cell corpses, in support of our hypothesis that the engulfed cell did not make new proteins (data not shown). Thus, expressing sli-1 outside of the engulfed cell rescues the sli-1 mutant phenotype, indicating that sli-1 acts in the engulfing cell.
Five proteins have been identified in C. elegans that inhibit the engulfment of apoptotic cells. Three of them, the myotubularin lipid phosphatase MTM-1, the adapter SWAN-1 and the RacGAP SRGP-1, act through the CED-10 Rac pathway [20]–[23]. It is unknown how the C. elegans progranulin, PGRN-1, suppresses engulfment defects [24]. Genetic and biochemical data indicate that ABL-1 inhibits ABI-1 in parallel to the CED-10 Rac pathway [14].
Since abl-1 and sli-1 both act independently of the ced-10 Rac pathway, we asked whether sli-1 and abl-1 act in the same pathway. We generated triple mutant strains containing mutations in an engulfment gene and in abl-1 and sli-1 and compared the engulfment defects and DTC migration defects to those of double mutant strains containing mutations in engulfment genes and either abl-1 or sli-1. We found that the engulfment defect of the null mutant ced-1(e1735) was suppressed to a greater degree by the combination of abl-1(ok171) and sli-1(n3538) than by either mutation alone (Table 4). The same phenomenon was observed for the partial loss of function ced-10(n1993) allele. The engulfment defect of the null mutant ced-5(n1812) was not suppressed by the abl-1 or sli-1 mutations together or alone, consistent with our prior results that neither sli-1 nor abl-1 loss-of-function can suppress null defects in the ced-10 Rac pathway. The ced-6(n2095) engulfment defect was suppressed by both the abl-1(ok171) and the sli-1(n3538) alleles, but they did not enhance each other. The ced-6(n2095); sli-1(n3538) strain had 10.2 unengulfed corpses while the ced-6(n2095); abl-1(ok171) sli-1(n3538) strain had 10.4 unengulfed corpses. While it is not clear why these mutations did not enhance each other in the ced-6 mutant background, the suppression by sli-1(n3538) is very strong and we suspect that we are near the threshold of the sensitivity of the engulfment assay so that further enhancement cannot be detected despite independent effects on engulfment.
For the DTC migration defect, ced-5(n1812) was suppressed by both abl-1(ok171) and sli-1(n3538) and was significantly more suppressed by the combination of the two mutations (Figure 5). By contrast, the DTC migration defect of ced-10(n1993) was suppressed so effectively by sli-1(n3538) that the addition of the abl-1(ok171) mutation did not enhance the suppression, similar to what was observed in engulfment with the ced-6(n2095)-containing strains. However, it appears that there is a trend towards increased suppression with sli-1 and abl-1 mutations together though the difference does not reach statistical significance (Figure 5).
abi-1 encodes the only C. elegans homolog of Abi, a member of the Wave Regulatory Complex (WRC). A combination of genetic and biochemical data suggest that ABL-1 and the CED-10 Rac pathway both act on the WRC through ABI-1 in parallel to each other: CED-10 Rac activates ABI-1 and ABL-1 inhibits it. Since SLI-1 acts in parallel to ABL-1, we asked whether it also acts on ABI-1. The only abi-1 mutations in existence (and abi-1 feeding RNAi) are quite weak and have no effect on engulfment alone but do enhance the engulfment defects of mutations in other engulfment genes. Therefore, we analyzed the effects of abi-1 mutation in combination with another engulfment mutation. Specifically, ced-1(e1735) null mutant animals containing combinations of mutations of abi-1 and/or sli-1 were assessed for the magnitude of their engulfment defects. sli-1(sy143) suppressed the engulfment defect of ced-1(e1735) animals in the presence or absence of the abi-1(tm494) mutation (Figure 6A). ced-1(e1735) L1 animals had 25.3 unengulfed corpses and ced-1(e1735); abi-1(tm494) animals had 35.0 corpses. ced-1(e1735); abi-1(tm494); sli-1(sy143) animals had 30.1 corpses. Similar findings were found for ced-5(n1812) mutants (Figure 6B). We also tested the effect of abi-1 on DTC migration using the ced-5(n1812) null mutation (Figure 6C). Similar to the findings with ced-1 in engulfment, sli-1(sy143) suppressed the DTC migration defect of ced-5(n1812) (48% vs. 29%) and sli-1(sy143) suppressed the DTC migration defect of an abi-1(tm494); ced-5(n1812) double mutant (49% vs. 26%). Thus, mutation of abi-1 did not completely suppress the effect of sli-1 on engulfment or DTC migration. abi-1(tm494) abolishes the ability of abl-1 null mutations to suppress defects in engulfment and DTC migration [14]. While these results do not prove that sli-1 acts in a different pathway from abi-1, the findings are in stark contrast to those for abl-1, since abi-1 mutation does not abrogate the effects of a sli-1 null mutation on engulfment and DTC migration. Thus, abi-1 might act independently of the WRC.
The finding that sli-1(sy143) suppresses the abi-1(tm494) engulfment defect in the presence of a ced-5(n1812) null mutation (Figure 6B) supports our model that sli-1 acts in parallel to the ced-10 Rac pathway rather than upstream of the ced-10 Rac pathway in engulfment. The ced-5(n1812) mutation totally inactivates the ced-10 Rac pathway. If sli-1 acted upstream of the ced-10 Rac pathway, the ced-5(n1812) mutation would block the ability of sli-1(sy143) to suppress the abi-1(tm494) engulfment defect, which we did not observe.
SLI-1 inhibits the LET-23 EGFR/LET-60 Ras pathway and is thought to do so by ubiquinating the LET-23 protein, targeting it either for destruction or sequestration [29], [30]. Mammalian Ras activates Rac. Therefore, it was plausible that SLI-1 might inhibit engulfment by suppressing the LET-23/LET-60 pathway and consequently decreasing activation of CED-10 Rac by LET-60. To test this possibility, we generated strains doubly mutant for engulfment genes and the gain-of-function mutation let-60(n1046gf). We would expect gain-of-function mutations in this pathway to suppress engulfment defects if sli-1 normally inhibits this pathway. We found no consistent effect on the number of unengulfed apoptotic cells in animals with or without the let-60(n1046gf) mutation (Figure 7). One allele of ced-12 was slightly enhanced while another allele of ced-12 and an allele of ced-2 were slightly suppressed. The only significantly modulated mutation was ced-6(n2095), which was suppressed. Possibly this effect reflects a gene- or allele-specific interaction with let-60. Regardless, this pattern does not phenocopy either sli-1 mutation. Thus, sli-1 does not appear to act through the let-23 EGFR/let-60 Ras pathway to inhibit engulfment.
To determine which domain of SLI-1 is required for its suppression of engulfment and DTC migration defects, we ectopically expressed truncated forms of SLI-1 under control of the C. elegans heat-shock promoters in sli-1 mutant animals. The SLI-1 protein contains three domains, an N-terminal domain that binds tyrosine kinases (and several other proteins), a RING finger, which mediates its E3 ubiquitin ligase function and a C-terminal domain, which contains several proline-rich regions. Minigenes encoding wild-type sli-1 and truncation mutants of sli-1 lacking each of the three domains expressed under heat-shock promoter control were injected into ced-10(n1993); sli-1(sy143) worms. These constructs were generated previously [30] and generously provided to us by Paul Sternberg. ced-10(n1993); sli-1(sy143) larvae harboring extrachromosomal arrays were incubated for one hour at 33°C, and their gonadal morphologies were analyzed 30 hours later in young adults. The arrays contained sli-1 minigenes encoding full-length sli-1 or sli-1 lacking the N-terminus, RING finger or C-terminus (sli-1wt, sli-1ΔN, sli-1ΔRING or sli-1ΔC, respectively). Figure 8 shows that the sli-1wt construct rescued the defect completely, while sli-1ΔRING and sli-1ΔC both partially rescued the defect and the sli-1ΔN did not rescue the defect at all. We also tested the sli-1ΔRING transgene in engulfment and found that it partially rescued the engulfment suppression defect (Table 3).
Thus, the N-terminal tyrosine kinase binding domain was strictly required for the function of sli-1 in DTC migration, whereas the RING finger and C-terminus were at least partially dispensable, suggesting that the ubiquitin ligase activity is unlikely to be central to the role of sli-1 in DTC migration. Consistent with our findings in DTC migration, the RING finger was also partially dispensable in engulfment.
We have demonstrated that SLI-1 negatively regulates the engulfment of apoptotic cells. sli-1 inhibits the engulfment process as well as the migration of distal tip cells during gonadogenesis and the engulfment-related cell-killing process. Our genetic analysis suggests that SLI-1 acts in a manner that does not require the known engulfment pathways. Ectopic expression experiments indicate that SLI-1 acts in engulfing cells and that its function is dependent on its N-terminal tyrosine kinase binding domain. Interestingly, these experiments demonstrate that the ubiquitin ligase function of SLI-1 is at least partially dispensable.
In mammals, the SLI-1 homolog Cbl interacts physically interacts with the CED-2-related protein Crk, Abl, Abi2 and regulates the activity of the CED-10 homolog Rac [28], [31]. In addition, in both mammals and worms, SLI-1 Cbl downregulates LET-23 EGFR by ubiquitination [29]. These interactions provided the rationale for our study of SLI-1 in engulfment initially. However, we found that the effects of SLI-1 on engulfment were independent of all of these proteins (with the possible exception of ABI-1; we were unable to test an abi-1 null mutant). This finding highlights the multiple roles signaling proteins play in the regulation of complex cell biological processes. Also, these data emphasize the value of genetic analyses in discerning the physiological relevance of physical interactions discovered in vitro for a particular process.
Like many other genetic suppressors, sli-1 mutation has no effect on normal engulfment. Specifically, only two engulfment suppressors, srgp-1 and pgrn-1, have been shown to increase the rate of clearance of apoptotic cells in wild-type animals whereas abl-1, swan-1 and mtm-1 do not do so [14], [20]–[24]. Notably, the srgp-1 and pgrn-1 effects are subtle ones seen in early embryos. Possibly, the engulfment process is so efficient that derepressing it by removing inhibitors has little or no demonstrable effect. Similarly, only srgp-1 and mtm-1 cause engulfment defects when overexpressed. However, overexpression of a protein does not always result in increased activity; activation of the protein might be required, explaining the lack of overexpression phenotypes. In the case of sli-1, overexpression is toxic to worms so our ability to discern whether overexpression caused increased cell corpse accumulation was limited.
The discovery of SLI-1 as an inhibitor of engulfment adds to the small list of engulfment inhibitory proteins. Moreover, our genetic analysis puts SLI-1 into a new genetic pathway. Specifically, sli-1 loss-of-function mutations suppress the engulfment defects of ced-1 pathway null mutations and the DTC migration defects of ced-10 Rac pathway null mutations. Thus, SLI-1 could act in a molecular pathway in parallel to both the ced-1 and ced-10 Rac pathways or it might act downstream of one or both pathways. However, the ced-1 pathway has no role in DTC migration, so it is unlikely that sli-1 acts downstream of the ced-1 pathway given its effect on that process. Also, sli-1 loss-of-function mutations do not suppress the engulfment defects of ced-10 Rac pathway null mutations and therefore cannot be downstream of the ced-10 Rac pathway. Thus, the simplest model consistent with the data is that sli-1 acts in parallel to both ced-10 Rac and ced-1 pathways.
abl-1, another inhibitor of engulfment and DTC migration defects, has a very similar pattern of interactions with the two core engulfment pathways, demonstrating that it, too, acts in parallel to the ced-1 and ced-10 Rac pathways. We show that abl-1 and sli-1 act in parallel to each other in these processes as well. Thus, SLI-1 defines a new pathway of inhibition of engulfment and DTC migration.
The genetic interactions between abl-1 and abi-1 and sli-1 and abi-1 differ considerably. Whereas even very weak loss-of-function of abi-1 completely suppresses the effects of abl-1 mutations on engulfment and DTC migration, the same abi-1 mutation only minimally suppresses the effect of sli-1 on these processes. These findings are consistent with a model in which sli-1 acts independently of the Wave Regulatory Complex in engulfment and DTC migration though we cannot conclude that since abi-1 null mutants were not used in the analysis.
Most of our understanding of the function of SLI-1 comes from mammalian studies of its homolog c-Cbl in cell culture. These studies have demonstrated a large number of protein-protein interactions. To discover which of these interactions might be relevant to the engulfment inhibitory function of sli-1, we tested which domains were required to rescue SLI-1 function. The only essential domain was the N-terminal TKB domain. While our studies do not preclude a role for the C-terminal proline-rich or RING finger domains, they do indicate that these domains are not central to the engulfment and cell migration functions of SLI-1.
The TKB domain includes three motifs: a four helix bundle, a Ca++ binding EF hand and an SH2 domain. These three motifs together define a unique domain that binds phosphotyrosines of protein tyrosine kinases [26]. This binding, in turn, allows the E3 ubiquitin ligase function of the RING finger of Cbl to ubiquitinate and target these tyrosine kinases for destruction or sequestration. However, since the RING finger domain, which is required for ubiquitination, is partially dispensable for inhibition of cell migration by SLI-1, the above mechanism cannot explain our results.
In addition to tyrosine kinases, several other proteins have been shown to interact with the N-terminal TKB domain. One of them is APS, an adapter protein that is involved in insulin signaling [47]. However there is no obvious APS homolog in C. elegans. Furthermore, APS signaling requires the C-terminus of Cbl in mammals and the phenotypes we describe only partially require the C-terminal domain. Another interactor, SLAP, the Src-like adapter protein, also binds to the N-terminus of Cbl [48]. It, too, has no obvious homolog in C. elegans.
A third TKB domain interactor is tubulin. Alpha and beta tubulin bind to the Cbl N-terminus [49], [50], and Cbl co-purifies with tubulin in B-cell lysates [51]. The idea that an interaction between SLI-1 and tubulin is involved in engulfment suppression is intriguing for several reasons. First, it would support a role for microtubules in apoptotic cell engulfment, which until now has been shown to be regulated solely by actin cytoskeletal rearrangement. Second, it would fit with our genetic findings concerning sli-1. Specifically, sli-1 inhibits both engulfment and DTC migration, two processes totally dependent on appropriate cytoskeletal regulation. Third, sli-1 appears to act in parallel to all known engulfment genes and engulfment inhibitors. That, too, would be consistent with sli-1 action affecting an entirely different molecular pathway, namely one regulating microtubules.
The discovery that sli-1 acts through a pathway in parallel to the two core engulfment pathways (ced-10 Rac and ced-1) suggests that there are still other cell biological processes involved in apoptotic cell engulfment yet to be discovered. Since the two core pathways were discovered over 20 years ago, it begs the question of why these processes were not identified previously. Possibly, defects in the unidentified processes result in embryonic lethality so they were not identified in genetic screens. Alternatively, these pathways are redundant with the core pathways and, therefore, would only be discovered in the absence of one or both of them. Regardless of the answer, the existence of other pathways suggests that very tight control of engulfment is required during development.
Much of the work on engulfment has been aimed at identifying which signals from the dying cell activate the ced-10 Rac and ced-1 pathways. Our findings suggest that in addition to the need for positive signals, engulfing cells require multiple inhibitory signals to prevent inappropriate engulfment. As discussed earlier, engulfment of dying cells promotes their programmed cell deaths. Potentially there are circumstances during development when cells are particularly susceptible to engulfment-mediated death, which, unless prevented, would result in excess cell death and developmental errors. Perhaps these inhibitory pathways exist as a failsafe mechanism to prevent such errors.
C. elegans strains were maintained at 22°C as described [52]. The N2 Bristol strain was used as the wild-type strain. Animals were grown on NGM plates and fed OP50 bacteria [4], [53]. The mutations and integrants used were: LGI: ced-1(e1735, n2091), ced-12(n3261, tp2); LGIII: abi-1(tm494), ced-6(n2095), ced-7(n1996); LGIV: ced-2(n5101), ced-3(n2427), ced-5(n1812), ced-10(n1993), dpy-13(e184sd), let-60(n1046gf); LGV: unc-76(e911), nIs96 [41]; LGX: abl-1(ok171), nIs106 [41], sli-1(n3538, sy143). Mutant alleles for which no citation is given were described previously [54]. Information about ok and tm alleles can be found at www.wormbase.org (tm alleles were kindly provided by S. Mitani, Tokyo Women's Medical University, Japan). The following balancer chromosomes were used: LGI; LGIII: hT2[qIs48], LGII: mIn1[mIs14], LGIV; LGV: nT1[qIs51].
We isolated ced-2(n5101) from a C. elegans deletion library; genomic DNA pools from the progeny of EMS or UV-TMP mutagenized animals were screened for deletions using PCR as described [55]. ced-2(n5101) removes 637 nucleotides from chromosome IV, 242 base pairs 5′ to the ced-2 ATG, the entire first exon (439 bp) and 12 bp of the first intron.
Unengulfed apoptotic corpses were visualized in the heads of young larvae as refractile discs directly using Nomarski differential interference contrast (DIC) microscopy [56], [57]. Apoptotic cell corpses were counted in the heads of first larval stage (L1) animals within 30 min of hatching, except for animals treated with RNAi (see below). Animals were anaesthetized in 30 mM sodium azide in M9 [53] and viewed using DIC optics on a Zeiss Inverted Axio Observer compound microscope (Thornwood, NY, USA). For animals treated with feeding RNAi, L1 animals were picked, and those with gonads that had not passed the 4-cell stage (all within 60 minutes of hatching) were viewed as described above. p values for pairwise comparisons were calculated using the Student's t test.
For quantitation of cell-death defects in the anterior pharynx, animals in the third larval stage (L3) were anaesthetized and viewed with DIC microscopy as described above. Briefly, the locations of the nuclei of the 16 cells that undergo programmed cell death in the anterior pharynx are known [39]. In wild-type animals by the L3 stage, all of those nuclei have disappeared; any nuclei in these locations in the animals examined at the L3 stage were scored as extra cells. p values for pairwise comparisons in the pharynges were calculated using Student's t test.
Single embryos were placed on agar pads, sealed with petroleum jelly and viewed at 20°C using a Zeiss Inverted Axio Observer compound microscope equipped with Nomarski DIC accessories, a Zeiss AxioCam HRm digital camera and Zeiss Axiovision image acquisition software. Pictures were taken every 3 min for 200 min, and images were analyzed beginning with the appearance of the first cell corpse and ending at the comma stage. The time of appearance of each corpse was recorded. For each time point, 60–80 serial z sections at 0.4 µm/section were recorded. Images were analyzed with ImageJ64 1.45 s (http://imagej.nih.gov/ij) using the plugin Cell Counter. p values for comparisons between strains were calculated using the Wilcoxon rank-sum test.
Adult animals 18 h after the mid-fourth larval stage (L4) were anaesthetized and viewed as described above in Quantitation of engulfment defects and gonads were visualized [44], [58]. Only gonads that were completely visualized were scored. Specifically, gonads that were partially occluded by other structures were not scored. DTC migration was scored as defective when the gonad was morphologically abnormal (extra turn, two arms or bizarre twists) or when the gonad was short or long. Gonadal length was defined as abnormal when the gonad tip was distal to the ipsilateral spermatheca (short) or distal to the contralateral spermatheca (long). The vast majority of abnormalities were in morphology rather than in length. p values for pairwise comparisons were calculated using Fisher's exact test.
For the transcriptional GFP fusion, a PCR product encoding the 5 kb genomic fragment upstream of the M02A10.3a (sli-1) start site was made with SalI/XbaI ends. The product was then digested with SalI and XbaI and ligated to pPD95.75 from the Fire Lab C. elegans kit (Addgene). The resulting plasmid contained the 5 kb upstream of M02A10.3a adjacent to gfp (GFP[S65C]). The plasmid was injected into gonads of N2 animals with the coinjection marker Punc-122::rfp (50 ng/µl for each with 50 ng/µl 1 Kb Plus DNA Ladder (Invitrogen) to a total concentration 150 ng/µl). Three independent transgenic lines were observed and photographed using fluorescence and DIC microscopy. For the translational GFP fusion, we used in vivo recombination (http://wormbook.org/chapters/www_reportergenefusions/reportergenefusions.html). Fosmid WRM0611cB12 was digested with MscI and SpeI, generating a 9 kb fragment which includes 5 kb of sequence upstream of M02A10.3a and 4 kb of the M02A10.3a sequence. To make the second fragment, a 5 kb full length M02A10.3a sequence was PCR amplified from fosmid WRM0611cB12 and then inserted into vector pDEST-MB14 using the Gateway method (Invitrogen), resulting in an in-frame fusion of M02A10.3a with GFP at its C-terminus. Then this plasmid was cut with PstI and SacII, making a 6 kb fragment including the C-terminal 4.5 kb of M02A10.3a fused with gfp and some additional sequence from pDEST-MB14. The 2 fragments were mixed with the co-injection marker Pmyo-2::rfp (50 ng/µl for each with 50 ng/µl 1 Kb Plus DNA Ladder to a total concentration 200 ng/µl) and injected into the gonads of ced-10(n1993);sli-1(sy143) animals. Three independent transgenic lines were analyzed. All lines demonstrated rescue of the sli-1 engulfment suppression defect.
Phspsli-1wt, Phspsli-1ΔN, Phspsli-1ΔRING, and Phspsli-1ΔC were described previously [30]. Briefly, sli-1ΔN encodes the first 64 amino acids of SLI-1 followed by a short linker (Leu Ala Leu) and then amino acid (aa) 350 through the end of the protein (aa 583). sli-1ΔRING encodes the first 393 amino acids followed by the following linker (Glu Thr Gly Thr Thr Phe Glu) and then amino acid 432 through the end of the protein. sli-1ΔC encodes the first 447 amino acids of the protein. The Phspgfp plasmids have been described previously [59]. Each minigene was expressed under the control of the hsp16/2 and hsp16/41 promoters. Phsp plasmids were injected into ced-10(n1993); sli-1(sy143) animals at a concentration of 20 ng/µl with a plasmid containing myo-2::rfp as a coinjection marker at 5 ng/µl and with 35 ng/µl of 1 Kb Plus DNA Ladder for a total concentration of 80 ng/µl per injection. The pharynges of transgenic animals were RFP-positive. For quantification of unengulfed apoptotic cell corpses, embryos were grown at 20°C, heat-shocked for one hour at 33°C, placed at 20°C for up to four hours after which cell corpses in the heads of newly hatched first larval stage (L1) animals were counted. For quantification of DTC migration defects, animals were heat shocked for one hour at 33°C and placed at 22°C for 30 hours. DTC morphology in young adults was then analyzed in an equal number of animals with and without the transgenic arrays. 200 gonad arms were analyzed per genotype. Two independent transgenic lines were analyzed for each transgene combination except for the engulfment analysis of Phspsli-1ΔRING, in which only one line was used. This was because only one of the lines produced viable L1 larvae after heat shock during embryogenesis. Attempts were made with five separate lines. We presume this line had lower expression levels based on the fact that high expression levels of SLI-1 proteins are toxic to worms. Also, the line that produced viable larvae had comparatively faint GFP staining.
Animals were fed bacteria that contained either the RNAi empty feeding vector L4440 [60] or an RNAi feeding vector with part of the abi-1 gene, B0336.6, cloned into it. We obtained the abi-1 feeding plasmid from Open Biosystems (Huntsville, AL, USA). The DNA sequence of the clone was determined to verify its accuracy. Feeding RNAi was performed as described [60], [61]. Briefly, bacteria were grown in liquid culture overnight and then transferred to NGM plates containing 1 mM isopropyl-D-β-thiogalactopyranoside (IPTG). Fourth-larval stage (L4) animals were placed on these plates and 24 h later were transferred to fresh plates. Progeny were tested for engulfment or DTC migration defects.
|
10.1371/journal.ppat.1003379 | Increase in Gut Microbiota after Immune Suppression in Baculovirus-infected Larvae | Spodoptera exigua microarray was used to determine genes differentially expressed in S. exigua cells challenged with the species-specific baculovirus SeMNPV as well as with a generalist baculovirus, AcMNPV. Microarray results revealed that, in contrast to the host transcriptional shut-off that is expected during baculovirus infection, S. exigua cells showed a balanced number of up- and down-regulated genes during the first 36 hours following the infection. Many immune-related genes, including pattern recognition proteins, genes involved in signalling and immune pathways as well as immune effectors and genes coding for proteins involved in the melanization cascade were found to be down-regulated after baculovirus infection. The down-regulation of immune-related genes was confirmed in the larval gut. The expression of immune-related genes in the gut is known to affect the status of gut microorganisms, many of which are responsible for growth and development functions. We therefore asked whether the down-regulation that occurs after baculovirus infection affects the amount of gut microbiota. An increase in the gut bacterial load was observed and we hypothesize this to be as a consequence of viral infection. Subsequent experiments on virus performance in the presence and absence of gut microbiota revealed that gut bacteria enhanced baculovirus virulence, pathogenicity and dispersion. We discuss the host immune response processes and pathways affected by baculoviruses, as well as the role of gut microbiota in viral infection.
| Baculoviruses are large DNA viruses that infect invertebrates, mainly insects from the order Lepidoptera. They were first discovered to cause insects' epizootics and are now used worldwide as biocontrol agents. Extensive studies on baculovirus biology led to the discovery that they can serve as expression vectors in insect cells; recently they have also been considered as vectors for gene therapy. Baculovirus infection, like many other oral infections, starts with the invasion of the gut by viruses; the gut is a compartment colonized by a community of resident microbiota. In this study, we observed that baculovirus infection leads to the decreased expression of immune-related genes in a Spodoptera exigua cell culture as well as in the larval gut. Gut microbial loads were found to increase after baculovirus infection. A series of bioassays showed that the baculovirus performs better in the presence of microbiota in the gut. Our study shows that baculovirus infection leads to increase of microbiota loads in the gut and that the gut microbiota play a significant role in insect immunity and susceptibility to viral infections. These findings suggest that gut microbiota can be manipulated to improve biocontrol strategies that employ baculoviruses.
| Baculoviruses are large DNA viruses that infect arthropods, mainly insects from the orders, Lepidoptera, Diptera and Hymenoptera. During their replication cycle, they produce two distinct morphological forms, the occlusion-derived virus that is responsible for transmission of the infection between insects and the budded virus that is responsible for spreading the infection within one insect host [1]. Baculoviruses have a long history of being used as microbial insecticides to control insect pest populations in forestry and agriculture [2]. With the development of insect cell cultures, it became possible to replicate baculoviruses in vitro and to investigate their replication cycle and the mechanisms of infection.
Baculovirus infection impacts a host in many ways. Not all tissues are equally affected and the infection process does not develop in the same way in different cell types. Typically, successfully infected host larvae die 3–10 days after the initial infection, and death is often accompanied by larval body liquefaction [3]. Although pathological effects are easily observed particularly in the later phase of infection, the processes that take place in the host in the early stages of infection are still not well understood.
Host-pathogen interaction can be viewed as an arms race between two enemies. The pathogen evolves to optimize host infection and its dispersion, and the insect attempts to improve protection against the pathogen. Insects' responses to virus infections include inducible reactions, such as immune responses as well as physical and chemical barriers that prevent viruses from establishing the infection [4]. Two types of immune responses are known, innate and adaptive. Whereas in vertebrates the immune system is composed of both, in invertebrates only innate responses are present. Innate responses include cellular and humoral response, which share the same signalling pathways [5]. Cellular responses refer to hemocyte-mediated responses like phagocytosis, nodulation and encapsulation. Humoral responses include identifying the invading microbes by pattern recognition proteins and subsequently synthesizing antimicrobial peptides (AMPs). The production of AMPs in insects is regulated by the signalling pathways Toll, Imd and JAK-STAT. Cellular and humoral responses depend on each other and interact in order to clear invading microbes from the organism. Both can lead to the activation of phenoloxidase cascade and subsequent melanization and to the production of reactive oxygen species [5].
Most of the studies on insect immune systems have been performed in Drosophila as a model organism. Research on Drosophila antiviral responses clearly indicate that antiviral responses differ from antimicrobial responses [6]. While in antibacterial and antifungal responses, Toll and Imd are the leading signalling pathways, in antiviral responses, the JAK-STAT pathway seems to play a main role [7]. Antiviral responses include the RNA interference (RNAi) machinery, which is especially active against RNA viruses; though recent studies show that it also contributes to fighting DNA viral infections [8], [9]. Although most of the studies of the insect immune system have been described in Drosophila, research on baculovirus infection of Lepidoptera suggest that they have a similar antiviral defence system. Host defense pathways implicated in resisting baculovirus infections include melanization and encapsulation [4]. Melanization depends on the prophenoloxidase (PPO) pathway, which, as in blood-clotting systems in vertebrates, leads to the isolation of the pathogen [10], [11]. For instance in Helicoverpa zea, which is semi-permissive to AcMNPV, the processes of melanization together with encapsulation were shown to be able to eliminate viral foci and attenuate disease progression [12]. Lately, it has also been suggested that factors other than only hemocytes' defensive responses may be involved in antiviral defense. Hirai et al. 2004 found that in a Bombycidae host, Antheraea pernyi, baculovirus infection induced the expression of hemolin, an insect immune protein, and that silencing of hemolin affected the progress of viral infection [13]. However, a more recent study was unable to detect changes in the expression of hemolin after infection with baculovirus in two species of Lepidoptera, suggesting that hemolin does not participate in the response to virus infection in all insects [14]. Recently, a number of transcriptional and proteomic studies of host gene expression have shown that immune-related genes are regulated after baculovirus infection [15]–[20], which may indicate that these genes are involved in the response to baculovirus.
After being ingested by the insect, baculoviruses reach the gut, one of the main physical barriers for the establishment of a systemic infection in the larvae. Immune aspects in insects were initially studied in the hemolymph and fat body, tissues traditionally attributed to immune responses. More recently and after recognizing the role that the gut plays in shaping immunity in mammals [21], research in insect immunity has also focused on the gut. Gut epithelial tissue has been found to express pattern recognition proteins as well as antimicrobial proteins, which indicates that the gut, in addition to the hemolymph, plays an important role in the immune responses of insects [22]–[24]. The immune system faces an enormous challenge especially in gut tissue as in addition to the invading pathogens, it must continuously confront microbiota present in the gut lumen. To date, little is known about the interaction between gut microbiota and baculovirus infection.
Gut microbiota are involved in various physiological functions, including digesting food and protecting against pathogen infections by occupying attachment sites, competing for nutrients and preventing colonization [25]. Recent studies also suggest that commensal microbiota present in the gut modulate insect immune responses [26], [27] and insect susceptibility to pathogens such as Bacillus thuringeiensis [28], [29]. In vertebrates, it is well recognized that gut commensal microbiota play a crucial role in the immune responses to pathogens and, to a great extent, shape gut immune system [21], [30], [31]. Studies show both promotion [32], [33] and suppression [34], [35] of pathogen invasion by gut microbiota in vertebrates. In invertebrate pathology, the effects of gut microbiota on pathogens have been also long studied, mainly in bacterial and fungal infections [36]–[39].
Numerous studies have shown that virulence and the pathogenicity of baculoviruses vary considerably depending on the food plant of the host [40]–[42]. Recently researchers have noted that plants that are consumed by insect larvae in large amounts shape gut commensal microbiota, and the observed differences in virulence depending on the food plant may be due to the differences in gut bacterial communities that in turn shape gut immunity. A new line of research, nutritional immunology, has emerged to study relationships between diet, immunity, and pathogenic processes and gut microbiota [43], [44]. Commensal bacteria communities in insects depend on diet and taxonomy [45]. Each insect species contains its particular microbiota, and the relationship of microbiota and the intestinal immune system is most often described as dynamic homeostasis. The mechanisms of maintaining gut homeostasis are obviously complex, taking into account continuously changing conditions such as food composition, pH, gut enzymes, fluctuating amounts of non-pathogenic bacteria, pathogenic bacteria and other pathogens, among other factors.
The expression of many host proteins is undoubtedly influenced by baculovirus infection. The knowledge of genes affected by virus infection may serve to improve baculoviruses as heterologous protein expression vectors [15] and as biocontrol agents [46]. Although previous studies report on changes in host gene expression after baculovirus infection [15]–[20], the consequences of such expression changes in insect physiology and resulting interactions with the virus have hardly been addressed [46]. In this study, we move beyond a descriptive nature of expression data to search for changes in the host gene expression which are reflected in host phenotype. We first present a comprehensive analysis of the host gene expression patterns of Spodoptera exigua cells infected with two baculoviruses, the highly specific S. exigua nucleopolyhedrovirus (SeMNPV) and the generalist Autographa californica nucleopolyhedrovirus (AcMNPV). A high throughput microarray analysis was performed in the Se301 cell line in order to take advantage of the homogeneous conditions that can be reached in cultured cells. Among the regulated genes after baculovirus infection, we observed the down-regulation of a large number of immune-related genes including antimicrobial peptides. After observing similar regulation by quantitative real-time polymerase-chain reaction (qRT-PCR) in larval guts, we investigated in more detail the impact of viral infection on gut microbiota, finding that baculovirus infection has an important effect and that intestinal microbiota play an important role in baculovirus pathogenesis.
Transcription profiles of close to 30,000 S. exigua unigenes were investigated in Se301 cells infected with the species-specific baculovirus SeMNPV and the non-specific, broad-range AcMNPV, in order to decipher host genes' responses and assess the extent of specific and general responses to viral infections. Only a very small number of genes were differentially regulated at 4 hours post infection (hpi) with the applied thresholds, 74 and 70 genes for SeMNPV and AcMNPV, respectively (Fig. 1A). At 12 hpi, the number of differentially regulated genes increased in cells infected with both SeMNPV and AcMNPV. However, the number of unigenes regulated after SeMNPV infection was three times higher than the number of genes regulated after AcMNPV infection (1000 and 344, respectively), suggesting that the gene transcriptional response to a species-specific virus is stronger and/or faster than the response to a non-specific virus. The highest number of differentially regulated unigenes was observed at 36 hpi, with an almost 4-fold difference between the response to SeMNPV and the response to AcMNPV. Over 8000 unigenes were differentially regulated after infection with SeMNPV, which exceeds 25% of all unigenes represented on the array. For AcMNPV infections, in contrast, only around 2600 unigenes (<10%) were differentially regulated. In neither case was the host gene expression shut-off observed, with 3177 up-regulated and 5053 down-regulated unigenes for SeMNPV and1341 up-regulated and 1250 down-regulated unigenes, for AcMNPV infections. For SeMNPV and AcMNPV infections, 43% and 27%, respectively, of differentially regulated unigenes showed homology to genes from public databases (Fig. 1B). Genes differentially regulated at 4, 12 and 36 h after SeMNPV and AcMNPV infections that have homology to genes present in public databases are listed in Supplementary Tables S2–S7.
In order to select differentially responding unigenes common for both baculovirus infections, the lists of differentially regulated unigenes after SeMNPV and AcMNPV infections were compared. To reduce the possible influence of differences in the development of infection between both baculoviruses, the lists of unigenes from all three time points post-infection were pooled. In total, we found 1279 unigenes that were regulated after SeMNPV infection as well as after AcMNPV infection (Fig. 1C). Except for a few unigenes, most of common unigenes showed the same patterns of regulation, meaning that unigenes up-regulated in the host after AcMNPV infection were also up-regulated after SeMNPV infection, and unigenes down-regulated in the host after AcMNPV infection were also down-regulated after SeMNPV infection. Moreover, the regulation of individual unigenes, independent of the direction (up or down), was in general stronger in SeMNPV infected than in AcMNPV infected host cells. A large portion of unigenes differentially regulated by both baculovirus infections showed homology to genes from public databases (the top 10 differentially regulated annotated unigenes are presented in Table 1). A considerable portion of unigenes differentially regulated (14%) showed homology to genes involved in pattern recognition and immune responses, including antimicrobial effector proteins, such as attacin, gloverin and cecropin. Interestingly, most of the regulated immune-related unigenes were down-regulated in response to baculovirus infection. Among the most down-regulated unigenes, we found genes with homology to peptidoglycan recognition protein D (>30-fold, for SeMNPV infection), prophenoloxidase-activating enzyme 3 (>30-fold), beta-1,3-glucan recognition protein 2a (>15-fold), lysozyme-like protein 1 (>12-fold), pattern recognition serine proteinase (>10-fold), gloverin (>10-fold), attacin (>7-fold), immune-related Hdd23 from Bombyx mori (>6-fold), and cecropin (>5-fold). Moreover, unigenes with homology to ABC transporters, known to be affected in other DNA virus infection, such as herpes, HIV or hepatitis B virus (Hinoshita et al., 2001), were found to be down-regulated.
Given the high proportion of immune-related genes among genes regulated after baculovirus infection (by both viruses), an immune-related gene search was performed for each virus separately (Fig. 2). Immune-related unigenes were selected based on the relationship of their gene ontology terms with immune responses as well as their sequence homology to known immune-related genes. The immune-related genes that were identified constitute about 1% (263) of all unigenes represented on the Sexi-array. Many of these immune-related unigenes are regulated after SeMNPV and AcMNPV infection, with 136 unigenes and 89 unigenes showing expression differences, respectively. The majority of the immune-related unigenes were found to be down-regulated, with 72% (98 unigenes) and 100% (89 unigenes) down-regulated after SeMNPV and AcMNPV infections, respectively. The list of immune-related unigenes and their differential regulation fold-change values are presented in Supplementary Table S8.
To validate the microarray results and extend them to in vivo conditions, 13 genes differentially regulated in Se301 after viral infection (several antimicrobial peptides, a pattern recognition protein, a phenoloxidase activating enzyme and a lectin) were selected; the response of these 13 genes to the viral infection was tested by qRT-PCR. Given that many immune-related genes are differentially regulated, which is generally attributed to insect defense against bacterial and fungal infections, we extended this study to two groups of larvae: one harboring laboratory levels of microbiota (referred to here as “with microbiota”) and another lacking microbiota (referred to here as “w/o microbiota”). Larvae “with microbiota” were reared on standard artificial diet [47] lacking antibiotics to maintain their natural laboratory microbiota levels, while larvae “w/o microbiota” were reared on the standard artificial diet supplemented with 0.2 g/l streptomycin; this supplement has previously been shown to remove a major part of culturable bacteria from the gut [28]. We also selected genes known to be involved in Toll and Imd pathways, and the JAK-STAT signalling pathway, the three main immune signalling pathways. Hypothesizing that gut tissue is where microbiota are located and interact with the insect, we searched for selected genes in larval guts using qRT-PCR.
Most of the selected unigenes showed the same patterns of regulation in infected S. exigua guts (Fig. 3A). In the genes studied, the biggest change was observed for the prophenoloxidase-activating enzyme. The peptidoglycan recognition protein (PGRP) and the β-glucan recognition protein (β-GRP) as well as antimicrobial effectors attacin and cecropin B also displayed clear down-regulation upon baculovirus exposure. Gloverin was down-regulated by infection in cell culture but no regulation in the larval gut tissue was observed. Among the three immune signalling pathway genes assayed in infected larvae, only the Imd gene was found to be down-regulated. Although the expression pattern was similar in larvae with and without microbiota, the level of regulation was in general stronger in larvae with microbiota.
A decrease in the expression level of immune-related transcripts, including immune response effectors such as antimicrobial peptides known to target bacteria, suggested that viral infection could be impacting the intestinal microbiota homeostasis. To test this hypothesis, S. exigua larvae were infected with SeMNPV, and culturable bacterial loads were determined and compared to the non-infected controls. A diet without antibiotics was used in this experiment to avoid the effects of antibiotics on viral infection. A significant increase (18.2-fold) in the intestinal bacterial loads was observed in the larvae infected with SeMNPV, in comparison to non-infected controls (Fig. 3B). Two types of bacterial colonies were observed among the culturable bacteria from the gut of S. exigua. 16S rRNA typing of these colonies revealed that they belong to the Enterococcus and Enterobacter genera, with approximately 100-fold higher counts for Enterococcus. In order to discount the possibility that bacteria could be leaking from the gut into the hemolymph due to the viral infection and lead to the observed differences in gene expression, we have also tested the bacteria in the hemolymph of our larval samples (data not shown). No significant differences in bacterial counts were observed between infected and control larvae.
We tested whether the decrease in the level of immune-related transcripts as well as the increase in the microbial loads in infected larvae has an impact on baculovirus pathogenesis. In other words, we aimed to determine if the down-regulation of immune-related genes in the intestine provides an advantage to the host or to the pathogen. For that purpose, S. exigua larvae with and without microbiota were infected with SeMNPV, and the infection process was monitored in terms of mortality and time to death, viral occlusion bodies (OBs) production and liquefaction.
Time to death, as a measure of virulence, was assessed for larvae with and without microbiota by comparing their survival plots using the Kaplan-Meier method. Larvae with microbiota died faster (median survival 132 h) than larvae without microbiota (median survival 150 h), when infected with 104 OBs/larva (p<0.0001, Gehan-Wilcoxon test). For this dose the mortality was 99% among the larvae with microbiota and 89% among the larvae without microbiota (Fig. 4A). When infected with a lower dose, 103 OBs/larva, a significant difference in survival time was again observed between the larvae with and without microbiota (p = 0.0138, Gehan-Wilcoxon test). The mortality was 71% for the larvae with microbiota and 65%for the larvae without microbiota (Fig. 4B).
OBs' productivity and liquefaction capability were also calculated for both larvae harboring gut microbiota and those lacking gut microbiota. Larvae with microbiota produced 2.8-fold (t-test, p<0.05, t = 5.8, df = 4) more OBs than did larvae without microbiota (Fig. 4C). Similarly, the release of OBs to the environment was higher (1.5-fold, t-test, p<0.05, t = 4.3, df = 4) for larvae with gut microbiota in comparison to larvae without gut microbiota (Fig. 4D).
The species-specific S. exigua microarray (Sexi-array) was developed in this study in order to elucidate sets of genes that are regulated after baculovirus infection. The microarray data from cells infected with two different nucleopolyhedroviruses, SeMNPV and AcMNPV, showed two main trends that occur in the cells after infection. The first was the lack of a general down-regulation of the host gene expression (host gene shut-off) that, according to published data, was expected to take place within 36 hpi. The second was the down-regulation of genes involved in pattern recognition and immune responses.
The present study extends the previous works in other lepidopteran species (both in vivo and in vitro) on the differential expression of host genes after infection with baculoviruses. Host gene shut-off at late times post-infection was commonly observed; however, the responses seem to be specific for the insect species, the infective virus and even for the studied tissue. The first report on how baculoviruses regulate hosts' RNA levels [48] showed that the expression of three conserved Spodoptera frugiperda genes, actin, hsp70 and histone, decreased dramatically between 12 and 18 hpi in Sf21 cells infected with AcMNPV. This early study has been recently complemented by a microarray-based comprehensive analysis of host genes in AcMNPV-infected Sf21 cells [15]. Out of 42,000 probes included in the array, 70% were differentially regulated, with nearly all of the genes being down-regulated. Earlier studies using the differential display approach [20] also showed a global down-regulation of Sf9 genes after AcMNPV infection. Our results show that baculovirus-infected Se301 cells did not entirely follow this pattern. We did observe that there were more genes down-regulated than up-regulated as the duration of infection increased; nevertheless, even at 36 hpi many genes were up-regulated. Similarly, and although tested at earlier times post-infection, no global down-regulation of host genes has been reported in Helicoverpa zea cells HzAM1 infected with Helicoverpa armigera single nucleopolyhedrovirus (HearNPV) [49] and after injection of S. exigua larvae with AcMNPV [16], at 18 and 12 hpi, respectively. These results suggest that the extent of host gene shut-off strongly depends on the specific host-virus interaction.
Insect immune responses were originally thought to be a response to bacteria and fungi [50], although recent studies have also shown that innate immunity may be involved in viral infections [51]. In our study, gene expression analysis revealed that immune-related genes are down-regulated after the infection of S. exigua cells with two different baculoviruses. We also observed this effect in the gut tissue of S. exigua larvae infected with SeMNPV. Confirming our results, Choi et al. [16] recently described in S. exigua (entire larvae) the down-regulation of immune-related genes 12 h after the injection of AcMNPV in the larval hemolymph. However, few studies focusing on the insect hemocytes or fat body have reported the induction of immune-related genes after baculovirus infection. Although a general immune induction due to the injection of foreign material could occur, Moreno-Habel et al. and Wang et al. [17], [52] found that gloverins were induced in the hemocytes of Trichoplusia ni and H. armigera larvae injected with AcMNPV. In in vitro studies, preincubating AcMNPV virions with Sf9 supernatant containing recombinant gloverins decreased viral infectivity, which led the authors to conclude that gloverin has antiviral activity in AcMNPV-infected Sf9 cells [52] and to speculate that the membrane-bound gloverin could act as a virus entry molecule. Hirai et al. [13] found hemolin had been induced in the pupal fat body after baculovirus injection. Considering the possible antiviral activity that has been associated with some immune-related proteins, the down-regulation of immune-related proteins observed in the gut could be a mechanism the virus relies on to improve its ability for a successful infection. Viruses, including baculoviruses, have been reported to manipulate host physiology for their own benefit [53]–[55].
On the other hand, the down-regulation of immune-related genes in the gut was expected to have an impact on gut microbiota. Evaluating the influence of baculovirus infection on bacterial loads in the gut, we found that baculovirus infection led to an increase in gut microbiota; this increase may be related to a decrease in antibacterial activity due to the observed down-regulation of immune-related genes. As speculated above, the down-regulation of immune-related genes in the gut as a result of virus infection may directly benefit the virus, but the role of the third participant in the system, gut microbiota, must also be accounted for. After observing that gut bacterial loads change after baculovirus infections, we measured how the baculovirus infection process proceeds in the presence and absence of gut microbiota. Four parameters of viral fitness were measured in larvae with and without gut microbiota after they were infected with a baculovirus: pathogenicity (measured as larval mortality), virulence (measured as time to death), viral multiplication, and viral dispersion (measured as OBs' productivity and liquefaction, respectively). For all tested parameters, the virus benefited from the presence of gut microbiota. Are increased loads of gut bacteria a side-effect of virus infection? Or are we observing a more complex mechanism according to which the down-regulation of immune defenses against gut bacteria is orchestrated by the virus for its own benefit? Likewise, if bacterial numbers are increasing, could a virus infection also be directly beneficial to the gut microbiota? These questions point to directions for future research. What remains clear is that the role of gut microbiota cannot be ignored when studying virus infection.
Assuming that a virus influences insect immunity for its own benefit, a question arises: what is the mechanism by which baculoviruses manipulate host immune responses? Recent studies have shown that baculoviruses encode microRNAs that target both viral and host genes [56]. In the Bombyx mori NPV genome, several microRNAs were computationally predicted [57], and the effect of one of them in increasing the viral load has been recently confirmed [58]. Among the targets predicted were host defense elements such as prophenoloxidase and hemolin, which suggests that virus miRNAs constitute a tool with which host immune systems can be manipulated. More genetic information from S. exigua would facilitate the identification of baculovirus-encoded miRNAs that may be targeting some of the down-regulated immune-related genes.
We also explored the regulation of the main immune-related pathways in insects, namely Toll, IMD and JAK-STAT, by analyzing the expression of key genes in these pathways. In our study, neither the Toll-like receptor nor the transcription factor STAT was found to be regulated. In contrast, the Imd gene was found to be regulated in both in vitro and in vivo experiments. In both cases, the decline in the expression of Imd was observed. We speculate that the IMD pathway may be involved in the response to baculovirus infection and that baculoviruses may manipulate microbiota loads by down-regulating this specific signalling pathway. Likewise, the Imd pathway has also been shown to be involved in the viral response to RNA viruses (Alfaviruses) [59].
Another question that remains to be elucidated is related to the mechanisms underlying the enhancement of baculovirus pathogenicity by gut microbiota. Bacterial colonization and pathogenicity relies on the synthesis and secretion of virulence factors. Virulence factors enable microorganisms to establish themselves on or within a host of a particular species and enhance their potential to cause disease. These virulence mechanisms are also relevant for commensal microbiota, which are not lacking virulence factors and, once they invade hemocoel, can ultimately lead to insect death [60]. Bacterial virulence factors might contribute to the success of viral infections. For example, various types of bacteria produce chitin-degrading enzymes [61], [62]. In our case, chitin degradation by gut bacteria may have contributed to the establishment of the viral infection. Baculoviruses also produce chitinase in order to degrade host tissues and facilitate viral dispersion [63]. Therefore, chitin-degrading factors from both bacteria and viruses, may have acted simultaneously, increasing the viral dispersion observed in our study. Both in a previous study and here we observed that laboratory culturable microbiota in our colonies of S. exigua are mainly composed of species of the genus Enterococcus [28] and Enterobacter. Virulence factors from Enterococcus spp. include several exoenzymes such as gelatinases, hyaluronidase, and serine proteases [64], [65]. Those enzymes may have contributed to the spread of the virus by degrading the extracellular matrix. Additional studies should be undertaken in order to discover bacterial elements that contribute to baculovirus pathogenicity.
Microbiota have to be considered when studying host-pathogen interactions especially for orally infecting pathogens [66]. In the case of viral pathogens, the influence of microbiota can be either protective [34], [55] or harmful [32], [33], [67]. Our study shows that the baculovirus infection process increases gut microbiota loads in the insect host and that such an increase enhances baculovirus fitness. The gut architecture of invertebrates should be taken into account when studying virus-host interactions and the system should be perceived as composed of four elements: virus-microbiota-host-diet. The challenge arises when we ask how we can manipulate the gut microbiota to optimize the use of baculoviruses for the biocontrol of insect pests. Manipulating symbiotic bacteria has been shown to reduce insect-vectored human diseases, such as dengue virus, after antiviral protection was associated with the presence of Wolbachia spp [68]. Interestingly, in the lepidopteran S. exempta, Wolbachia spp. were found to increase the host's susceptibility to baculovirus infection. An approach based on exploiting bacterial endosymbionts for more effective pest control has been proposed [69]. Similarly, the manipulation of gut microbiota could be considered in biocontrol strategies in order to increase pest susceptibility to baculoviruses as well as a way to reduce viral susceptibility in other organisms.
Baculovirus-free S. exigua larvae were kindly provided by Primitivo Caballero from the Public University of Navarra, Pamplona, Spain. The colony was regularly checked for the presence of baculovirus infection by qRT-PCR and proved to be negative for SeMNPV. The insects were reared at 25°C, relative humidity of 70% and a photoperiod of 16 h∶8 h (light∶dark), on artificial diet [47]. The viruses used in the experiments were SeMNPV (SP2 and Ox4 isolates) [70], [71] provided by Primitivo Caballero from the Public University of Navarra, Pamplona, Spain and AcMNPV (C6 isolate) [72]. S. exigua Se301 cells were maintained in HyQ-SFX insect cell culture medium supplemented with 5% FBS at 25°C, according to standard cell maintenance procedures.
S. exigua L4 larvae were orally infected with SeMNPV and AcMNPV virus at doses of 104OBs/larvae and 105OBs/larvae, respectively. Larvae were offered a suspension of OBs added to a small piece of diet. Larvae that ingested diet plugs containing virus suspension within approximately 2 h were provided fresh diet and considered in the assay. Four days post-infection (dpi), the hemolymph was collected from infected larvae by cutting prolegs. The hemolymph was filtered by 0.45 µm filters, and Se301 cells were infected at a high multiplicity of infection (MOI) with the budded viruses (BVs) present in the hemolymph for four hours. Five days post infection (dpi), the BV-containing supernatant was collected from the cells and filtered as before; virus titers were then calculated in Se301 cells using an end point dilution assay [73]. Viruses prepared in such a way were subsequently used to infect Se301 cells. A control sample was prepared in the same way, although L4 S. exigua larvae were mock-infected.
A 44K Agilent oligonucleotide chip was designed to include more than 40,000 probes from close to 30,000 S. exigua unigenes. The sequences of S. exigua were obtained from a S. exigua transcriptome sequencing project described elsewhere [74]. The S. exigua Agilent custom microarray (Sexi-array) was designed based on obtained sequences using eArray application from Agilent. Most of the unigenes were represented by two 60-mer oligonucleotide probes, designed to target different fragments of each unigene.
Se301 cells were infected by incubation with SeMNPV and AcMNPV at a MOI of 5 for four hours. At 4, 12 and 36 hours post-infection (hpi), cells were collected and total RNA was extracted using RNAzol reagent (Molecular Research Center, Inc., Cincinnati, OH), according to the manufacturer's protocol. To further purify the RNA, an RNAeasy Kit (Qiagen, Hilden) was used following the manufacturer's protocol. The quality of RNA was assessed by an Agilent 2100 Bioanalyzer using the EukaryoteTotal RNA Nano protocol. Three independent biological replicates were performed for each treatment and each time point.
Agilent One-Color Spike-in Mix was added and 600 ng of total RNA was used for cRNA (complementary RNA) synthesis. 1.65 µg of the resultant cRNA was fluorescently labelled with cyanine-3-CTP, fragmented and hybridized to S. exigua microarray slides following the One-Color Microarray-Based Gene Expression Analysis (Quick-Amp labelling) protocol from Agilent. S. exigua microarrays were scanned using a G2505B Agilent scanner and data were extracted using Agilent Feature Extraction 9.5.1 software. Before data analysis, hybridization quality control reports were verified as correct. RNA labelling and hybridization as well as array scanning and data extraction were performed by the Microarray Analysis Service of Principe Felipe Research Centre (CIPF), Valencia, Spain.
Data analysis was performed using free Babelomics 4.3 software (http://babelomics.bioinfo.cipf.es/) [75]. First, arrays were normalized using RNA spike-in probe data and quantile normalization methods. Normalized arrays of the samples treated with the virus were compared to controls at 4, 12 and 36 h after treatment and expressed as a fold-change in the gene expression. The thresholds of fold-change ≥2 and p-value<0.05 were applied. T-tests were used to compare gene expression between the arrays, both those infected with virus and the control, for each time point and each treatment.
S. exigua larvae reared on two types of diets were used in the in vivo validation experiment. One group was reared on the standard diet lacking antibiotics to maintain natural laboratory levels of microbiota in their guts. The other group was reared on the standard diet supplemented with streptomycin (0.2 g/l) to deplete intestinal microbes. Streptomycin was previously shown to remove a major part of culturable bacteria from the larval guts [28].
S. exigua L4 larvae from both types of diet were individually challenged with SeMNPV at a dose of 104 OBs/larvae by adding virus suspension to a diet plug. Control larvae were fed with water. Six larvae were used per each treatment and condition, and the experiment was performed three times. Guts from six larvae were collected at 72 hpi and pooled, and total RNA was extracted using RNAzol reagent. 1 µg of RNA was used for cDNA synthesis. RNA was first treated with DNase I (Invitrogen) and subsequently reverse-transcribed to cDNA using oligo-d(T) primer and SuperScript II Reverse Transcriptase (Invitrogen) according to the manufacturer's protocol. Thirteen genes were selected for validation with qRT-PCR, including several antimicrobial peptides and some immune-related genes. Additionally, as representatives of the three main signalling pathways in insect immunity, S. exigua homologs for Toll, Imd, and JAK-STAT pathway genes [74] were also included in the analysis. An ATP synthase gene served as a reference gene for qRT-PCR data normalization. Selected gene primer sets were designed using Prime Express software from Applied Biosystems (Supplementary Table S1). qRT-PCR was carried out in the ABI PRISM 7000 from Applied Biosystems. All reactions were performed using Power SYBR Green PCR Master Mix (Applied Biosystems) in a total reaction volume of 20 µl. Four µl of the 1∶5 diluted cDNA templates was added to each reaction. Forward and reverse primers were added to a final concentration of 300 pM. Expression ratios were calculated based on the formula (Ctgene of interest, treated−Ctreference gene, treated)/(Ctgene of interest, control−Ctreference gene, control), 2−ΔΔCt, which assumes 100% efficiency of all amplification reactions. The standard deviation of the ΔCt values of treated and control samples was calculated as (s12+s22)1/2, where s1 is a standard deviation of gene of interest Ct and s2 is a standard deviation of reference gene Ct. The standard deviation of ΔCt was then incorporated into the fold-difference calculation, after Applied Biosystems “Guide to Performing Relative Quantitation of Gene Expression Using Real-Time Quantitative PCR”.
S. exigua L4 larvae were infected orally with SeMNPV at a dose of 104 OBs/larva. Control larvae were mock-infected. At 4 dpi, larvae were dissected and their guts isolated. For each sample, the guts of five larvae were pooled. The experiment was performed 3 times in quadruplicate. 1 ml of 0.1% Triton X-100 in PBS was added to each sample and the samples were vortexed vigorously. 100 µl of serial dilutions was plated on LB-agar plates and grown o/n at 37°C. Colony-forming units (CFUs) were counted and calculated as CFU/ml.
Bacterial loads were also counted in the hemolymph of infected and control larvae. Larvae infected as above were bled at 4 dpi and 100 µl of serial dilutions of hemolymph were plated on LB-agar plates and grown o/n at 37°C. CFUs were counted and calculated as CFU/ml.
Microarray data analyses were performed in Babelomics as described above. qRT-PCR data were analyzed according to Applied Bioscience guide. The rest of the bioassays were analyzed using GraphPad Prism version 5. Bacteria loads in the gut counts were analyzed using the non-parametric Mann Whitney test and virus performance assays were analyzed with unpaired t-tests. Data are reported as means ± standard deviation (SD). Differences between compared groups were considered significant when p<0.05.
|
10.1371/journal.pntd.0002073 | High Content Screening of a Kinase-Focused Library Reveals Compounds Broadly-Active against Dengue Viruses | Dengue virus is a mosquito-borne flavivirus that has a large impact in global health. It is considered as one of the medically important arboviruses, and developing a preventive or therapeutic solution remains a top priority in the medical and scientific community. Drug discovery programs for potential dengue antivirals have increased dramatically over the last decade, largely in part to the introduction of high-throughput assays. In this study, we have developed an image-based dengue high-throughput/high-content assay (HT/HCA) using an innovative computer vision approach to screen a kinase-focused library for anti-dengue compounds. Using this dengue HT/HCA, we identified a group of compounds with a 4-(1-aminoethyl)-N-methylthiazol-2-amine as a common core structure that inhibits dengue viral infection in a human liver-derived cell line (Huh-7.5 cells). Compounds CND1201, CND1203 and CND1243 exhibited strong antiviral activities against all four dengue serotypes. Plaque reduction and time-of-addition assays suggests that these compounds interfere with the late stage of viral infection cycle. These findings demonstrate that our image-based dengue HT/HCA is a reliable tool that can be used to screen various chemical libraries for potential dengue antiviral candidates.
| Dengue, a re-emergent human disease that places nearly half of the world's population at risk, threatens to further expand in geographical distribution. The lack of an available effective dengue vaccine has encouraged the search for antiviral drugs as an alternative approach. In recent years, drug discovery through high-throughput screening has become a trend in the search for dengue antivirals. In this study, we developed an image-based dengue high-throughput/high-content assay using prevalent viral strains of three dengue serotypes (DENV1, DENV2 and DENV3) isolated from dengue outbreaks in South America and a laboratory-adapted strain of DENV4. We demonstrated the usefulness of our image-based dengue HT/HCA in identifying potential dengue antivirals by screening a small subset of chemical compounds for inhibition of dengue virus infection in a human-derived host cell line (Huh-7.5), and partially characterized their activities against dengue infection in a mosquito host cell line (C6/36), a distantly-related virus (hepatitis C virus), and an unrelated virus that is transmitted by the same mosquito vector (chikungunya virus).
| Dengue virus (DENV) is an important mosquito-borne pathogen responsible for causing dengue fever (DF) and the more severe, life-threatening dengue hemorrhagic fever/shock syndrome (DHF/DSS) [1]. DENV is a small, enveloped virus belonging to the genus Flavivirus, family Flaviviridae. [2]. Dengue virions are approximately 50 nm in diameter [3], containing a single-stranded positive RNA of ∼11 kilobase with a genomic organization: 5′-C-preM-E-NS1-NS2A/B-NS3-NS4A/B-NS5-3′, and flanked by the 5′ and 3′ UTRs (untranslated regions) [4], [5]. There are 4 serotypes of dengue (DENV1, DENV2, DENV3, DENV4), with two or more serotypes commonly found to co-circulate in many dengue endemic areas [6], [7]. The presence of more than 1 dengue serotype in a geographical area contribute to the persistence of epidemics, as immunity acquired against one dengue serotype does not confer long-term protective immunity against heterologous serotypes [8]. Conversely, individuals that have acquired humoral immunity against one dengue serotype may be pre-disposed to DHF/DSS when subsequently infected with a heterologous serotype through antibody-dependent enhancement [9], [10]. Since its re-emergence in 1953, DENV has spread rapidly across 5 continents and more than 100 countries, mostly in tropical and subtropical regions. Present estimates by the World Health Organization place nearly 2.5 billion people at risk of dengue, with approximately 50 million cases of dengue infection and 20 thousand mortalities occurring annually [11].
Due to the global burden of dengue, numerous studies have been done to elucidate the nature of dengue infection and the underlying mechanisms of DF and DHF/DSS. Although ADE is the most widely accepted theory for the occurrence of DHF/DSS, several studies have suggested higher viremia titers, virus serotype and host genetic background as determinants of DHF/DSS [12], [13]. Extensive genetic, cellular and immunological studies have investigated the role of host innate immunity in the events leading to DHF/DSS [14]–[16]. However, a clear understanding on the immunopathology of dengue infections remains elusive.
With the number of DHF/DSS cases rising every year, the demand for a preventive, prophylactic, or therapeutic measure against DENV is growing rapidly. Significant progress has been achieved in the development of a dengue vaccine, including one vaccine candidate (ChimeriVax) that has passed through Phase II clinical trials [17]. However, its long-term efficacy and safety has not been established and mass production of this vaccine candidate to meet the growing demand remains a daunting task. A new approach that is gradually gaining interest is the development of dengue antivirals. For the last 6 years, several drug candidates for HCV and other RNA viruses have been pursued for repositioning as potential drug candidates for dengue [18]. Though at present, none of these compounds have gone beyond pre-clinical trials.
Recent advancements in high-throughput screening (HTS) technologies have contributed to increasing efficiency in the drug discovery process. These include in silico HTS, in vitro enzymatic assays, cell-based reporter assays, image-based whole infection assays, among others [19]. In the work reported here, we describe the development of an image-based high-throughput/high-content assay (HT/HCA) screening method for anti-dengue compounds using an infectious virus system and an innovative approach in image analysis. Using this dengue HT/HCA system, we screened a BioFocus kinase inhibitor library of 4,000 small molecules against DENV1-4 to identify compounds that possess antiviral activity against all 4 serotypes during infection of a human host cell. We counter-screened the primary hits against DENV infection in Aedes albopictus clone C6/36, hepatitis C virus (Family Flaviviridae, genus Hepacivirus) infection, and chikungunya virus (Family Togaviridae, genus Alphavirus) infection to partially characterize the compounds as having a host-specific or virus-specific target. The dengue hit compounds were clustered based on their chemical structures and, together with the activity profile, used to identify scaffolds whose antiviral activity against the 4 serotypes vary depending on the chemical substituents. These scaffolds identified from our dengue HT/HCA screening could be used as potential starting points for the development of dengue antivirals.
The mosquito cell line C6/36 Aedes albopictus clone (CRL-1660), mouse hybridoma cells D1-4G2-4-15 (HB-112), and human hepatocyte Huh-7.5 (PTA-8561, U.S. Patent Number 7455969) were obtained from the American Type Culture Collection. HuH-7 (JCRB0403) was kindly provided by Dr. Katja Fink. Three South American isolates of dengue viruses: Den1 BR/90 (GenBank AF226685.2), BR DEN2 01-01 (GenBank JX073928), BR DEN3 290-02 (GenBank EF629369.1), and the World Health Organization laboratory strain DEN4 TVP-360 were generously provided by Dr. Claudia N. Duarte dos Santos. Hepatitis C virus (HCV) genotype 2a (JFH-1) expressing the NS5a-GFP fusion protein was kindly provided by Dr. Marc Windisch and the chikungunya virus (CHIKV-118-GFP) was a generous gift from Dr. Olivier Schwartz. C6/36 was maintained at 28°C in Leibovitz's L-15 media (Gibco/Invitrogen, USA) supplemented with 5% Fetal Bovine Serum (FBS, Gibco/Invitrogen, USA), 0.26% Tryptose Phosphate Broth (TPB, Sigma-Aldrich, USA) and 25 µg/mL Gentamicin Sulfate (Gibco/Invitrogen, USA) and passaged every 3–4 days. Huh-7.5 was maintained under humidified conditions at 37°C, 5% CO2 in Dulbecco's minimum essential medium/Hank's F-12 (DMEM/F12, 1∶1) (Gibco/Invitrogen, USA) supplemented with 10% FBS and 100 U/mL Penicillin/100 µg/mL Streptomycin (antibiotic solution, Gibco/Invitrogen, USA) and passaged every 3–4 days. HuH-7 was cultured under humidified conditions at 37°C, 5% CO2 in RPMI 1640 containing 25 mM HEPES (WelGene, South Korea) supplemented with 10% FBS and antibiotic solution. Low passaged dengue viruses were propagated for 7 days in C6/36 maintained at 28°C in Virus Medium (VM: Leibovitz's L-15 medium supplemented with 1% FBS, 0.26% TPB, 25 µg/mL Gentamicin) according to previously described methods [20] and titrated by focus formation assay (FFA) using C6/36 as previously described [21]. Dengue virus titers were expressed as focus forming units per mL (ffu/mL). Flavivirus group-specific αE monoclonal antibody 4G2 [22], used as detecting antibody, was prepared from culture supernatant of D1-4G2-4-15 maintained under humidified conditions at 37°C, 5% CO2 in RPMI 1640 containing 25 mM HEPES (WelGENE, South Korea) and supplemented with 10% FBS, 1 mM sodium pyruvate (Sigma-Aldrich, USA), antibiotic solution and 250 ng/mL Amphotericin B (Sigma-Aldrich, USA). 4G2 was concentrated by ammonium sulfate precipitation following previously described methods [23] and purified by protein G affinity chromatography (GE Amersham, Sweden) according to manufacturer's instruction. Total antibody protein was determined by spectrophotometric analysis using the formula of Warburg and Christian [24].
A small target-focused chemical library, comprising of 4,000 synthesized compounds based on ligand binding of known kinase binding sites, was sourced from BioFocus (Galapagos, Belgium). Reference compounds were purchased from TOCRIS Bioscience (Bristol, UK): AZ 10417808, BIBU 1361, LE 135 and MPP and Sigma-Aldrich (USA): Chloroquine, ribavirin and recombinant human Interferon-αA (IFN-α2A). All compounds from the BioFocus kinase inhibitor library and reference compounds were prepared in 100% dimethyl sulfoxide (DMSO, Sigma-Aldrich, USA), with the exception of IFNαA that was prepared in Dulbecco's Phosphate-buffered saline (DPBS, WelGENE, South Korea) containing 5% FBS.
For dispensing of liquid media containing the host cell and viruses, the Thermo Scientific WellMate (Fischer Brand, USA) was used. Dispensing of antibody solutions and other liquid reagents for IFA, including the washing steps, was done using the 96/384-head BioTek EL406 automated liquid washer/dispenser (BioTek, USA).
For miniaturization of the image-based dengue HT/HCA, the following conditions were optimized: a) host cell seeding density, b) multiplicity of infection (M.O.I.) and c) incubation period of infection. For the optimum host cell density, Huh-7.5 cells were prepared at various cell densities and seeded in a 384-well plate, μ-clear black (Greiner Bio-one, Germany). The cells were cultured between 2–4 days at 37°C, 5% CO2. For dengue virus infection, the optimum cell seeding density of Huh-7.5 was inoculated with DENV1, DENV2, DENV3 or DENV4 at various M.O.I. (0.1–5) and cultured between 2–4 days at 37°C, 5% CO2.
An immunofluorescence assay (IFA) used to detect dengue infection was optimized for the dengue HT/HCA. Briefly, cells were fixed with 4% (w/v) paraformaldehyde (PFA) for 20 min at room temperature (Rm T). PFA-fixed cells were treated with 0.25% (v/v) Triton-X for 20 min at Rm T. DENV-infected cells were detected by probing with 4G2 mAb prepared in blocking buffer: DPBS containing 5% FBS for 30 min at 37°C, followed by AlexaFluor-conjugated goat anti-mouse IgG (H+L) (Invitrogen Molecular Probes, USA) prepared in blocking buffer for 30 min at 37°C. Cell nuclei counterstained with 5 µg/mL 4′,6-diamidino-2-phenylindole (DAPI, Sigma-Aldrich, USA). Two washing cycles of DPBS was done after each step of the IFA. After the final washing, digital images were acquired using a high-throughput confocal fluorescence imaging system (Evotec Technologies High-Throughput Cell Analyzer Opera, Perkin Elmer, USA). The digital images were taken from 3 different fields of each well at 20× magnification.
Acquired images were analyzed using our in-house developed image-mining platform (IM). This platform is designed to do high-content screening and directly access the database of images that were sequentially analyzed with specially designed algorithms developed as a customized plug-in to the IM platform. The results of all the analyses were stored in a centralized database. The IM plug-in for dengue HT/HCA works by independently analyzing two separate channels acquired with the Evotec Technologies High-Throughput Cell Analyzer Opera using different algorithms and converging these results to yield the final readout. One channel (DAPI-channel) captures the signal emitted by DAPI-stained nuclei at 450 nm, while the other channel (A488-channel) captures the signal emitted by the AlexaFluor 488 dye bound to the dengue E protein-antibody complexes confined in the cytoplasm of dengue-infected cells at 540 nm. To define the percentage of dengue-infected cells and, conversely the non-infected cells, a modified watershed method was applied. Compared to the original watershed algorithm [25] that uses the morphological gradient image, a weighted gradient image is used for the topographic surface, whose weights are defined as the ridge values computed from the eigen values of the original image. The percentage of non-infected cells, which is also defined as the percent inhibition or percent activity, is derived by using the formula: [1−(A488-positive cells/total cells)]×100%.
The dengue HT/HCA was validated by 1) by infecting Huh-7.5 cells in 384-well plates spotted with 0.5% DMSO with DENV or MOCK plated and 2) observing the dose-response curves of reference compounds previously reported to have anti-dengue activity. In both validation experiments, Huh-7.5 was mixed with DENV1, DENV2, DENV3 or DENV4 at a M.O.I. of 0.5 or VM only for MOCK-infection and dispensed in designated wells using the Thermo Scientific WellMate automated liquid dispenser. For the first validation experiment, the statistical reliability of the dengue HT/HCA was determined by calculating the Z'-factor for the percent infection or percent inhibition. Briefly, the Z'-factor of a defined parameter is calculated using the formula: 1−[(3σp+3σn)/(|μp−μn|)], where the μp, μn, σp and σn are the means (μ) and standard deviations (σ) of the positive (p) and negative (n) controls [26]. For the second validation experiment, the reference compounds were prepared in 2-fold serial dilutions and dispensed in duplicate wells in the 384-well plate, followed by the Huh-7.5 and DENV mixture (M.O.I. 0.5). Percent infection, percent inhibition, and percent cell number were determined using the customized IM platform plug-in. Scatter-plot distribution was generated using TIBCO Spotfire 4.5.0 (TIBCO Software Inc., Somerville, MA). The ten-point dose-response curves (10-pt DRCs) were plotted using the non-linear regression formula: log (inhibitor) vs. response – variable slope (4 parameters), available in GraphPad Prism 5.04 (GraphPad Software Inc., San Diego, CA). The top and bottom values of the reference compounds were unconstrained when the curve fittings of the 10-pt. DRCs were generated.
The BioFocus kinase inhibitor library were screened against DENV1, DEN2, DENV3 and DENV4 at 10 µM in 0.5% (v/v) DMSO. MOCK-infected Huh-7.5 and IFN-α2A (500 U/mL) were used as positive controls, and the 0.5% DMSO vehicle was used as negative control. Four sets of 15 384-well plates (13 plates designated for test compounds and 2 plates designated for DMSO vehicle control) were used for the primary screening of the BioFocus kinase inhibitor library against each dengue serotype. Each of the 4,000 compounds in the library was tested in single wells. For data normalization and quality control of the screening, each test compound plate contained 16 replicates of the positive and negative controls. After dispensing the test compounds IFN-α2A and DMSO vehicle in the 384-well plates, Huh-7.5 was mixed with DENV1, DENV2, DENV3, or DENV4 to achieve a M.O.I. of 0.5 and dispensed at 5×103 cells/well using the Thermo Scientific WellMate automated liquid dispenser. For the MOCK-infected Huh-7.5, the cells were mixed with VM and dispensed under the same conditions as previously stated. Virus infection in the presence of the compounds proceeded at 37°C, 5% CO2 for 96 hrs. Compound activity based on percent inhibition and cell toxicity was assessed by IFA and IM analysis as described above. Scatter-plot distribution of the entire screening was generated using TIBCO Spotfire 4.5.0 (TIBCO Software Inc., Somerville, MA).
The calculated activity was normalized to a percent inhibition (PI) based on the MOCK-infected cells (100% activity, or zero infection) and dengue-infected cells (0% activity, or maximum measured infection percentage) controls according to the formula:where: PImeasured - percent inhibition of test compound
μPIDENV-infect - average percent inhibition readout of dengue-infected control
μPIMOCK-infect - average percent inhibition readout of non-infected control
Similarly, the percent cell number was normalized based on the measured cell number in MOCK-infected (100% cell number) controls according to the formula: % cell number = (Cmeasured/μCMOCK-infect)×100%, where Cmeasured is the measured cell number in the test well and μCMOCK-infect is the average cell number in the MOCK-infected controls. The statistical validity of the dengue high-throughput screening was determined by calculating for the Z'-factor using the 0.5% DMSO-treatment and MOCK-infected Huh-7.5 as negative and positive controls, respectively. In addition, other parameters, including DRC of a reference compound, were used to evaluate the assay performance. For the primary screening a Z'-factor ≥0.5 and a coefficient of variation (CV) among the controls ≤10% was used to validate the results of the assay. The hit (i.e. a compound that demonstrates inhibition of infection) selection criteria for the primary screening was set at ≥80% inhibition of dengue viral infection in at least 1 dengue serotype and with the corresponding percent cell number at ≥50%.
The hits identified from the primary screening were tested at 10 µM for inhibition of dengue virus infection in C6/36. Briefly, cells were inoculated with DENV1∼4 at an M.O.I. of 0.5 and seeded in 384-well plates spotted with the reference and primary hit compounds and incubated for 96 hrs at 28°C. Detection of dengue-infected cells by IFA and image acquisition using Evotec Technologies High-Throughput Cell Analyzer Opera was carried out following the method described above. Compound activity was determined by measuring percent inhibition and percent cell toxicity using the IM platform as previously described.
The hits identified from the primary screening were tested at 10 µM against HCV genotype 2a (JFH-1) infection in Huh-7.5 using an in vitro HCV cell culture system (HCVcc). Cells were seeded in 384-well plates and cultured under humidified conditions at 37°C for 24 hrs. Reference and primary hit compounds were added, followed by inoculation with HCV at an M.O.I. of 1 and incubated for another 72 hrs at 37°C. HCV-infected cells were identified by detection of NS5A-GFP expression using ImageXpress Ultra (Molecular Devices, USA) and analysis using the IM platform as previously described. Compounds resulting in ≥50% inhibition of HCV genotype 2a infection and percent cell number ≥50% were considered as positive hits for anti-HCV activity.
The hits identified from the primary screening were also tested at 10 µM against CHIKV-118-GFP infection in HuH-7 cells and evaluated by resazurin reduction assay (RRA). Resazurin (7-Hydroxy-3H-phenoxazin-3-one 10-oxide) is reduced to the red fluorescent resorufin by redox enzymes produced by viable cells, and is a good indication of metabolic capacity, and by extension cell viability. The amount of converted resorufin was measured as relative fluorescence readout (RFU) at excitation/emission of 531/572 nm using a fluorescence spectrophotometer (Victor3 V Spectrophotometer, Perkin Elmer, USA). Briefly, cells were inoculated with CHIKV-118-GFP at an M.O.I. of 0.5 and seeded in 384-well plates containing reference and primary hit compounds and incubated under humidified conditions for 72 hrs at 37°C. Resazurin solution was added to a final concentration of 10 µM and further incubated for another 12 hrs prior to measurement of RFU. The percent activity of the compounds, reflected by the percent cell viability, was quantified by normalizing against the RFUs of MOCK-infected cells and CHIKV-118-GFP-infected cells. Compounds resulting in normalized RFU ≥70% were considered as positive hits for anti-CHIKV activity.
To confirm the compound activity against dengue viruses, the selected hits from the primary screening were tested in a 10-pt. DRC (2-fold serial dilution from 50 µM) using the same assay described for the dengue HT/HCA. Each concentration of the hit compounds was tested in duplicate wells. Data generated from image analysis of the 10-pt. DRC was plotted and analyzed using the non-linear regression formula: log (inhibitor) vs. response – variable response (4 parameters) in GraphPad Prism 5.04. The EC50 value, defined as the effective concentration resulting in a 50% inhibition of DENV infection, was used to evaluate compound activity. Compound toxicity was determined by testing the hit compounds in a 10-pt. DRC against Huh-7.5 in the absence of viral infection and measuring the cell viability using resazurin reduction assay as described above. The CC50 value, defined as the compound concentration resulting in a 50% reduction in cell viability (based on normalized RFU values) compared with the MOCK-infection, was used to evaluate cell toxicity. Confirmed hits were selected based on their Selectivity Index (SI), a dimensionless value that indicates the magnitude between cytotoxic concentration and effective concentration, and is calculated as: SI = CC50/EC50.
Cluster analysis was done using a molecule-clustering module from Pipeline Pilot (Accelrys Software Inc., San Diego, CA, USA). The active scaffolds of compounds confirmed to have anti-dengue activity through dose response curves were selected for structural analysis. Structural relationship among the hit compounds was analyzed using the Tanimoto coefficient structural similarity [27].
Several phases were involved in developing the image-based dengue high-throughput/high-content assay (HT/HCA). A schematic workflow diagram of the assay development and assay method is shown in Figure S1. The first phase involved miniaturization of the assay to the 384-well plate format, including host cell seeding density and viral infection conditions. Selection criteria for the appropriate cell seeding density was included having a sufficiently high number cells but with enough spatial distribution for proper identification and accurate segmentation by the IM platform plug-in. After testing various seeding densities of Huh-7.5, the seeding density of 5×103 cells per well was selected (data not shown). DENV infection in Huh-7.5 was visualized by immunofluorescence assay (IFA) detection of the dengue E protein using the 4G2 mAb and confocal imaging using the Evotec Technologies High-Throughput Cell Analyzer Opera. For the DENV infection, a M.O.I. of 0.5 and incubation time of 96 hrs was used since it allows for multiple rounds of virus replication and facilitates the screening of active compounds that target different stages of the dengue virus life cycle (Figure S2).
A flowchart of the image analyses is shown in Figure 1. Defining the cell nuclei was done as follows (Figure 1A–D): after applying a Gaussian low pass filter [28] with relatively high sigma value (to the nucleus size) on the DAPI channel (Figure 1A), the local maxima (Figure 1B) were subsequently located. A k-means clustering method [29] was then utilized to separate the background and foreground to obtain the nuclei mask image (Figure 1C) and the local maxima located in the background were removed, leaving the remaining maxima as those representing the number of cells in the image. Starting from the local maxima constrained by the nuclei mask, along with the slightly blurred nucleus image as a distance map, the region for single nuclei were defined (Figure 1D) with the watershed method [25].
Identifying the dengue-infected cells and the percentage of non-infected cells were done as follows (Figure 1D–I): from the image obtained from the A488-channel (Figure 1E), a weight map was calculated based on the edge features of this channel (Figure 1F). After applying an open-by-reconstruction operator and Gaussian low-pass filter to alleviate the noise, a foreground mask was attained (Figure 1G). Starting from the separated nuclei borders (Figure 1D) constrained by the foreground mask, and along with the weight map, the region of the signals were defined and marked with four different colors (Figure 1H), delineating the borders of the cells. Finally, the dengue-infected cells are identified as those having an A488 signal within the defined cell borders above a pre-defined threshold level, and are delineated by blue line segments (Figure 1I).
The first assay validation evaluated the Z'-factors for DENV1-4 infection of Huh-7.5 in the 384-well plate format. Figure 2A shows a representation of the validation process done for the DENV2 HT/HCA. Cells, virus, and a reference control were dispensed in 384-well following a designed template pattern (upper left panel). After the viral infection period and IFA, IM analyses of the acquired images revealed the infection percentage, cell number based on nuclei detection and other pre-defined parameters. An IM analysis showing the relative percentage of DENV2 infection is represented by a generated heat map (lower left panel). The Z'-factor was calculated using the average and standard deviations of the percent infection of the positive and negative controls (right panel). MOCK-infected Huh-7.5 was designated as positive control while the DENV-infected Huh-7.5 was used as the infection control. All wells contained 0.5% DMSO vehicle to simulate the culture conditions used in the screening. The calculated Z'-factors for the DENV1, DENV2, DENV3, and DENV4 HT/HCA in 384-well plates showed a range between 0.50 and 0.75 (Figure S3). According to Zhang et al. [26] a Z'-factor ≥0.5 indicates a statistically reliable separation between positive and negative controls.
The second assay validation tested a panel of reference compounds previously reported to have antiviral properties against different strains of DENV2. This panel includes: AZ10417808, BIBU1316, MPP, LE 135 [30], Ribavirin [31], Chloroquine [32] and IFN-α2A [33]. MOCK-infection and 0.5% DMSO were used as positive and negative controls, respectively. The compounds' antiviral activities and cell toxicities against BR DEN2 01-01 infection of Huh-7.5 were determined by DRC. Figures 2B shows the generated heat map for percent DENV2-infected cells and percent cell viability. Figure 2C shows the DRC of the reference panel, with the percent infection normalized against DENV2-infected Huh-7.5 and percent cell viability normalized against MOCK-infected Huh-7.5. It was observed that all compounds in the reference panel showed inhibition of DENV2 infection in a dose-dependent manner. At very low concentration of the reference compounds, the percent cell viability of DENV2-infected cells did not exceed 75% compared with the MOCK-infected cells, as a consequence of DENV2-associated cytopathic effect. The resulting EC50 of the reference compounds against the DENV2 infection of Huh-7.5 using our dengue HT/HCA varied from those previously reported. Furthermore, most of the compounds in our reference panel exhibited significant cell toxicities at EC50 compared with the MOCK-infected and DENV2-infected controls. Conversely, IFN-α2A concentration ≥EC50 resulted in higher cell numbers compared with MOCK-infected Huh-7.5. Discrepancies between the EC50 of the reference compounds obtained in this study with the previous reports may be attributed to factors such as intrinsic differences between the DENV2 strains and the type of host cell used. Nonetheless, the results of the assay validation demonstrate the statistical reliability of our developed dengue HT/HCA. None of the compounds in the reference compound panel exhibited the ideal EC50 and CC50 values for use in the dengue HT/HCA. While IFN-α2A has shown strong antiviral properties against dengue infection, having a multi-target mode of action restricts its application as a reference drug. Based on these observations, MOCK-infection and IFN-α2A were used as positive controls for the screening of the compound library, but only MOCK-infection was used for calculating Z'-factors and validating the reliability of the entire screening process.
The compounds screened with our dengue HT/HCA is a subset of 4,000 small molecules belonging to the BioFocus kinase inhibitor library of chemical compounds designed to interact with one of the seven representative subsets of kinases according to protein conformations and ligand binding modes [34]. The library was screened at 10 µM against DENV1, DENV2, DENV3 and DENV4, and primary hits were selected based on the criteria: ≥80% activity and ≥50% cell number (Figure 3). The 50% cell number threshold was chosen to allow a wider range of compounds that are slightly cytotoxic at 10 µM, but may still be active at lower concentrations, to be selected. Primary hits were selected for activity against at least 1 dengue serotype. Out of the 4,000 small molecules screened, 157 compounds qualified for further confirmation and counter-screening, giving a hit rate of 3.9%. The primary hits were selected according to activity (≥80% inhibition), irrespective of their cytotoxicity levels. Among the 157 primary hits, 40 compounds (25.5%) showed inhibition of all 4 serotypes, 19 (12.1%) against 3 serotypes, 30 (19.1%) against 2 serotypes, and 68 (43.3%) against 1 serotype.
The inhibitory properties of the dengue primary hits were further investigated by testing these compounds at 10 µM against DENV infection of C6/36, HCV genotype 2a infection of Huh-7.5, and CHIKV-118-GFP infection of HuH-7. The activity profile of these dengue primary hits is summarized in Figure S4. Thirty-nine of the dengue primary hits (24.8%) exhibited ≥50% inhibition against at least 1 DENV serotype in the C6/36 host, suggesting that the targets of these compounds are required for successful DENV infection in both human and insect host cells. It is important to note that even though the other 118 dengue primary hits (75.1%) did not inhibit DENV infection in C6/36 at the same concentration, the putative role of their targets in DENV infection in the insect cells have not been ruled out. Conversely, 103 dengue primary hits (65.6%) showed ≥50% inhibition of HCV genotype 2a infection of Huh-7.5 at 10 µM, with only 19 hits exhibiting <50% cell number in the host cell. In contrast to the high number of overlapping hits between DENV and HCV genotype 2a, only 9 (5.7%) of the dengue primary hits exhibited detectable activity against CHIKV-118-GFP in the resazurin reduction assay. These hits had low antiviral activity, and were excluded after conducting DRC analysis (data not shown).
The activities of the 157 primary hits were confirmed by 10 pt. DRC against DENV1, DENV2, DENV3 and DENV4 infection in Huh-7.5. Cluster analysis of the top 53 compounds exhibiting the lowest EC50 values were performed using a molecule-clustering module from Pipeline Pilot yielded 4 enriched clusters plus singletons. Core structures of the two scaffolds were heterocyclic ring of imidazopyridine and the other two scaffolds were thiazole-based compounds. One of the thiazole scaffold clusters, consisting of 11 compounds, had 4-(1-aminoethyl)-N-methylthiazol-2-amine as a common core structure. The profile of these compounds (EC50, CC50 and Selectivity Index) against the four dengue serotypes and their chemical structures are shown in Table 1 and Figure 4, respectively. The compounds showing a wide spectrum of anti-dengue activity against all four serotypes have only pyridine or pyrimidine ring by amine linkage to the core scaffold and addition of substituents on the ring narrows the spectrum of activity, especially against DENV4. In addition, 9 out of 10 compounds in this cluster have an extra carbon next to aminoethyl linkage at 4th position of thiazole, followed by a phenyl group and trifluoro-, methoxy-, amine or chloride substituents on para position of the phenyl group.
From the time of its re-emergence 60 years ago, dengue has spread across the globe, placing nearly 40% of the world's population at risk of infection. Coincidentally, the geographical distribution of the four serotypes has also expanded, with all serotypes reported to co-circulate in most of the dengue-endemic countries [7]. This has serious implications in the rise of DHF/DSS cases, as it is presently understood that antibody-dependent enhancement combined with elevated cytokine responses resulting from subsequent infection with a heterologous serotype are involved in disease severity [35]. The most advanced dengue vaccine candidates try to address this issue by constructing chimeric YF/DENV virus (ChimeriVax-DEN), incorporating the prM and E genes of each DENV serotype in the yellow fever (YF) 17D backbone, and used these to prepare tetravalent cocktails [36]. However, while the tetravalent ChimeriVax-DEN vaccine has shown good immunological responses in clinical trials [17], its long-term safety and efficacy has not been fully established. In contrast to the dengue vaccine approach, therapeutic drug approach circumvents the immunopathological complication of dengue, and directly addresses the acute viral infection. The work reported here describes the development of a high-throughput/high-content assay screening for potential anti-dengue drugs using image-based quantitation of inhibition of dengue virus infection in vitro as a measure of antiviral activity. This dengue HT/HCA was used to screen 4,000 small molecules from the BioFocus kinase inhibitor library against DENV1-4, and revealed a number of compounds that inhibit more than 80% infection of all 4 serotypes of dengue in vitro. More than 60% of these compounds were also found to inhibit more than 80% infection of HCV genotype 2a infection. Interestingly, most of the compounds belonging to the 4-(1-aminoethyl)-N-methylthiazol-2-amine cluster exhibited measurable antiviral activities against dengue viruses in Huh-7.5, but did not demonstrate strong inhibition of HCV genotype 2a infection in the same host cells.
Recently, high throughput assays (HTA) have been used to find several drug candidates with anti-dengue properties [37]. Structure-based dengue virtual screening (i.e. in silico high-throughput screening or HTS) is a target-based HTA that analyzes the binding potential of chemical compounds against a target dengue viral protein. Using combinatorial libraries and docking programs that predict the chemical interactions, binding potentials of the compounds with known crystal structures and associated ligands of the target protein are computed. This approach has been used extensively in discovering potential inhibitors of DENV E protein binding and fusion [38], [39] NS5 2′O-Methyltransferase [40], NS3 protease [41], [42] and its complex, NS2B/NS3protease (NS2B/NS3pro) [43]. Another target-based HTS approach using enzymatic assay has identified BP2109 as a potential inhibitor of NS2B/NS3pro complex [44]. One drawback of in silico-based HTS is that predicted chemical interactions do not take into account other biological factors. This often results in selection of compounds with high binding properties in silico, but weak activities once tested in vitro. Similarly, target-based enzymatic assays are performed in a highly controlled environment that facilitates optimum enzymatic function of the target, which can be dramatically different from its biological setting. Thus, compounds found to be highly active against the target using this approach may not exhibit the same effect when tested using a cell-based assay, since the assays do not factor in cellular uptake, availability of the target, and other environmental conditions [45]. The in silico and target-based HTA approaches are designed to find active compounds against a specific target. While this helps to simplify the process of identifying the probable mechanism of action, it is inherent in the assays to exclude compounds that may be active against other targets involved in viral infection. Hence, their application to the comprehensive screening of active compounds against viral infection is limited.
In contrast to target-based HTA, cell-based HTA is more robust as it covers a wider aspect of the viral infection process. This type of HTA uses either infectious viruses to follow one or multiple rounds of viral infection, or viral replicons to observe the events surrounding viral replication. Cell-based dengue HTA takes advantage of the various intracellular and intercellular events that occur during viral infection. Hence, antiviral properties can be attributed to either compound activity against viral or cellular targets. Cell-based flavivirus immunodetection (CFI), which measures viral protein expression after infection by ELISA, and luciferase reporter viral replicon assay were used to identify inhibitors of viral RNA synthesis namely, the adenosine nucleoside inhibitor NITD008 [14], NS4B inhibitor NITD-618 [46] and NITD-982, an inhibitor of host dihydroorotate dehydrogenase (DHODH) [47]. A modified type of the viral replicon assay using dengue-1 virus-like particles (DENV1-VLP) assembled by packaging the dengue viral replicon tagged with a Renilla luciferase 2A reporter gene (Rluc2A) in DENV1 structural proteins generated using the Semliki Forest Virus (SFV) expression system was reported to be useful in identifying inhibitors of dengue viral entry, translation and replication [48]. In addition to the expression of viral proteins during infection or replication, other indications of viral infection can be used to assess compound activity such as cell death and reduced metabolic activity. A dengue cytopathic effect (CPE)-based HTA that uses luminescence assay to determine cellular viability by measuring cellular ATP was previously reported [49]. These cell-based HTA are described as “single-readout” assays, since a single value is generated during the assay corresponding to the effect of a particular treatment.
Image-based high-content assay (HCA) is also a form of cell-based assay. Unlike CFI and replicon-based luciferase reporter assays however, it requires more sophisticated equipment like a high-throughput confocal microscope for acquiring images and special software for analyzing image data. When adapting image-based HCA for high-throughput screening, it is more labor intensive and requires more stringent criteria for data analysis. Nonetheless, image-based HCA has one clear advantage over single-readout assays – the amount of information that can be generated from images of a single treatment is not limited to a single value. Aside from the degree of viral infection and cell viability, other interesting information can be extracted from images such as morphological changes in host cell, protein localization, among others [50]. Like other cell-based HTA, image-based assay can be used to screen compounds with diverse modes of activity. This was demonstrated in an image-based HCA screening of 5,362 compounds with diverse chemical structures against DENV2, revealing 73 active compounds, most of which have previously characterized cellular interactions [30]. Dasatinib, a c-Src kinase inhibitor that disrupts the assembly of dengue virions in virus-induced membranous replication complexes, was also identified after screening a kinase inhibitor library using image-based HCA [51].
The image-based dengue HT/HCA developed in this study was used to screen 4,000 compounds belonging to the BioFocus kinase inhibitor library. Primary screening against all four dengue serotypes required 16,000 experiment points (4,000 compounds×4 dengue serotypes), excluding the positive and negative controls. The entire primary screening took 11 days in total: 3 days for expansion of Huh-7.5 from a single T175 tissue culture (TC) flask to 8 T175 TC flasks, 4 days for host cell plating and virus infection in the 384-well plates containing the compounds, 2 days for image acquisition, 1 day for image analysis using the IM platform, and 1 day for data analysis. Two previous image-based screening campaigns for dengue antivirals were conducted with lab-adapted DENV2 (New Guinea C) whole virus [30], [51]. One of these campaigns [51] further investigated the hit compounds by testing the antiviral activities against other lab-adapted dengue serotypes. Our image-based dengue HT/HCA screening campaign differs from the previous image-based HTA in three aspects: First, we used a novel target-focused chemical library (BioFocus kinase inhibitor library) whose collection of small molecules has not been thoroughly screened and characterized. Second, the compounds were screened against low passage strains of field isolated dengue viruses (with the exception of DENV4 tvp360), which allows the identification of compounds that may be active towards prevalent strains. Third, the entire 4,000 compound subset of the BioFocus kinase inhibitor library was screened against all four dengue serotypes. By screening all the compounds of the library against the four dengue serotypes, we can identify novel compounds that may be active against all four serotypes or specific toward any of the serotypes. Such findings can have biological implications on the differences between the viral infection process of the 4 serotypes at the cellular and molecular level. This offers an advantage over the primary screening using the DENV2 serotype and subsequent confirmation of activity with the other serotypes since the latter is already biased towards compounds active against DENV2, resulting in the “loss” of potential hit compounds that do not inhibit this particular serotype.
Screening of the BioFocus kinase inhibitor library using our image-based dengue HT/HCA resulted in the identification of 4 major clusters exhibiting inhibitory properties against dengue virus infection in vitro. Among them, one cluster consisting of 11 compounds having a 2-aminothiazole as a core scaffold, showed antiviral activities of varying degrees against the infection of DENV1, DENV2, DENV3, DENV4 in the human hepatoma cell line Huh-7.5. The inability of these compounds to inhibit dengue infection in C6/36 initially suggests that the target is most likely a factor involving dengue virus infection in human cells. However, this discrepancy may also be attributed to other factors, such as difference in membrane permeability between the two different host cells or molecule uptake of the compounds into the host cell. Such differences in the physiology between the human and insect cells may affect the efficacy of these compounds in inhibiting dengue viral infection, but has not been thoroughly investigated in this study. Interestingly, none of the hit compounds from the 2-aminothiazole cluster significantly inhibited the infection of HCV and CHIKV in hepatoma cells, suggesting that the inhibitory property is more specifically directed towards dengue virus infection.
One major drawback when using cell-based assays in high-throughput screening is the effect of toxicity to the host cells, and by extension, viral infection. Compound toxicity can have a profound effect in the viral infection process, and may lead to inaccurate assessment of the antiviral activity. Since the HTS is conducted using only a single concentration of the compounds, it is impossible to avoid encountering those that exhibit moderate to high level of toxicity. Hence, a confirmatory assay that tests a range of concentration is necessary to verify if these hit compounds indeed have antiviral activities. In addition, secondary assays are used to confirm compound activity and predict the mechanism of action. For cell-based assays that utilize image-based technology, determining compound toxicity with high certainty is more difficult. In the absence of biological markers that detect mitochondrial activity, cell apoptosis, cell starvation, the only indicator of compound toxicity is the relative cell number compared with non-treated controls. This can be misleading if the toxicity does not result in abolition of the cells or degradation of the cell nuclei since the cell number will not reflect the actual number of viable cells. Thus, it is essential to confirm compound toxicity by measuring production of ATP or relative oxygen species as an indicator of cell viability [52].
Time-of-addition assay, a strategy to determine the stage of inhibition during the viral infection cycle, has been used to characterize the mode of action of some inhibitors of dengue viral entry (NITD Compound 6), viral replication (NITD-982) and early translation (NITD-2636) [39], [47], [53]. The inhibitor is added at different time points during viral infection and monitored for expression of the viral proteins, replication of the genome or production of infectious progeny virions. Among the 2-aminothiazole hit compounds identified in this study, CND1203 exhibited a strong antiviral activity against all 4 dengue serotypes, and blocked the formation of dengue virus plaques in Huh-7.5 at 25 µM in the plaque reduction assay (Figure 5). Compound CND1203 was used for the time-of-addition assay, adding 25 µM at different time points (−2 hpi, 0 hpi, 0.5 hpi, 1 hpi, 2 hpi, 4 hpi) of the dengue virus infection (M.O.I. 5) in Huh-7.5. In contrast to the strong inhibition of compound CND1203 against dengue viruses in the plaque reduction assay, none of the treatments inhibited dengue infection in the time-of-addition assay, suggesting that the compound does not interfere with viral entry (data not shown). Aminothiazole-based compounds have previously been implicated in the inhibition of HCV replication by binding to an allosteric site on the viral polymerase [54]. However, the structure of these active anti-HCV compounds differ from the 2-aminothiazole hit compounds reported in this study in terms of the substitutions on the scaffold. As a consequence, it is unlikely that our hit compounds interact with the viral polymerase.
Although compound CND1203 did not inhibit DENV entry, its ability to inhibit the spread of DENV infection and formation of virus plaques in Huh-7.5 clearly suggest that the mode of action is at the post-entry stages. Two kinase inhibitors were previously reported to block virus assembly of dengue virions: Dasatinib, a thiazolyaminopyrimidine that inhibit c-Src protein kinase, did not interfere with dengue RNA replication, but disrupted the proper assembly of dengue virions within virus-induced cell membranous replication complex [51]. SFV785, a trifluorinated N-methylanaline derivative that selectively inhibits NTRK1 and MAPKAPK5 kinase activity, altered the distribution of structural envelope protein from the reticulate network to enlarged discrete vesicles, consequently affecting the co-localization with the DENV replication complex and disrupting the assembly of progeny virions [55]. Interestingly, the 2-aminothiazole hit compounds identified in this study shares the aminothiazole moiety of Dasatinib, as well as the rings on each of the molecule. Based on the structural similarity with Dasatinib, c-Src kinase may be a candidate target of the 2-aminothiazole hit compounds, which would imply a post-genomic replication mode of action. Further investigation is necessary to support this hypothesis.
The persistence of dengue outbreaks around the world, and the lack of an available dengue vaccine reinforce the need to find and develop therapeutic drugs to address this major health concern. The use of HT/HCA screening technologies can expedite the drug discovery of potential dengue antivirals by facilitating the screening of large chemical libraries. The work reported here features an innovative image-based HT/HCA system that can be used as a reliable tool in screening for antiviral compounds against all four DENV serotypes. Furthermore, the compounds identified in the present study can serve as a potential starting point for the development of dengue antivirals.
|
10.1371/journal.ppat.1000249 | Surviving Mousepox Infection Requires the Complement System | Poxviruses subvert the host immune response by producing immunomodulatory proteins, including a complement regulatory protein. Ectromelia virus provides a mouse model for smallpox where the virus and the host's immune response have co-evolved. Using this model, our study investigated the role of the complement system during a poxvirus infection. By multiple inoculation routes, ectromelia virus caused increased mortality by 7 to 10 days post-infection in C57BL/6 mice that lack C3, the central component of the complement cascade. In C3−/− mice, ectromelia virus disseminated earlier to target organs and generated higher peak titers compared to the congenic controls. Also, increased hepatic inflammation and necrosis correlated with these higher tissue titers and likely contributed to the morbidity in the C3−/− mice. In vitro, the complement system in naïve C57BL/6 mouse sera neutralized ectromelia virus, primarily through the recognition of the virion by natural antibody and activation of the classical and alternative pathways. Sera deficient in classical or alternative pathway components or antibody had reduced ability to neutralize viral particles, which likely contributed to increased viral dissemination and disease severity in vivo. The increased mortality of C4−/− or Factor B−/− mice also indicates that these two pathways of complement activation are required for survival. In summary, the complement system acts in the first few minutes, hours, and days to control this poxviral infection until the adaptive immune response can react, and loss of this system results in lethal infection.
| As one of the most successful pathogens ever, smallpox caused death and disfigurement worldwide until its eradication in the 1970s. The complement system, an essential part of the innate immune response, protects against many pathogens; however, its role during smallpox infection is unclear. In this study, we investigated the importance of the complement system in mousepox infection as a model for human smallpox disease. We compared mice with and without genetic deficiencies in complement following infection by multiple routes with ectromelia virus, the causative agent of mousepox. Deficiencies in several complement proteins reduced survival of ectromelia infection. Sera from these same complement-deficient mice also have reduced ability to neutralize ectromelia virus in vitro. In complement-deficient mice, ectromelia virus disseminated from the inoculation site earlier and produced higher levels of virus in the bloodstream, spleen, and liver. The increased infection in the liver resulted in greater tissue damage. We hypothesize that the complement-deficient mice's reduced ability to neutralize ectromelia virus at the inoculation site resulted in earlier dissemination and more severe disease. We have demonstrated that surviving ectromelia virus infection requires the complement system, which suggests that this system may also protect against smallpox infection.
| Poxviruses remain a threat to the human population despite the eradication decades ago of naturally circulating variola virus, the causative agent of smallpox. Smallpox, with its up to 30% mortality rate, could devastate the large unvaccinated population if released accidentally or by bioterrorists [1]. Closely related monkeypox virus has also emerged as a human pathogen [2]. To understand the virulence of smallpox, investigators have turned to related poxviruses like ectromelia virus (ECTV), the causative agent of mousepox. Variola virus and ECTV have a narrow host-range and cause significant morbidity and mortality [3],[4]. The numerous available mousepox-susceptible and -resistant mouse strains allow the components of the protective immune response to poxviruses to be dissected in the natural host.
Disease severity varies among inbred mouse strains, and comparisons of these strains have elucidated factors essential for survival. Mice naturally acquire ECTV via cutaneous abrasions, which is mimicked experimentally with footpad inoculation [4]. Through this route, ECTV infection is 100% lethal in susceptible strains (BALB/c, DBA/2, and A/J) but asymptomatic in the resistant C57BL/6 strain. The C57BL/6 strain has a stronger TH1 type cytokine response and a more robust cytotoxic lymphocyte response than susceptible strains [5]. Lethal infection occurs in C57BL/6 mice that lack CD8+ T cells [6],[7], B cells [7],[8], macrophages [6], natural killer (NK) cells [9],[10], interferon (IFN)-γ [11]–[13], IFN α/β receptor [13], perforin [14],[15], and granzyme A or B [16]. Survival, therefore, requires both the adaptive and innate immune response.
The innate immune system defends the host during the early phase of an infection and shapes the adaptive response [17]–[19]. The complement system is an essential component of the innate immune system, and evidence from human disease and animal models implicates complement as a critical part of host defense against several virus families [20]–[24].
The complement system consists of cell-surface and serum proteins that interact to destroy invading microorganisms and infected host cells [19], [25]–[27]. Three distinct pathways activate this cascade: classical, lectin, and alternative (pathway diagram in the results section). Antibody binding to antigen triggers the classical pathway. Mannan-binding lectin (MBL) and related proteins recognize repetitive carbohydrate motifs on pathogens and infected cells to initiate the lectin pathway [28]. Spontaneously activated C3 initiates the alternative pathway, especially if deposited on surfaces deficient in regulatory proteins [29]. The alternative pathway also serves as a positive feedback loop by forming additional C3 convertases from the C3b produced by any pathway. All three pathways converge at the step of C3 cleavage to C3a and C3b, and they share a common terminal pathway that generates the C5a anaphylatoxin and the membrane attack complex (MAC).
Complement system activation can exert multiple antiviral effects [25],[27]. Opsonization of the virion may block attachment or promote destruction by phagocytosis. The MAC disrupts the membrane integrity of the virion or infected cells. The anaphylatoxin cleavage products, C3a and C5a, attract and activate proinflammatory and immune effector cells [30]. Finally, complement activation induces and instructs the adaptive response and augments the neutralizing activity of antibody [18], [31]–[33]. To evade these antiviral activities, viruses use multiple strategies to hinder complement activation [26],[27],[34].
In their large double-stranded DNA genomes, poxviruses encode factors that modify the immune response [35]. Study of immunomodulatory molecules has provided insights into viral pathogenesis and revealed novel facets of the host's immune response [36]–[39]. Variola virus, monkeypox virus, and ECTV each produce an orthologous complement regulatory protein that has structural and functional homology to host proteins [33], [40]–[45]. Loss of this complement regulatory protein may account for the reduced virulence seen in the West African vs. Congo basin strains of monkeypox virus [45],[46]. The limits of the monkeypox animal models, however, have made this a difficult hypothesis to test. Loss of the complement regulatory protein affects local lesion size of cowpox and vaccinia virus, but these are non-lethal infection models [33],[47]. Additionally, an incomplete understanding of the role of complement during poxviral infections has complicated the investigation into how these proteins enhance virulence.
Complement influences poxviral infections, but an essential role for survival has not been demonstrated. One study described increased inflammation at the inoculation site of cowpox virus in C5−/− mice; however, no mortality occurred in these mice [48]. Additionally, an allele for genetic resistance to ECTV mapped to the chromosomal region containing C5 [49].
Using complement-deficient mice, the mousepox model offers an opportunity to characterize the role of this system during infection in the natural host. Use of a model where the host and pathogen have co-evolved is particularly important given the species specificity of many poxviruses and of complement proteins, regulators, and receptors [3],[50]. In this study, we focused on the role of C3, the complement cascade's central component. Resistant C57BL/6 mice that genetically lack C3 inadequately control ECTV infection and have increased morbidity, viral burdens, and mortality. Our in vitro and in vivo evidence suggests that the complement system neutralizes ECTV early in infection and contributes to survival.
The route of infection influences the interaction between poxviruses and the host [51]. Half of the 16 mutant vaccinia viruses assessed using two routes of inoculation, ear pinna or intranasal, had a detectible phenotype by only one route. ECTV infections of C57Bl mice by the intranasal, intraperitoneal, or intravenous routes result in severe disease and mortality, while the footpad and intradermal routes cause minimal disease [52]. To examine the role of complement in vivo, wild-type and C3−/− mice were infected by three routes: footpad, ear pinna, and intranasal.
Approximately 95% (54 of 57) of the wild-type mice survived when inoculated with 40,000 pfu of ECTV, the highest dose employed in the footpad infections (Figure 1A). In contrast, C3−/− mice had about 90% mortality at that dose. They also had significantly increased mortality (P<0.0001) at lower doses, even when inoculated with only 4 pfu. The median time to death increased as the dose decreased from 7 days at 40,000 pfu to 9, 10, or 13 days at the lower doses of 4,000, 400, or 4 pfu, respectively.
The C3−/− mice also showed increased morbidity over the course of the infection. Unlike the wild-type mice on day 7 post-infection with 40,000 pfu, the C3−/− mice displayed clinical signs of infection, including fur ruffling and hunchbacked posture. Consistent with these observations, C3−/− mice lost more weight at 400 and 40,000 pfu than wild-type (Figure 1B). The few surviving C3−/− mice at the 400 pfu dose required ∼3 additional weeks compared to the wild-type mice to return to their initial weight. All surviving mice in Figure 1A were held for at least 40 days to monitor recovery, and a subset of C3−/− mice (n = 4 at 400 pfu) were held to day 119 post-infection. The mice that survived the acute illness recovered weight steadily and showed no signs of relapse.
The ear pinna studies used a dose of 700 pfu to mimic the low inoculum thought to transmit the natural poxvirus infection [2]. The infection caused 72% mortality in the C3−/− mice (28 of 39) (Figure 1C), compared to 25% in the wild-type mice (6 of 24, P = 0.0002). The surviving C3−/− mice lost less weight and recovered to the initial weight earlier if inoculated by the ear pinna compared to the footpad route (Figure 1B and 1D). In contrast to the increased morbidity and mortality observed, C3 deficiency caused no gross differences in the primary lesion; C3−/− and wild-type mice had similar levels of footpad swelling or necrosis at the ear pinna inoculation site (data not shown).
To examine the role of C3 in intranasal infection, the dose was lowered to 100 pfu due to the increased susceptibility of the wild-type mice with this route. C3 deficiency increased the mortality rate from 40% to 80% (P<0.0001, Figure 1E). Similar to the other routes, the surviving C3−/− mice had more severe disease than wild-type, as they lost more weight and took longer to recover (Figure 1F).
ECTV replicates at the inoculation site and in the draining lymph node to generate the primary viremia that infects the spleen and liver [4]. Virus released from these target organs causes a secondary viremia, which seeds distal sites like the skin, generating the characteristic pox lesions.
To begin to dissect how C3 contributed to protection against ECTV, we examined viral burden in two key tissues, the spleen and liver. Wild-type and C3−/− mice were inoculated in the footpad with either 400 or 40,000 pfu, and then spleen and liver tissue were collected on day 7 post-infection. All animals had detectible virus in either the spleen or liver. At the two doses, the C3−/− mice had a 1–2 log higher mean titer than wild-type mice in both tissues (Figure 2A and 2B).
In wild-type mice, both doses produced similar maximal tissue titer; however, the higher dose increased the uniformity of the group and, thereby, increased the mean titer. At the 40,000 pfu dose, the splenic viral burden in the C3−/− mice was ∼150-fold higher (P = 0.0002, Figure 2A). Reducing the dose to 400 pfu resulted in ∼25-fold lower viral titer in the C3−/− mice, yet it was still ∼25-fold higher than the wild-type controls (P = 0.03). In contrast, both doses produced similar liver titers in the C3−/− mice. The lower dose revealed an 80-fold increase in the liver titer of the C3−/− mice compared to the wild-type mice (P = 0.01, Figure 2B), while the higher dose showed less of a difference between the strains (15-fold) due to the increased titer in the wild-type mice. Illustrative of the impact of C3, the C3−/− mice at 400 pfu had higher titers than the wild-type mice given 40,000 pfu, a 100-fold more virus.
These increases in viral titer prompted further exploration of how C3 deficiency impacts viral spread. C3 could control viral replication early at the inoculation site by directly inactivating free virus or by recruiting inflammatory cells through release of anaphylatoxins. The lack of C3 in the blood to neutralize or opsonize the virus could also result in greater viremia, thereby producing the higher titers observed in the target organs on day 7. Alternatively, C3's well-established ability to facilitate induction of antibody and T cell responses could explain the observed difference [21], [22], [53]–[57]. To elucidate when the infections in the C3−/− and wild-type mice diverge, we inoculated via the ear pinna route and examined the viral burden in the blood, spleen, and liver on days 2, 4, 7, and 10 post-infection. The ear pinna route was selected for further analysis because it is a cutaneous route of inoculation that mimics a natural infection of the epithelium where complement may promote containment.
Using whole blood enables an unbiased detection of all virus, whether free in the plasma or in infected cells. Quantitative PCR was employed to detect viral DNA in blood on days 2, 4, and 7 (Figure 2E). A few day 2 samples contained viral DNA, but most were below the detection limit. The C3−/− mice had 2.0- and 2.5-fold higher levels of viral DNA than wild-type mice had on days 4 and 7 (P = 0.004 and 0.03, respectively).
Despite the low levels of viremia on day 2, infectious virus was present in the spleen of over 70% of the C3−/− mice (13 of 18) compared to 28% of the wild-type mice (5 of 18, P = 0.006, Figure 2C). By day 7 post-infection, the C3−/− mice had 45-fold higher viral titers in the liver (P = 0.01, Figure 2D), and there was also a similar trend in the spleen (6-fold, P = 0.09). The wild-type mice regained weight starting on day 10 (Figure 1D), and by then over 80% had cleared the virus from the spleen or liver (9 of 11, Figure 2C and 2D). In contrast, less than half of the C3−/− mice survived to day 10 (Figure 1C), and of these, over 75% had ongoing infection of the spleen and liver (7 of 9, P = 0.008 and 0.01, respectively).
In summary, C3 deficiency resulted in earlier dissemination to spleen and in higher peak titers in the liver. The viral infection also continued to day 10 in the C3−/− when it had been cleared by most wild-type mice.
In susceptible mouse strains, ECTV causes extensive hepatic and splenic necrosis [58],[59]. We compared C3−/− and wild-type mice for histopathological changes in the liver on days 4, 7, and 10 post-infection.
On day 4, the liver histopathology appeared normal in 4 of 5 wild-type and 3 of 4 C3−/− mice (data not shown). By day 7, all animals had a diffuse lymphocytic infiltrate in addition to discrete inflammatory foci (Figure 3). These lesions varied in size and were smaller and less frequent in the wild-type (Figure 3A and 3B) compared to the C3−/− mice (Figure 3C and 3D). They often occurred near the portal triad, and some contained areas of coagulative necrosis. An inflammatory infiltrate encircled the discrete necrotic foci (Figure 3B and 3C) and bordered the areas of bridging necrosis (Figure 3D). In contrast to the liver, no major differences were observed in the spleen at this time (data not shown). Using blinded samples, we counted the necrotic and non-necrotic foci and evaluated the location and severity of the necrosis in the liver (Figure 4).
There were prominent differences between the C3−/− and wild-type mice relative to the number inflammatory foci and in the degree of necrosis. The C3−/− mice had twice as many total foci (8 vs. 18 per field, P = 0.02, Figure 4A) and 5-fold more foci containing regions of necrosis (3 vs. 15 per field, P = 0.03, Figure 4B). The majority of inflammatory foci contained necrotic areas in two-thirds of the C3−/− mice compared to only one-fifth of the wild-type mice (Figure 4C). The C3−/− mice had larger foci with more extensive necrosis (P = 0.02, Figure 4D). Most wild-type mice had small foci with either no necrosis or only piecemeal necrosis (0 and 1 on necrosis severity scale, Figure 3A and 3B, respectively). In contrast, the C3−/− mice had confluent areas of necrosis that coalesced into bands of bridging necrosis (2 and 4 on the necrosis severity scale, Figure 3C and 3D, respectively). Given that necrosis most frequently occurred in zone 1 of the liver, it likely originated there and then extended into zones 2 and 3 (Figure 4E).
The increased hepatic necrosis in the C3−/− mice resulted in higher levels of liver enzymes, aspartate aminotransferase (AST) and alanine aminotransferase (ALT), in the serum on day 7 (P = 0.008, 0.0503, respectively, Figure 4F). The AST and ALT levels positively correlated with the viral burden (Figure 4G).
Most C3−/− mice died between day 7 and 10 (Figure 1C). Two C3−/− mice that were sacrificed on day 10 had ∼5–7 inflammatory foci per field, while the 5 wild-type mice had only occasional foci (data not shown). At this time point, infectious ECTV persisted in the C3−/− mice; whereas, wild-type mice had cleared the infection (Figure 2D).
To explore the interaction between C3 and ECTV in vivo, we examined how mouse complement affects ECTV virions in vitro. Purified intracellular mature ECTV was incubated with either EDTA-treated plasma or sera from naïve C57BL/6 mice. Infectious virus was detected as plaques on a BS-C-1 monolayer. EDTA-treated plasma was reconstituted with a buffer containing calcium and magnesium to allow for complement activation.
Reconstituted wild-type plasma neutralized approximately 90% of the virus (Figure 5A, P<0.001). Heat inactivation or buffer lacking calcium and magnesium abolished neutralization. Wild-type sera concentrations of 10, 25, or 50% neutralized 70–80% of the ECTV (Figure 5B, P<0.0001). These observations implicate the complement system in neutralizing ECTV.
To further define if complement neutralized ECTV, sera from mice genetically deficient in a complement component or antibody were used in this assay (Figure 4C–4G). The neutralizing activity was reduced by ∼50% with deficiency of either C3 or C4 (Figure 5C). However, mixing C3−/− and C4−/− sera produced results equivalent to wild-type sera.
This requirement for C4 for full ECTV neutralization was further dissected. The C1q subunit of C1 interacts with antibody to trigger the classical pathway. MBL, a C1q analog, initiates the lectin pathway. MBL A−/− x MBL C−/−, C1q−/−, and antibody-deficient (μMT) sera were compared (Figure 5D). μMT or C1q−/− sera only partially neutralized ECTV, comparable to C4−/− sera. Conversely, wild-type levels of neutralization occurred independent of MBL A and C. These data suggest that natural antibody activated the classical complement pathway to neutralize ECTV.
Further analysis revealed three key points relative to natural antibody. First, heat-inactivated wild-type sera behaves like buffer alone, which indicates that natural antibody alone lacks neutralizing activity; instead, complement activity was required to neutralize ECTV (Figure 5A and 5G). Second, heat-inactivated wild-type sera, as a source of natural antibody, restored the neutralizing activity of μMT sera (Figure 5F). Consistent with this finding, μMT or heat-inactivated wild-type serum did not effectively neutralize ECTV independently, but they did so in combination. Third, the modest but significantly greater neutralization in the normal compared to heat-inactivated μMT sera suggests that antibody-independent (alternate pathway) complement activation also occurred.
C3b deposited by any pathway interacts with factor B (FB) and factor D (FD) to generate the alternative pathway C3 convertase, which amplifies C3b production. Alternative pathway activation itself likely explains the neutralization observed in the μMT, C1q−/−, or C4−/− sera. Interestingly, FB−/− or FD−/− sera neutralized less ECTV than wild-type sera (Figure 5E), which indicates that the alternative pathway enhanced complement-mediated neutralization initiated by the classical pathway.
C3b could neutralize ECTV by directly preventing attachment to or entry into the cell or by disrupting the virion's membrane through formation of the C5 convertase and the MAC. C5 initiates the terminal pathway that forms the MAC, and no lytic activity occurs in the absence of C5. C5−/− sera from C57BL/10 mice were used to define the contribution of the MAC to neutralization (Figure 5G). C5−/− sera neutralized a significant portion of virus (P<0.001), however, less than C5+/+ sera (P<0.05). These findings suggest that opsonization by C4b and C3b mediated most of the neutralization; although, the MAC also contributed.
To conclude, these findings demonstrate that naïve wild-type mouse sera neutralized ECTV. We propose that natural antibodies bound to ECTV and triggered the classical pathway. This led to C4b deposition, formation of the C3 convertase, and C3b deposition on the virus. The alternative pathway amplified the C3b placed on the virion by the classical pathway. Most ECTV neutralization occurred through opsonization by C4b and C3b, with a minor contribution from the MAC.
Both the classical and alternative pathways contributed to ECTV neutralization in vitro. To examine the importance of each pathway in vivo, we compared C4−/− and FB−/− mice to C3−/− and wild-type mice. We challenged C4−/− mice via the ear pinna route and monitored survival and weight loss. Over 90% of the C4−/− mice succumbed to the infection (P<0.0001, Figure 6A). The C4−/− and C3−/− mice had comparable mortality and weight loss (Figures 6C and 1D).
Intranasal ECTV infection also produced similar results in the C3−/−, C4−/−, or FB−/− mice. Each complement-deficient strain had a higher mortality rate compared the wild-type mice (P<0.0001), and there were no significant differences among the three strains (Figure 6B). The complement-deficient strains also lost weight at a similar rate (Figures 6D and 1F). Thus, control of ECTV in vivo required both the alternative and classical pathways, analogous to the in vitro results.
Complement poses a barrier to the systemic spread of pathogens, particularly through the bloodstream [17]. The major role of complement could be to neutralize ECTV recognized by natural antibody. Our prior experiments established that B cell-deficient μMT mice challenged with a high dose of ECTV by the footpad route all died early in infection (94% by day 8) (Figure 6E). Their early death suggests that B cells contribute to survival prior to the rise of specific antibody on day 7 [5].
Based on our in vitro data and the data of others [60],[61], we hypothesized that natural antibody contributes to early protection. Consequently, providing μMT mice with natural antibody should prolong their survival. Based on the work of Ochsenbein et al. [61], μMT mice infected with a high dose of ECTV were treated with naïve sera from either μMT or wild-type mice (Figure 6F). Treatment with wild-type sera increased the median day of death from 7 to 9; however, sera lacking natural antibodies (μMT) had no effect. On day 8 post-infection, over half of the mice receiving wild-type sera outlived both other groups and 16 of 17 mice from the prior experiments (Figure 6G). Thus, natural antibody delayed, but did not prevent, lethal ECTV infection in μMT mice.
We investigated the impact of complement deficiency using the ECTV mouse model. Deficiency of C3, C4, or FB resulted in acute lethal infection, establishing a requirement for multiple complement pathways in host defense against this pathogen. Specifically, C3 deficiency permitted ECTV to disseminate earlier, reach a higher titer in the target organs, and induce greater liver damage. Consistent with these in vivo results, naïve mouse sera neutralized ECTV infectivity in vitro, and sera lacking either classical or alternative pathway components had decreased activity. Several lines of evidence indicate that natural antibody initiated the classical complement cascade in the wild-type mouse. Substantial neutralization occurred in sera without lytic activity, which points to opsonization as the predominant mechanism of neutralization. Based on these results, we propose that natural antibody binds viral antigen to activate the classical pathway, followed by engagement of the alternative pathway's feedback loop to opsonize the virus.
The ECTV model system provides several advantages for analyzing the role of complement in poxviral pathogenesis. First, the mouse-specific pathogen ECTV has coevolved with and causes severe disease in the natural host, analogous to variola virus in humans. Second, the role of complement and the pathways involved can now be more rigorously dissected in vivo and in vitro with the availability of complement-deficient mice. Additionally, the in vitro experiments employed sera from the same strains used to characterize the effect of complement deficiency in vivo, and the neutralizing capacity in vitro paralleled the in vivo mortality. Third, viral pathogenesis, morbidity, and mortality can be assessed by multiple routes of infection and across a range of viral inoculum to demonstrate a broad requirement for complement.
Complement-deficient mice succumbed to acute ECTV infection with the majority of deaths occurring between days 6–10. Based on time to death following footpad inoculation, C3 deficiency resembled immunodeficiencies of other important components of the antiviral response, specifically CD8+ T cells [6],[7], NK cells [9],[10], and IFN-γ [12]. In contrast, mice deficient in CD4+ T cells [6], CD40, or CD40 ligand (CD154) [7] survive the acute phase but do not clear the virus. The CD40−/− and CD154−/− mice ultimately die ∼4 to 8 weeks post-infection. This differs from surviving C3−/− mice, which recovered and did not show signs of ongoing illness for up to 4 months of observation. The early death of the complement-deficient mice highlights the complement system's essential contribution to survival during the first few days of infection.
To characterize how complement protects the host from lethal infection, we analyzed the impact of C3 deficiency on the kinetics of viral spread. ECTV replication at the inoculation site and in the draining lymph node produces a viremia that seeds the primary target organs, the liver and spleen [4]. Several observations from this study increase our understanding of complement's role in controlling poxviral infection.
First, as early as day 2, C3 deficiency allowed for greater spread of ECTV from the inoculation site to the spleen. This indicates that complement is a key player in the initial hours of infection, likely to control ECTV at the inoculation site. Second, we detected higher levels of viral DNA in the blood on days 4 and 7. Consistent with our in vitro data, these results establish that C3−/− mice poorly control viral dissemination through the bloodstream. This higher viremia could result from increased replication in tissues and/or decreased clearance of virus from the bloodstream. Third, the liver viral titers on day 7 were ∼50-fold higher in the C3−/− mice. The greater viremia likely produced more extensive infection, but a delayed adaptive immune response may also have contributed to this observation. The viral titer correlated with serum levels of ALT, which suggests that ECTV caused hepatic necrosis either directly through lytic infection or indirectly through the antiviral immune response. An inflammatory infiltrate surrounded the necrosis in the C3−/− mice, which contrasts with susceptible Balb/c mice where necrosis occurs in the absence of a lymphocytic infiltration [62]. In summary, we propose that mice lacking C3 have reduced ability to control ECTV locally and in the bloodstream, leading to higher levels of infection and greater tissue damage in the liver.
Complement could delay viral dissemination by opsonizing and thereby neutralizing virions at the inoculation site or in the circulation and by promoting the inflammatory response including the recruitment of phagocytic cells. To assess if complement could directly neutralize ECTV, we examined the interaction between purified ECTV and mouse complement in vitro. Naïve plasma or sera neutralized ECTV in a complement-dependent manner, even at a concentration as low as 10%. Sera from mice deficient in specific complement components demonstrated that maximal neutralization required both the classical and alternative pathways. μMT sera, lacking antibody, resembled the sera deficient in the classical pathway components, C1q or C4, and addition of a natural antibody source restored neutralization activity. Opsonization led to neutralization of the majority of virus; however, the modest but significant difference between the C5−/− and C5+/+ sera indicates that the MAC contributed to viral damage. Interestingly, no complement component deficiency tested fully abolished neutralization. The residual activity suggests that the classical and alternative pathways functioned independently, likely because both C4b and C3b opsonized and, consequently, neutralized ECTV. However, the system was most effective when the two pathways and the MAC worked cooperatively.
The reconstitution of the neutralization activity in the μMT sera with heat-inactivated wild-type sera suggests that natural antibody is important in the neutralization process. Consistent with this observation and prior studies with other viruses [60],[61], natural antibody passively transferred into μMT mice lengthened survival during the acute infection. In our experiments, most μMT mice died early, with 100% mortality by day 9 at the highest inoculum. The mice that survived the acute infection eventually died at ∼2 months post-infection. Our findings differ from prior studies, which described mortality at either 2–4 weeks [8] or 2 months [7] post-infection. More of our mice survived the acute infection at the lower doses. This suggests that the observed discrepancy could be secondary to differences in the viral stock or dose, as both differed among the three groups. The death of the μMT mice, despite natural antibody treatment, indicates that B cells help control the infection by additional mechanisms.
Our in vitro experiments provide a model for understanding the fate of the viral inoculum in our in vivo experiments, since they both used the same stock of purified intracellular mature virus (IMV). To understand the spread of infection, a second infectious form must be considered. During viral replication in the host cell, extracellular enveloped virus (EEV) is produced by enveloping the IMV with an additional unique membrane derived from the Golgi complex and late endosomal compartment [63]. In studies of vaccinia virus, the host's complement regulators, present in the outermost membrane, protect the EEV from human and rabbit complement; in contrast, the IMV is sensitive to complement [64]. Our study builds on this observation by determining the contribution of each complement activation pathway to the neutralization of IMV infectivity, and it implicates natural antibody as the primary initiating factor [61]. We also show that natural antibody by itself is ineffective but requires augmentation by the complement system to neutralize ECTV. Additionally, the neutralization observed with vaccinia virus and ECTV points to the IMV form being inherently susceptible to complement-mediated neutralization.
The relative importance of IMV vs. EEV during infection in vivo is not well established. However, the IMV's sensitivity to complement neutralization suggests that ECTV likely travels through areas featuring efficient complement activation, such as the blood stream, in the EEV form or within infected cells. At extravascular sites, where complement levels are lower than in circulation, infected cells may produce sufficient soluble poxviral complement regulatory protein to protect the IMV.
Most poxvirus disease models initiate infection with the complement-sensitive IMV. If complement activity in the mouse behaves as it does in vitro, then inoculated ectromelia IMV should be recognized by natural antibody and coated with C4b and C3b, resulting in neutralization of viral infectivity at the site of injection and inhibition of spread. This line of reasoning could explain why the mortality increases in the wild-type mice as the invasiveness of the route decreases [52]. Percutaneous inoculation would likely result in neutralization, while application to the mucosal membranes might enable ECTV to enter host cells before being neutralized by complement. Once internalized, ECTV produces its regulatory protein and EEV to evade complement and propagate the infection. Additionally, based on the in vitro data, complement deficiency would greatly limit this initial neutralization, which likely contributes to the early spread and greater mortality observed in the complement-deficient mice.
A sub-neutralizing concentration of complement opsonins could target the virion for immune adherence and phagocytosis in vivo, particularly in blood with its high complement levels. Furthermore, the liver sinusoids are lined with Kupffer cells bearing CRIg (Complement Receptor of the Ig superfamily), which mediates phagocytosis of C3-opsonized pathogens [65]. Indeed, the liver clears over 95% of intravenously administered ECTV from the circulation within 5 min of injection [66]. In the following hour, most of the viral antigen in the liver becomes undetectable by immunofluorescence, and viral infectivity decreases by over 90%. This rapid removal suggests that the virus has been recognized as foreign and tagged for immune adherence and destruction. Opsonization by complement followed by uptake via the recently described CRIg provides a mechanistic explanation for these important observations made nearly five decades ago [66].
These observations influence the interpretation of poxviral infections initiated with an IMV-rich inoculum by the intravenous route. The liver's Kupffer cells may sequester most of the inoculated virus within minutes and destroy much of it within an hour, thereby inhibiting systemic dissemination. Not only is the dose effectively reduced by ∼10-fold, but the neutralized IMV also provides the immune system with an immediate source of antigen. These issues have particular relevance for the monkeypox and variola virus non-human primate models that commonly use the intravenous route to test vaccines for human use [67]–[72].
The early mortality of the C3−/−, C4−/−, and FB−/− mice demonstrates an essential role for the classical and alternative pathways in the initial stages of poxvirus infection. Despite equivalent mortality levels, further analysis may reveal different functions for each complement pathway in vivo, as such differences exist in the immune response to other viruses [23]. The similarity between the in vivo mortality and in vitro serum neutralization experiments suggests that complement neutralizes ECTV and thereby limits its spread. Undoubtedly, complement deposition triggers other effector functions, such as recruiting inflammatory cells, promoting phagocytosis, and priming the adaptive immune response. The precise contribution of each of these to protection in vivo remains unexplored.
However, these experiments establish that complement is essential to the immune response to poxviruses. It accounts for why the virus encodes a potent complement regulatory protein. A virus lacking this regulator would be at risk for greater host complement activity and attenuation. This is consistent with the theory that loss of this regulator contributes to the reduced virulence of some strains of monkeypox virus [46]. To conclude, the complement system is critical for slowing down viral spread and decreasing tissue titers and damage.
Plaque-purified Moscow strain ECTV was propagated in murine L929 cells. Intracellular mature viral stocks were purified through a sucrose cushion as described [73] and titrated on BS-C-1 cells, an African green monkey kidney cell line [74]. Both cell lines were cultured in Dulbecco's modified Eagle's media (DMEM, BioWhittaker, Walkersville, MD) supplemented with 10% heat-inactivated fetal calf serum (FCS, HyClone, Logan, UT), 2 mM L-glutamine, and antibiotics.
The following strains on a C57BL/6 background were acquired: C3−/− [56],[75] and FB−/− [76],[77] from H. Molina, Washington University Medical School; C4−/− [78] from M. Carroll, Harvard Medical School; B cell-deficient μMT [79] from H. W. Virgin, Washington University Medical School; C1q−/− [80] from M. Botto, Imperial College School of Medicine; FD−/− [81] from Y. Xu, University of Alabama, Birmingham; and MBL A−/− x MBL C−/− (B6.129S4-Mbl1tm1Kata Mbl2tm1Kata/J) and wild-type from Jackson Laboratories. The C5+/+ and C5−/− C57BL/10 mice (B10.D2-Hc1 H2d H2-T18c/nSnJ, B10.D2-Hc0 H2d H2-T18c/oSnJ) were also obtained from Jackson Laboratories. Age-matched mice of both sexes were used in the footpad and ear pinna studies (6–11 weeks-old) and the μMT survival experiments (10–11 weeks-old). Male mice were used in the intranasal (8–12 weeks) and sera transfer (10–12 weeks) studies. Some wild-type and μMT mice used in the footpad studies were purchased from Jackson Laboratories. The rest of the mice were bred at Washington University in a specific pathogen-free facility. The animals were transported to the biohazard suite at Saint Louis University at least a week prior to infection. All experiments were performed following the animal care guidelines of the two institutions.
Mice were inoculated with 10 µl ECTV diluted in PBS to the indicated dose using a 29 gauge insulin syringe into the ear pinna and hind footpad or a 20 µl pipettor for the intranasal route. Mice were anesthetized for inoculation using CO2/O2 for the footpad route and ketamine/xylazine for the ear pinna and intranasal routes. Individual mice were marked by ear punching or shaving. After infection and before sacrifice in the mortality studies, mice were manipulated only to obtain weights. Serum was collected from surviving animals at the end of the experiment. The survival curves include only animals that generated an antiviral antibody response, which was detected by ELISA in >95% of the mice [82]. In the passive transfer experiment, mice received intraperitoneally 1 ml of wild-type or μMT C57BL/6 sera on day −1 and 0.5 ml every two days starting on day 0.
Blood was collected via cardiac puncture. Spleen and liver tissues were harvested aseptically, frozen immediately on dry ice, and stored at −70°C. Tissues were homogenized in PBS-1% FCS to ∼10% (weight/vol) using 1 ml glass homogenizers. They were frozen and thawed three times, sonicated, and titrated on BS-C-1 monolayers [74].
DNA was isolated from whole blood collected in EDTA microtainer tubes (BD, Franklin Lakes, NJ) using the High Pure PCR Template Preparation Kit (Roche). The kit's whole blood protocol was used with the following modifications. The 40 µl Proteinase K, 200 µl Binding Buffer, 150 µl PBS, and 50 µl of whole blood in EDTA were added sequentially and then vortexed. The incubation at 70°C was extended to 12 min. The sample was applied to the column by centrifugation at 8,000 g for 2 min and eluted in 50 µl.
Quantitative PCR was performed on viral DNA using Power SYBR Green PCR Master Mix on a 7500 Real Time PCR System (Applied Biosystems, Foster City, CA) [83]. The primers (10 pmol) SP028 (GTAGAACGACGCCAGAAT AAGAATA, 5′ at 120627 bp) and SP029 (AGAAGATATCAGACGATCCACAATC, 5′ at 120462 bp) were used to amplify 165 bp of gene EV107. The amplification product cloned into a plasmid vector (pGEM-T, Promega) was used as a standard to estimate copies of DNA/µl in blood. Three to four wells were used for each sample.
Tissue samples were fixed in 10% buffered formalin, embedded in paraffin, sectioned, and stained with hematoxylin and eosin by the Digestive Diseases Research Core Center, Washington University. The number of inflammatory foci and the magnitude of tissue necrosis were evaluated in blinded samples. Inflammatory foci in a 10× visual field were counted for ∼7 fields/mouse liver.
AST and ALT levels were measured in samples of frozen sera by the Department of Comparative Medicine at Saint Louis University using a standard clinical analyzer.
Mouse EDTA plasma and sera were collected on ice from male C57BL/6 mice in microtainer tubes (BD), separated by centrifugation, and then pooled, aliquoted, and frozen at −70°C. Plasma and sera were diluted on ice into GVB± Ca++/Mg++ (#B102, B103, Complement Technology, Tyler, TX) or GVB without Ca++/Mg++, respectively, to 2× the desired final concentration (vol/vol). Purified ECTV was sonicated and diluted in PBS (without Ca++/Mg++) to ∼5×104 pfu/ml. A 1∶10 dilution in the buffer used to dilute the complement source, GVB±Ca++/Mg++, produced a final concentration of ∼5 pfu/µl. An equal volume of virus (30 µl≈150 pfu) was added rapidly to the diluted complement at RT. Samples were vortexed, centrifuged for 5 sec, and incubated at 37°C for 60–90 min. Samples were diluted by addition of 700 µl of DMEM-2% FCS, vortexed, and applied to BS-C-1 monolayers in 6-well plates. After 1 hr, 3 ml/well of 37°C overlay media (1% carboxymethylcellulose in culture media) was added. After 3–5 days, the cells were fixed with 1 ml/well of an 11% formaldehyde/ 0.13% crystal violet/ 5% ethanol solution for over 1 hr, rinsed, and dried. The number of plaques was scored visually using a light box. The EDTA plasma or sera data were normalized to the buffer only control or heat-inactivated sera, respectively.
All statistical analysis was performed using GraphPad Prism software version 5.01 (GraphPad Software, San Diego, CA). The survival curves were analyzed by the log-rank test. The Mann-Whitney test was used to determine the statistical significance of the viral titers, viremia, liver histology, and liver enzymes. Either 1-way ANOVA followed by Tukey multiple comparisons test or 2-way ANOVA was used for the analysis of the complement neutralization assays.
|
10.1371/journal.pntd.0006293 | A rabies lesson improves rabies knowledge amongst primary school children in Zomba, Malawi | Rabies is an important neglected disease, which kills around 59,000 people a year. Over a third of these deaths are in children less than 15 years of age. Almost all human rabies deaths in Africa and Asia are due to bites from infected dogs. Despite the high efficacy of current rabies vaccines, awareness about rabies preventive healthcare is often low in endemic areas. It is therefore common for educational initiatives to be conducted in conjunction with other rabies control activities such as mass dog vaccination, however there are few examples where the efficacy of education activities has been assessed. Here, primary school children in Zomba, Malawi, were given a lesson on rabies biology and preventive healthcare. Subsequently, a mass dog vaccination programme was delivered in the same region. Knowledge and attitudes towards rabies were assessed by a questionnaire before the lesson, immediately after the lesson and 9 weeks later to assess the impact the lesson had on school children’s knowledge and attitudes. This assessment was also undertaken in children who were exposed to the mass dog vaccination programme but did not receive the lesson. Knowledge of rabies and how to be safe around dogs increased following the lesson (both p<0.001), and knowledge remained higher than baseline 9 weeks after the lesson (both p<0.001). Knowledge of rabies and how to be safe around dogs was greater amongst school children who had received the lesson compared to school children who had not received the lesson, but had been exposed to a rabies vaccination campaign in their community (both p<0.001) indicating that the lesson itself was critical in improving knowledge. In summary, we have shown that a short, focused classroom-based lesson on rabies can improve short and medium-term rabies knowledge and attitudes of Malawian schoolchildren.
| Rabies is a fatal disease that claims the lives of approximately 59,000 people every year. Children under the age of 15 make up 40% of all human rabies deaths yet this is preventable through a combination of vaccinating dogs against rabies and education. Numerous studies have shown that people in rabies endemic areas lack sufficient knowledge about rabies, and there are many misconceptions about its treatment and prevention. Whilst many organisations run vaccination and education campaigns, few have assessed their impact on rabies knowledge, attitudes or practices (KAP). Fewer still have assessed the impact on children. This study investigated the impact of a rabies lesson on school children’s knowledge and attitudes about rabies in conjunction with a rabies vaccination campaign in Zomba, Malawi. We found that a rabies lesson improved school children’s knowledge about rabies and how to be safe around dogs. We observed that knowledge remained higher several weeks later. Knowledge about both canine rabies and bite prevention was greater amongst school children who had received the lesson compared to school children who had not received the lesson, but had been exposed to a rabies vaccination campaign in their community. This indicates that the lesson itself was critical in improving knowledge.
| Of the estimated 59,000 people who die from rabies annually [1], the vast majority result from a bite from a rabid dog. Children are at greater risk of suffering dog bites than adults [2,3] and as a result approximately 40% of all human rabies deaths occur in children aged under 15 years old [4,5]. Elimination of the rabies virus can be achieved through annual vaccination of 70% of the dog population and human exposures to rabies virus will continue to occur until elimination has been achieved. Prompt post-exposure treatment is effective at preventing rabies, however incomplete adherence to recommended protocols has resulted in many deaths.
School based rabies education is an efficient way of reaching large numbers of children. Lessons containing simple messages can improve rabies prevention through appropriate behaviour, such as immediately washing bite wounds and seeking post-exposure vaccination.
Whilst many governments and NGOs advocate the integration of education components in rabies elimination programmes [6], few have published studies documenting the effectiveness of their interventions [7–9].
Knowledge, attitudes and practices (KAP) studies can be used to assess how effective education initiatives are by comparing responses prior to, and after, an intervention. Studies have shown that knowledge of rabies and rabies prevention can vary greatly across rabies endemic countries. For example, in one rabies KAP study in Tanzania only 5% of those interviewed knew of the importance to thoroughly wash dog bite wounds [10]; over 35% of respondents in Ethiopia did not know the symptoms of rabies in people [11]; and in Cambodia only 48% of people knew that vaccination could protect dogs from rabies [4]. A lack of understanding about the risk of rabies and preventive measures reduces the perceived need for control measures. It also reduces engagement with elimination efforts and the likelihood of taking appropriate action to prevent rabies in the event of exposure. Most rabies KAP studies have focused on adult populations [4,10–20] despite the disproportionally high incidence of rabies in children [8]. Only two studies have evaluated the efficacy of lessons to improve rabies KAP in children; Kanda et al. in Sri Lanka [9] and Dzikwi et al. in Nigeria [21]. Furthermore, studies assessing longer term knowledge retention after a short lesson and comparison with control populations exposed to mass dog vaccination campaigns alone are lacking. The current study was therefore conducted to investigate the immediate and medium-term impact of short lessons on primary school children’s understanding of rabies prevention in Zomba City, Malawi.
Zomba city, a rabies education naïve area in Southern Malawi, was chosen as the study site (Fig 1). Zomba City, located in Zomba District of southern Malawi is the fourth largest city in Malawi, with a human population of 114,000 (based on 3.2% annual growth rate since the 2008 census [22]). Mission Rabies is an international NGO working to establish effective rabies control activities in Malawi, including mass dog vaccination, community rabies education and enhanced canine rabies surveillance initiatives. Mass dog vaccination and education campaigns were conducted in Blantyre city in May 2015 and May 2016. No previous education or vaccination activities had taken place in Zomba city prior to this study. Local reports of high rates of canine and human rabies and an absence of rabies vaccination or education activities prompted requests from local authorities for expansion of Mission Rabies activities to the Zomba region. This study was undertaken during the initial stages of work in Zomba.
The 17 public primary schools in Zomba City had a total of 25,824 registered school children in 2016. All schools were included in the study. Fifteen received rabies education classes shortly before a mass dog vaccination campaign in the region. Two control schools did not receive education classes prior to the vaccination campaign. The control schools were closed during the period before vaccination so were not available to conduct the rabies education classes before the campaign. To investigate whether the choice of control introduced a bias, being a convenient sample, demographics of children in the intervention schools and the controls schools were compared.
Lessons were delivered in the national language, Chichewa, by trained Malawian Mission Rabies education officers to children in the seventh school year (standard 7). The interactive lessons lasted for approximately 45 minutes and included school children participation, requiring volunteers to join in short demonstrations and question-answer sessions. Learning points included which animals can transmit rabies, rabies symptoms and prevention, and safety around dogs. Education officers received four days of training in how to deliver a standardised lesson before commencing the study. The Mission Rabies school education campaign teaches all year groups, with lessons ranging from 20 minutes for younger years to 60 minutes for older year groups. A single year group was included in this study to standardise the type of lesson delivered.
The education programme was evaluated using self-administered paper questionnaires. Rabies lessons and initial questionnaires were conducted between 11th and 17th July 2016. In the 15 schools where rabies lessons were given, the same standardised questionnaire was completed by school children at three time-points; a “pre” questionnaire prior to receiving the lesson, a “post” questionnaire by the same children immediately following the lesson and a “retention” questionnaire, in the same class at the same schools 7.5 to 10.5 weeks later. The pre-questionnaire was used to assess children’s baseline knowledge and attitudes; the post-questionnaire assessed instant impact of the lesson on school children’s knowledge and attitudes; and the retention questionnaire assessed longer-term learning.
The control group completed the same questionnaire only once (“control” questionnaire), after the vaccination campaign that took place between 6th and 17th August 2016. This approach allowed us to assess the impact of the rabies lesson itself rather than just the exposure to the wider dog rabies vaccination programme. Retention and control questionnaires were completed between 8th and 29th September 2016.
The initial draft of the questionnaire was designed using input from NGO staff and publications conducting similar questionnaires directed at adults, followed by a process of informal feedback and refinement through application in two schools in Blantyre city. The questionnaire was written in English and translated to Chichewa. It was independently back translated to English for comparison with the original version to ensure question integrity was maintained.
The questionnaire consisted of four sections. Section A identified demographic information. Section B assessed dog ownership practices and understanding of animal welfare (to be reported outside of this study). Section C contained questions about rabies. Section D asked about prior rabies education and dog vaccination history. This section was used by the NGO and did not form part of this study. A copy of the questionnaire can be found in S1 Appendix.
To minimise age related bias the questionnaire was given to standard 7 school children only, as described above. Sample size was determined using a sample size calculator [24] with the following parameters: 95% confidence level; 5% margin of error; and a response distribution of 50%. These parameters were chosen to give the most conservative sample size. The target population contained 2,844 standard 7 school children registered with the education department in Zomba, therefore a sample size of 339 was required. This was the equivalent of 22.6 school children per school for the educated group and 169 per school for the control group. To account for incomplete questionnaires and varying school attendance this was increased to 30 per school for the educated group and 190 per school for the control group.
Even though each year group has an average of 169 registered students, which would lend itself well to systematic random sampling of every fifth student from the class register, attendance was much lower. For this reason, the teacher selected 30 children to take part, or as many as were present in the class where there were less than 30 school children. The teacher was instructed to select school children at random and not to choose by ability. All school children had the opportunity to decline taking part at each stage of the study.
Anonymity was maintained throughout the study by allocating each participant a unique identification code (UIC) consisting of a school code followed by a sequential number. The UIC was entered on the questionnaire, a consent letter for school children’s parents/guardians and on the data entry smartphone application. Less than half of the school children could be reliably matched between the pre and retention questionnaire due to school children absence at the time of the retention questionnaire or because school children forgot their UIC. Where there were fewer school children present for the retention than for the pre/post questionnaire, additional school children were invited to take the questionnaire with the provision that they had been present for the rabies lesson. The UIC was adapted to take this into consideration.
All questionnaires were presented to children on paper. Results from the questionnaires were entered into digital versions of the questionnaire created in a smartphone application (the Mission Rabies App) that allows remote data collection through smartphones [25]. Where possible, data were entered through single- and multi-select options to minimise transcription error [26]. Data were uploaded to a Microsoft SQL database on a secure cloud based server, from which it could be downloaded remotely via a password protected website as a csv file, which was imported into Excel 2013 (Microsoft Inc., Redmond, WA) and R Studio [27] for analysis.
To allow statistical analysis numerical scores were allocated to answers based on accuracy of response for questions which had correct or incorrect answers. A completely correct answer scored 2, a mostly correct answer 1, a missing or wrong answer 0 and an incorrect answer -1. For example, the latter includes answers that could have deleterious consequences for the child or an animal, or if an incorrect species susceptible to rabies was identified. Scores of 0 were given to answers that are not correct but would not have deleterious consequences. For example, the risk of getting rabies from milk is theoretical but people are not encouraged to drink milk from a rabid animal [28,29]. Therefore, responses that milk could transmit rabies were given a score of 0, as whilst it is wrong it is not harmful. Scores allocated to each question response can be found in S1 Table. When questions allowed for multiple answers the score was the sum of all answers selected. Questions were grouped into categories that assessed knowledge and attitudes towards rabies and safety around dogs (S2 Table).
Data were analysed using the R statistical software version 3.3.2 [30] with paired t-tests comparing matched pre and post questionnaire responses [31]. Results from 13 pre-questionnaires could not be matched to post questionnaires and these data was excluded when performing paired t-tests between these questionnaires. Two tailed two sample t-tests [31] were used to compare data that could not be matched: pre to retention, pre to control and control to retention scores. To deal with the issue of multiple testing, the threshold cut-off for significance was adjusted to a p-value < 0.003 based on the Bonferroni correction [31]. Mixed effects multiple linear regression [31] was used to determine the effect of demographics on baseline questionnaire scores. Children who had not replied to any of the questions considered in the model were removed from the regression analysis. Children’s age, gender, religion and dog ownership status were considered in the model as fixed effects. The school each student studied at was introduced in the model as a random effect. Variables selection was carried out using manual backward elimination and variables retained in the final regression model were chosen based on their effect on the Akaike information criterion (AIC).
The study and questionnaire were approved by the University of Edinburgh’s Human (Research) Ethical Review Committee (HERC). In Malawi, we obtained permission from the Department of Education. Signed consent was granted by head teachers of all schools prior to commencing the study. Consent letters were given to school children to give to their parents/guardians with clear instructions on how to remove their child’s data from the study if desired.
The number of school children who completed the questionnaire at each stage is shown in Fig 2. Only 122 out of 386 school children who completed the pre-questionnaire and could remember their UIC completed the retention questionnaire. School children who had been present for the lesson but did not complete the questionnaire and those who could not remember their UIC comprised the remainder of school children completing the retention questionnaire (Fig 2). S1 Fig shows the distribution of missing data in the questionnaire responses.
The mean age of school children was 13 years old with a range between eight and 15 years old for pre, post and control groups, and nine and 15 years old for retention groups. Though approximately equal, slightly more females than males completed the questionnaire.
Approximately 50% of the educated school children were Catholic, with other Christian denominations accounting for at least 30%. This differed in the control group where more school children belonged to other Christian denominations followed by Catholicism. Across all groups between nine and 15% of school children were Muslim. Across all questionnaires the majority of school children either owned or had contact with dogs (pre 78.8%, n = 304; post 79.3%, n = 302; retention 83.6%, n = 317; control 82.3%, n = 284).
The main reason for dog ownership was guarding (pre 86.1%, n = 242; post 86.3%, n = 276; retention 81.8%, n = 233; control 86.1%, n = 223). Most dogs were kept tied up or in a cage outside (combined results: pre 76.2%, n = 294; post 74.5%, n = 284; retention 68.9%, n = 261; control 58.6%, n = 202).
S2 Fig shows graphical comparisons of means/proportions and confidence intervals of demographic characteristics between intervention (educated school children) and control (school children not exposed to the rabies lesson).
Pre-questionnaire results were used to determine baseline knowledge and attitudes. Overall, this was assessed by pooling the scores for all questions assessing knowledge or attitudes. Accounting for school differences as a random effect, the mixed effects multiple linear regression model showed that male students had higher baseline scores. Similarly, when compared to Catholic students, students who said they did not belong to a religious group had higher scores. School children that did not own or have contact with dogs had lower overall baseline scores when compared to those who owned dogs. Results of the regression model are presented in Table 1. The variables selection process used is explained in Table 2.
Safety around dogs was assessed by asking school children how to behave around dogs to avoid being bitten. Of all responses, 86.6% identified an appropriate behaviour (stand still, be calm, cover face and play with friendly dogs) (S3 Fig).
School children achieved a mean score of 19.23 (sd 7.89) out of 71 for rabies knowledge. Seven out of the eight questions assessing rabies knowledge allowed multiple responses, yet between 43.0% and 72.5% of school children gave single answers to these questions. Most school children knew that people can get rabies (86.5%, n = 334). 91.0% (n = 628) of responses for the question ‘Which animals can get rabies?’ were correct with 49.0% (n = 338) of school children selecting dogs, whilst only 83.3% (n = 685) of responses were correct for ‘Which animals can give rabies to people?’. Being bitten was the most commonly selected option for how people can get rabies with 59.4% (n = 318) of responses. Many school children could identify symptoms of rabies (95.6%, n = 723 correct responses). Of the responses for ‘What do you do if bitten by a dog?’ 24.2% (n = 172) of responses were correct and 68.4% (n = 486) were completely correct. Over half of the responses (54.0% n = 249) given to the question asking children to identify ways to prevent dogs from getting rabies were that a dog should be vaccinated annually. School children also identified that vaccinating dogs and people would prevent people from getting rabies (32.2% n = 170 and 39.2%, n = 207 responses respectively). S3 Table details rabies knowledge responses across all questionnaires. Most school children believed that rabies was serious and that dogs should be vaccinated giving an overall rabies attitude score of 3.81, out of 4.
Overall score results, as well as scores for each category are presented in Fig 3. Furthermore, results of t-test comparisons are shown in Table 3. Results for overall score, safety around dogs and rabies knowledge all demonstrated a significant improvement in score immediately after the lesson. This reduced over time but remained significantly greater than baseline at the retention questionnaire. Control scores were significantly different to retention scores but not significantly different to pre-questionnaire scores.
Some question responses illustrated an improvement in knowledge more than others. For example, 32% of children identified that they should inform an adult if they were bitten by a dog after the lesson compared to 17% before the lesson. They also knew to clean the wound for 15 minutes (proportion of responses per questionnaire: pre 1.8% n = 13, post 9.9% n = 81, retention 8.8% n = 65) and apply antiseptic (proportion of responses per questionnaire: pre 4.5% n = 32, post 10.3% n = 84, retention 9.3% n = 69). This knowledge diminished over time but remained greater than baseline levels 9 weeks after the lesson. Conversely school children failed to identify that they should go for 5 vaccines and go to the hospital (S4 Fig).
Following the lesson there was an increase in the number of animals that school children identified could get rabies. The number of responses increased for all mammal species, though the greatest difference was in cats, bats and monkeys. This knowledge waned overtime though remained greater than baseline (for all mammal species other than dogs) and was greater than control groups (except for mongoose). Equally, there was a very similar increase in the number of animals school children identified that could transmit rabies to people. Again, the number of responses increased for all mammals with the greatest increase for cats, bats and monkeys. Children’s responses to the question assessing how rabies can be transmitted to people followed a similar pattern with an increase in correct responses that diminished over time but remained greater than baseline and the control group for all but one of the correct answers. There was an increase in response rate to saliva and scratches, and to a lesser extent licking wounds, after the lesson. However, this was coupled with a reduction in children’s responses indicating that bites can transmit rabies to people. There were no significant changes between school children groups or questionnaires for attitudes towards rabies (Table 3) as most school children had initially shown strong attitudes. Similarly, over 80% of responses to all questionnaire types indicated that children knew that people could get rabies.
This study examined the ability of a rabies school lesson to enhance both immediate and medium-term rabies knowledge and attitudes in primary school children. To the authors’ knowledge, this is the first study to evaluate medium-term rabies knowledge retention post intervention in children. Additionally, assessment of rabies understanding gained from a lesson in comparison to that generated by high-profile dog rabies vaccination campaigns alone has not been previously reported. We have shown that a short rabies lesson can significantly improve knowledge about rabies biology and preventive healthcare and that this information is largely retained two months later. Children receiving rabies lessons scored better than children who had only been exposed to the vaccination campaign.
For many people in Malawi and other sub-Saharan countries there remain many barriers to accessing appropriate post exposure prophylaxis following a dog bite. These include geographic and economic constraints for individual bite victims. This is exacerbated by frequent shortages of human rabies vaccine at major hospitals and rabies immunoglobulin, for example the latter is not available in Malawi. Until improvements are made in access to effective vaccines, prompt and thorough local wound treatment may be the only opportunity for bite victims to reduce their risk of rabies. Washing bite wounds for a period of at least 15 minutes with water and soap, detergent, povidine iodine or other virucidals can significantly reduce the risk of contracting rabies from an infectious animal [32,33]. This study showed that both before and after the lesson, more children answered that it was necessary to get vaccinated than to wash the wound, highlighting the need to raise common knowledge of this potentially lifesaving step. Hampson et al. also reported the lack of awareness of prompt bite wound washing in Tanzania [34]. In the current study, there was a fivefold increase in the proportion of children answering to wash bite wounds for 15 minutes, from 2% to 10% following the lesson. This accompanied a marked increase in the proportion of children reporting the importance of telling an adult from 17% to 32%. Although it is hoped that increasing the likelihood of children reporting bite wounds to adults improves their chances of receiving appropriate PEP, this may not be the case. Despite the statistically significant increases in correct responses, the proportions of students giving completely correct responses remained lower than ideal. The lesson should be reviewed aiming to achieve even greater improvement in children’s knowledge, especially in crucial issues such as the proportion of children knowing that dog bite wounds must be thoroughly washed. This might be achieved through greater emphasis on critical aspects of the lesson, repeating education programmes more often and combining them with other interventions. Community education activities targeting adults take place in conjunction with the school education campaign. Further investigation is warranted to assess whether there is a change in knowledge and practices among the adult population. Given the risk of other water borne diseases such as schistosomiasis, it may be beneficial to incorporate simple messages about the importance of using clean water.
There was no significant change after the lesson in the proportion of children reporting the need for five post-exposure vaccinations. This may be because they selected other behaviours such as telling an adult. Other factors such as: the cost of travel; distance to vaccine distribution points; physician recommendations; and availability of vaccine, are more likely to influence a child’s likelihood of completing a full course of rabies vaccinations regardless of their knowledge on how many doses are needed [34].
Demographic factors associated with baseline rabies knowledge and attitude scores were investigated using a mixed effects multiple linear regression model. Our results showed that male students had higher baseline scores. Additionally, school children that did not own or have contact with dogs had lower overall baseline scores when compared to those who owned dogs. During our experience in the field, we observed that dogs are more often brought to the vaccination clinics by boys rather than girls. Additionally, several authors have suggested that children that have experience with animals have greater knowledge about the welfare needs of animals [35,36]. It is likely that boys, which have more interaction with dogs and are more likely to take up responsibilities relevant to maintaining the health of dogs, might result in them knowing more about dog welfare, diseases and risks relating to interacting with them. Lastly, religion was included in the model as a potential explanatory variable as it was hypothesised that Muslim children, who do not traditionally keep dogs [37,38] and may therefore know less about dog welfare needs, would have lower baseline knowledge. Nevertheless, the model showed no such difference. On the other hand, students who said they did not belong to a religious group had higher knowledge scores compared to Catholic students.
A novel component of this study was the assessment of both the immediate and mid-term (after about 2 months) impact of a rabies lesson on knowledge and attitudes. Dzwiki et al. tested effects of a lesson and educational leaflets 2 weeks after the intervention [21], whilst Matibag et al. reassessed KAP 4 weeks after distributing leaflets and acknowledged the need to test longer term KAP [7]. Assessing mid to long term knowledge retention has important implications to determine if educational interventions warrant repetition and at what frequency. Whilst there was a slight drop in overall knowledge and attitude scores 9 weeks post lesson, the overall score remained significantly higher than at baseline. The lesson given by the NGO in this study is repeated annually and it would be interesting to assess rabies knowledge and attitude scores at the next visit to determine level of retention after one year and the effects of a repeat lesson.
A child’s risk of contracting rabies may be reduced if they are better able to recognise signs of aggression in dogs and know what they can do to avoid being bitten by aggressive dogs. This study found that children scored better on questions about how to stay safe around dogs after the lesson. The use of roleplay and theatre was a low cost and engaging way to convey aspects of dog behavior without the need for electricity or video equipment as has been described in other studies (19). Additional improvements to knowledge that may reduce a child’s risk of rabies included understanding which animals can transmit the disease and that rabies is transmitted in saliva.
Queen Elizabeth Central Hospital in Blantyre, Malawi, documented the highest number of child rabies deaths from any African hospital [39], making the findings of this study more significant. It is anticipated that improving child rabies knowledge and attitudes will result in reduced child rabies deaths. The finding that rabies knowledge was significantly greater in children who received the lesson compared to children who had only experienced a mass dog vaccination campaign, highlights the benefits of conducting school focused education activities on rabies. In the authors’ experience, children often play a positive role in mass dog vaccination campaigns by bringing dogs for vaccination and increasing vaccination coverage. Therefore, timing school education initiatives shortly before mass vaccination campaigns could have the dual benefit of increasing turnout to vaccination camps and increasing their chances of seeking appropriate PEP in the event of exposure.
This study had some limitations. Due to low attendance of children in each classroom, systematic random sampling was impossible to employ and teachers were asked to choose students to take part in the questionnaire. To overcome the possible bias that can arise from this it was emphasised that the selection should not be based on student’s ability. Additionally, it was not possible to ensure that the school children who completed the pre/post questionnaire also completed the retention questionnaire. This prevented direct comparisons between individuals, however it was still possible to compare results as a population. Despite this the results emphasised the efficacy of a rabies lesson among school children in urban Malawi. Further work is needed to investigate whether the findings of this study can be repeated in rural populations. Furthermore, controls for this study were chosen conveniently. To investigate potential bias introduced by this, demographics of intervention and control groups were compared. The only difference found regarding the demographics between the groups, was the proportion of Catholic students compared to Christians from other denominations. Based on our regression analysis, this is not expected to have an effect on baseline rabies knowledge, so it is unlikely to introduce bias to our work. One important result of this study is that we have been able to show significant knowledge retention in our intervention group. Despite that, the lack of a control group where students would be exposed to the educational campaign, but not the vaccination campaign makes it difficult to distinguish whether retention was due to the rabies lesson alone or was enhanced by exposure to the vaccination campaign. Finally, the change in risk of rabies or exposure to dog bites in children following the lesson could not be evaluated in the current study. Therefore, future investigation into changes in risk reduction behavior following the lesson is warranted.
Although the described education activities took place under the consent and support of the Department of Education and it is delivered by Malawian Education officers, there remain challenges in scaling this initiative to larger, particularly rural areas in a cost-effective way. The use of external education officers who systematically tour schools in a region, delivering specific health care messages has benefits in bringing up-to-date information and teaching methods. However, the duration of each class and benefit delivered needs careful consideration. Future work should explore how lessons such as those described here could be integrated into the national curriculum and effectively disseminated to teachers to give in remote areas across the country. Further study could focus on the effect of lessons delivered by remotely trained Malawian school teachers as a part of a framework that could be implemented nationally.
This study assessed the impact of a rabies lesson on an education naïve population and demonstrated that this is an effective way to improve knowledge in primary school children in an urban setting. Some aspects of the lesson were more effective than others in teaching children about rabies and its prevention. Knowledge remained greater than baseline suggesting the lesson allowed mid-term learning, though research to determine long-term knowledge retention is warranted. The educational component of this study took place alongside a rabies vaccination campaign and it was demonstrated that the vaccination campaign did not alter children’s knowledge about rabies or how to be safe around dogs. Rabies attitudes did not alter after the lesson but this was because children had already identified that rabies was serious prior to the lesson. The lesson format presented in this study was effective at teaching school children about rabies and its prevention using few resources and training. This study was conducted in the city of Zomba. We therefore believe that this lesson could be successfully used throughout urban Malawi providing children with effective techniques to reduce child mortality from this fatal disease. Future research is needed to assess the efficacy of this lesson in the rural setting and whether the increase in knowledge correlates to reduced risk of rabies.
|
10.1371/journal.pbio.2005952 | Spatiotemporal coordination of cell division and growth during organ morphogenesis | A developing plant organ exhibits complex spatiotemporal patterns of growth, cell division, cell size, cell shape, and organ shape. Explaining these patterns presents a challenge because of their dynamics and cross-correlations, which can make it difficult to disentangle causes from effects. To address these problems, we used live imaging to determine the spatiotemporal patterns of leaf growth and division in different genetic and tissue contexts. In the simplifying background of the speechless (spch) mutant, which lacks stomatal lineages, the epidermal cell layer exhibits defined patterns of division, cell size, cell shape, and growth along the proximodistal and mediolateral axes. The patterns and correlations are distinctive from those observed in the connected subepidermal layer and also different from the epidermal layer of wild type. Through computational modelling we show that the results can be accounted for by a dual control model in which spatiotemporal control operates on both growth and cell division, with cross-connections between them. The interactions between resulting growth and division patterns lead to a dynamic distributions of cell sizes and shapes within a deforming leaf. By modulating parameters of the model, we illustrate how phenotypes with correlated changes in cell size, cell number, and organ size may be generated. The model thus provides an integrated view of growth and division that can act as a framework for further experimental study.
| Organ morphogenesis involves two coordinated processes: growth of tissue and increase in cell number through cell division. Both processes have been analysed individually in many systems and shown to exhibit complex patterns in space and time. However, it is unclear how these patterns of growth and cell division are coordinated in a growing leaf that is undergoing shape changes. We have addressed this problem using live imaging to track growth and cell division in the developing leaf of the mustard plant Arabidopsis thaliana. Using subsequent computational modelling, we propose an integrated model of leaf growth and cell division, which generates dynamic distributions of cell size and shape in different tissue layers, closely matching those observed experimentally. A key aspect of the model is dual control of spatiotemporal patterns of growth and cell division parameters. By modulating parameters in the model, we illustrate how phenotypes may correlate with changes in cell size, cell number, and organ size.
| The development of an organ from a primordium typically involves two types of processes: increase in cell number through division, and change in tissue shape and size through growth. However, how these processes are coordinated in space and time is unclear. It is possible that spatiotemporal regulation operates through a single control point: either on growth with downstream effects on division, or on division with downstream effects on growth. Alternatively, spatiotemporal regulation could act on both growth and division (dual control), with cross talk between them. Distinguishing between these possibilities is challenging because growth and division typically occur in a context in which the tissue is continually deforming. Moreover, because of the correlations between growth and division it can be hard to distinguish cause from effect [1]. Plant development presents a tractable system for addressing such problems because cell rearrangements make little or no contribution to morphogenesis, simplifying analysis [2].
A growing plant organ can be considered as a deforming mesh of cell walls that yields continuously to cellular turgor pressure [3,4]. In addition to this continuous process of mesh deformation, new walls are introduced through cell division, allowing mesh strength to be maintained and limiting cell size. It is thus convenient to distinguish between the continuous expansion and deformation of the mesh, referred to here as growth, and the more discrete process of introducing new walls causing increasing cell number, cell division [5–8].
The developing Arabidopsis leaf has been used as a system for studying cell division control within a growing and deforming tissue. Developmental snapshots of epidermal cells taken at various stages of leaf development reveal a complex pattern of cell sizes and shapes across the leaf, comprising both stomatal and non-stomatal lineages [9]. Cell shape analysis suggests that there is a proximal zone of primary proliferative divisions that is established and then abolished abruptly. Expression analysis of the cell cycle reporter construct cyclin1 Arabidopsis thaliana β-glucuronidase (cyc1At-GUS) [10] shows that the proximal proliferative zone extends more distally in the subepidermal as compared with the epidermal layer. Analysis of the intensity of cyc1At-GUS, which combines both epidermal and subepidermal layers, led to a one-dimensional model in which cell division is restricted to a corridor of fixed length in the proximal region of the leaf [11]. The division corridor is specified by a diffusible factor generated at the leaf base, termed mobile growth factor, controlled by expression of Arabidopsis cytochrome P450/CYP78A5 (KLUH). Two-dimensional models have been proposed based on growth and cell division being regulated in parallel by a morphogen generated at the leaf base [12,13]. These models assume either a constant cell area at division, or constant cell cycle duration.
The above models represent important advances in understanding the relationships between growth and division, but leave open many questions, such as the relations of divisions to anisotropic growth, variations along both mediolateral and proximodistal axes, variation between cell layers, variation between genotypes with different division patterns, and predictions in relation to mutants that modify organ size, cell numbers, and cell sizes [14].
Addressing these issues can be greatly assisted through the use of live confocal imaging to directly quantify growth and division [15–22]. Local rates and orientations of growth can be estimated by the rate that landmarks, such as cell vertices, are displaced away from each other. Cell division can be monitored by the appearance of new walls within cells. This approach has been used to measure growth rates and orientations for developing Arabidopsis leaves and has led to a tissue-level model for its spatiotemporal control [16]. Live tracking has also been used to follow stomatal lineages and inform hypotheses for stomatal division control [23]. It has also been applied during a late stage of wild-type leaf development after most divisions have ceased [24]. However, this approach has yet to be applied across an entire leaf for extended periods to compare different cell layers and genotypes.
Here, we combine tracking and modelling of 2D growth in different layers of the growing Arabidopsis leaf to study how growth and division are integrated during organ morphogenesis. We exploit the speechless (spch) mutant to allow divisions to be followed in the absence of stomatal lineages, and show how the distribution and rates of growth and cell division vary in the epidermal and subepidermal layers along the proximodistal and mediolateral axes and in time. We further compare these findings to those of wild-type leaves grown under similar conditions. Our results reveal spatiotemporal variation in both growth rates and cell properties, including cell sizes, shapes, and patterns of division. By developing an integrated model of growth and division, we show how these observations can be accounted for by a model in which core components of both growth and division are under spatiotemporal control. Varying parameters of this model illustrates how changes in organ size, cell size, and cell number are likely interdependent, providing a framework for evaluating growth and division mutants.
To develop an integrated model of growth and division, we first tracked the epidermis of spch mutants, as they exhibit a simplified pattern of cell lineages [23]. Cell division dynamics were monitored by measuring spatiotemporal variation in two components: competence and execution. Competence refers to whether a cell has the potential to divide at some point in the future, whereas execution refers to a cell undergoing division (i.e., being cleaved into two).
Tracking cell vertices on the abaxial epidermis of spch seedlings imaged at about 12-h intervals allowed cells at a given developmental stage to be classified into those that would undergo division (competent to divide, green, Fig 1A), and those that did not divide for the remainder of the tracking period (black, Fig 1A). During the first time interval imaged (Fig 1A, 0–14 h), division competence was restricted to the basal half of the leaf, with a distal limit of about 150 μm (all distances are measured relative to the petiole-lamina boundary, Fig 1). To visualise the fate of cells at the distal limit, we identified the first row of nondividing cells (orange) and displayed them in all subsequent images. During the following time intervals, the zone of competence extended together with growth of the tissue to a distance of about 300 μm, after which it remained at this position, while orange boundary cells continued to extend further through growth. Fewer competent cells were observed in the midline region at later stages. Thus, the competence zone shows variation along the proximodistal and mediolateral axes of the leaf, initially extending through growth to a distal limit of about 300 μm and disappearing earlier in the midline region.
To monitor execution of division, we imaged spch leaves at shorter intervals (every 2 h). At early stages, cells executed division when they reached an area of about 150 μm2 (Fig 2A, 0–24 h). At later stages, cells in the proximal lamina (within 150 μm) continued to execute division at about this cell area (mean = 151 ± 6.5 μm2, Fig 2B), while those in the more distal lamina or in the midline region executed divisions at larger cell areas (mean = 203 ± 9.7 μm2 or 243.0 ± 22.4 μm2, respectively, Fig 2A, 2B and 2D). Cell cycle duration showed a similar pattern, being lowest within the proximal 150 μm of the lamina (mean = 13.9 ± 0.8 h) and higher distally (mean = 19.4 ± 1.8 h) or in the midline region (18.9 ± 2.1 h, Fig 2C and 2E). Within any given region, there was variation around both the area at time of division execution and the cell cycle duration (Fig 2F and 2G). For example, the area at execution of division within the proximal 150 μm of the lamina had a mean of about 150 μm2, with standard deviation of about 40 μm2 (Fig 2F). The same region had a cell cycle duration with a mean of about 14 h and a standard deviation of about 3 h. Thus, both the area at which cells execute division and cycle duration show variation around a mean, and the mean varies along the proximodistal and mediolateral axes of the leaf. These findings suggest that models in which either cell area at the time of division or cell cycle duration are fixed would be unable to account for the observed data.
To determine how cell division competence and execution are related to leaf growth, we measured areal growth rates (relative elemental growth rates [25]) for the different time intervals, using cell vertices as landmarks (Fig 1B). Areal growth rates varied along both the mediolateral and proximodistal axis of the leaf, similar to variations observed for competence and execution of division. The spatiotemporal variation in areal growth rate could be decomposed into growth rates in different orientations. Growth rates parallel to the midline showed a proximodistal gradient, decreasing towards the distal leaf tip (Fig 1C and S1A Fig). By contrast, mediolateral growth was highest in the lateral lamina and declined towards the midline, becoming very low there in later stages (Fig 1D and S1B Fig). The region of higher mediolateral growth may correspond to the marginal meristem [26]. Regions of low mediolateral growth (i.e., the proximal midline) showed elongated cell shapes. Models for leaf growth therefore need to account not only for the spatiotemporal pattern of areal growth rates but also the pattern of anisotropy (differential growth in different orientations) and correlated patterns of cell shape.
Cell size should reflect both growth and division: growth increases cell size while division reduces cell size. Cell periclinal areas were estimated from tracked vertices (Fig 1E). Segmenting a sample of cells in 3D showed that these cell areas were a good proxy for cell size, although factors such as leaf curvature introduced some errors (for quantifications see S5 Fig, and ‘Analysis of cell size using 3D segmentation’ in Materials and methods). At the first time point imaged, cell areas were about 100–200 μm2 throughout most of the leaf primordium (Fig 1E, left). Cells within the proximal 150 μm of the lamina remained small at later stages, reflecting continued divisions. In the proximal 150–300 μm of the lamina, cells were slightly larger, reflecting larger cell areas at division execution. Lamina cells distal to 300 μm progressively enlarged, reflecting the continued growth of these nondividing cells (Fig 1E and Fig 3A). Cells in the midline region were larger on average than those in the proximal lamina, reflecting execution of division at larger cell areas (Fig 1E and Fig 3C). Thus, noncompetent cells increase in area through growth, while those in the competence zone retain a smaller size, with the smallest cells being found in the most proximal 150 μm of the lateral lamina.
Visual comparison between areal growth rates (Fig 2B) with cell sizes (Fig 2E) suggested that regions with higher growth rates had smaller cell sizes. Plotting areal growth rates against log cell area confirmed this impression, revealing a negative correlation between growth rate and cell size (Fig 4B). Thus, rapidly growing regions tend to undergo more divisions. This relationship is reflected in the pattern of division competence: mean areal growth rates of competent cells in the lamina were higher than noncompetent cells, particularly at early stages (Fig 3I). However, there was no fixed threshold growth rate above which cells were competent, and for the midline region there was no clear difference between growth rates of competent and noncompetent cells (Fig 3I). Plotting areal growth rates for competent and noncompetent cells showed considerable overlap (S6 Fig), with no obvious switch in growth rate when cells no longer divide (become noncompetent). Thus, high growth rate broadly correlates with division competence, but the relationship is not constant for different regions or times.
To determine how the patterns and correlations observed for the epidermis compared to those in other tissues, we analysed growth and divisions in the subepidermis. The advantage of analysing an adjacent connected cell layer is that unless intercellular spaces become very large, the planar cellular growth rates will be very similar to those of the attached epidermis (because of tissue connectivity and lack of cell movement). Comparing the epidermal and subepidermal layers therefore provides a useful system for analysing division behaviours in a similar spatiotemporal growth context. Moreover, by using the spch mutant, one of the major distinctions in division properties between these layers (the presence of stomatal lineages in the epidermis) is eliminated.
Divisions in the abaxial subepidermis were tracked by digitally removing the overlying epidermal signal (the distalmost subepidermal cells could not be clearly resolved). As with the epidermis, 3D segmentation showed that cell areas were a good proxy for cell size, although average cell thickness was greater (S11 Fig, see also ‘Analysis of cell size using 3D segmentation’ in Materials and methods). Unlike the epidermis, intercellular spaces were observed for the subepidermis. As the tissue grew, subepidermal spaces grew and new spaces formed (Fig 5A–5D). Similar intercellular spaces were observed in subepidermal layers of wild-type leaves, showing they were not specific to spch mutants (S8 Fig).
Vertices and intercellular spaces in the subepidermis broadly maintained their spatial relationships with the epidermal vertices (Fig 5C, 5E and 5F). Comparing the cellular growth rates in the plane for a patch of subepidermis with the adjacent epidermis showed that they were similar (S9 Fig), although the subepidermal rates were slightly lower because of the intercellular spaces. This correlation is expected, because unless the intercellular spaces become very large, the areal growth rates of the epidermal and subepidermal layers are necessarily similar.
The most striking difference between subepidermal and epidermal datasets was the smaller size of the distal lamina cells of the subepidermis (compare Fig 6A with Fig 1E, and Fig 3A with Fig 3B). For the epidermis, these cells attain areas of about 1,000 μm2 at later stages, while for the subepidermis they remain below 500 μm2. This finding was consistent with the subepidermal division competence zone extending more distally (Fig 6B), reaching a distal limit of about 400 μm compared with 300 μm for the epidermis. A more distal limit for the subepidermis has also been observed for cell cycle gene expression in wild type [10]. Moreover, at early stages, divisions occurred throughout the subepidermis rather than being largely proximal, as observed in the epidermis, further contributing to the smaller size of distal subepidermal cells (S10 Fig). Despite these differences in cell size between layers, subepidermal cell areal growth rates showed similar spatiotemporal patterns to those of the overlying epidermis, as expected because of tissue connectivity (compare Fig 6C with Fig 1B). Consequently, correlations between growth rate and cell size were much lower for the subepidermis than for the epidermis (Fig 4B and 4C).
This difference in the relationship between growth and cell size in different cell layers was confirmed through analysis of cell division competence. In the subepidermis, at early stages there was no clear difference between mean growth rates for competent and noncompetent cells (Fig 3J cyan, green), in contrast to what is observed in the epidermis (Fig 3I cyan, green), while at later stages noncompetent cells had a slightly lower growth rate (Fig 3J yellow, red).
To determine how the patterns of growth and division observed in spch related to those in wild type, we imaged a line generated by crossing a spch mutant rescued by a functional SPCH protein fusion (pSPCH:SPCH-GFP) to wild type expressing the PIN3 auxin transporter (PIN3:PIN3-GFP), which marks cell membranes in the epidermis [23]. The resulting line allows stomatal lineage divisions to be discriminated from non-stomatal divisions (see below) in a SPCH context. At early stages, wild-type and spch leaves were not readily distinguishable based on cell size (S12 Fig). However, by the time leaf primordia attained a width of about 150 μm, the number and size of cells differed dramatically. Cell areas in wild type were smaller in regions outside the midline region, compared with corresponding cells in spch (Fig 7A). Moreover, cell divisions in wild type were observed throughout the lamina that was amenable to tracking (Fig 7B, 0–12 h), rather than being largely proximal. Divisions were observed over the entire lamina for subsequent time intervals, including regions distal to 300 μm (Fig 7B, 12–57 h). These results indicate that SPCH can confer division competence in epidermal cells outside the proximal zone observed in spch mutants.
To further clarify how SPCH influences cell division, we used SPCH-GFP signal to classify wild-type cells into two types: (1) Stomatal lineage divisions, which include both amplifying divisions (cells express SPCH strongly around the time of division and retain expression in one of the daughter cells) (S1 Video, orange/yellow in Fig 7C) and guard mother cell divisions (SPCH expression is bright and diffuse during the first hours of the cycle, transiently switched on around time of division, and then switched off in both daughters). (2) Non-stomatal divisions, in which SPCH expression is much weaker, or only lasts <2 h, and switches off in both daughter cells (S2 Video, light/dark green in Fig 7C).
If cells with inactive SPCH behave in a similar way in wild-type or spch mutant contexts, we would expect non-stomatal divisions to show similar properties to divisions in the spch mutant. In the first time interval, non-stomatal divisions (green) were observed within the proximal 150 μm (Fig 7C, 0–12 h), similar to the extent of the competence zone in spch (Fig 1A, 0–14h). The zone of non-stomatal divisions then extended to about 250 μm and became restricted to the midline region. After leaf width was greater than 0.45 mm, we did not observe further non-stomatal divisions in the midline region, similar to the situation in spch leaves at a comparable width (Fig 1A, 58-74h, 0.48 mm). These results suggest that similar dynamics occur in the non-stomatal lineages of wild type and the spch mutant.
To determine how SPCH modulates division, we analysed stomatal and non-stomatal divisions in the lamina. Considerable variation was observed for both the area at which cells divide (25–400 μm2) and cell cycle duration (8–50 h) (S13 Fig). The mean area at which cells execute division was greater for non-stomatal divisions (about 165 ± 28 μm2 [1.96 × standard error]) than stomatal divisions (about 80 ± 6 μm2) (S13 Fig). Similarly, cell cycle durations were longer for non-stomatal divisions (about 25 ± 3 h) compared with stomatal divisions (about 18 ± 1 h). These results suggest that in addition to conferring division competence, SPCH acts cell autonomously to promote division at smaller cell sizes and/or for shorter cell cycle durations.
Given the alteration in cell sizes and division patterns in wild type compared to spch, we wondered if these may reflect alterations in growth rates. When grown on agar plates, spch mutant leaves grow more slowly than wild-type leaves (S14A Fig). The slower growth of spch could reflect physiological limitations caused by the lack of stomata, or an effect of cell size on growth—larger cells in spch cause a slowing of growth. However, the tracking data and cell size analysis of spch and wild type described above were carried out on plants grown in a bio-imaging chamber in which nutrients were continually circulated around the leaves. Growth rates for wild type and spch leaves grown in these conditions were comparable for much of early development, and similar to those observed for wild type on plates (compare Fig 7D with Fig 1B, S14 Fig). These results suggest that the reduced growth rates of spch compared with wild type at early stages on plates likely reflect physiological impairment caused by a lack of stomata rather than differences in cell size. As a further test of this hypothesis, we grew fama (basic helix-loop-helix transcription factor bHLH097) mutants, as these lack stomata but still undergo many stomatal lineage divisions [27]. We found that fama mutants attained a similar size to spch mutants on plates, consistent with the lack of stomata being the cause of reduced growth in these conditions (S14 Fig).
Plots of cell area against growth rates of tracked leaves grown in the chamber showed that, for similar growth rates, cells were about three times smaller in wild type compared with spch (compare Fig 4A with Fig 4B). Thus, the effects of SPCH on division can be uncoupled from effects on growth rate, at least at early stages of development.
At later stages (after leaves were about 1 mm wide), spch growth in the bio-imaging chamber slowed down compared with wild type, and leaves attained a smaller final size. This later difference in growth rate might be explained by physiological impairment of spch because of the lack of stomata, and/or by feedback of cell size on growth rates. This change in later behaviour may reflect the major developmental and transcriptional transition that occurs after cell proliferation ceases [9].
The above results reveal that patterns of growth rate, cell division, and cell size and shape exhibit several features in spch: (1) a proximal corridor of cell division competence, with an approximately fixed distal limit relative to the petiole-lamina boundary; (2) the distal limit is greater for subepidermal (400 μm) than epidermal tissue (300 μm); (3) a further proximal restriction of division competence in the epidermis at early stages that extends with growth until the distal limit of the corridor (300 μm) is reached; (4) larger and narrower cells in the proximal midline region of the epidermis; (5) a proximodistal gradient in cell size in the epidermal lamina; (6) a negative correlation between cell size and growth rate that is stronger in the epidermis than subepidermis; (7) variation in both the size at which cells divide and cell cycle duration along both the proximodistal and mediolateral axes; and (8) variation in growth rates parallel or perpendicular to the leaf midline. In wild-type plants, these patterns are further modulated by the expression of SPCH, which leads to division execution at smaller cell sizes and extension of competence, without affecting growth rates at early stages. Thus, growth and division rates exhibit different relations in adjacent cell layers, even in spch, in which epidermal-specific stomatal lineages are eliminated, and division patterns can differ between genotypes (wild type and spch) without an associated change in growth rates.
These observations argue against spatiotemporal regulators acting solely on the execution of division, which then influences growth, as this would be expected to give conserved relations between division and growth. For the same reason, they argue against a single-point-of-control model in which spatiotemporal regulators act solely on growth, which then secondarily influences division. Instead, they suggest dual control, with spatiotemporal regulators acting on both growth and division components. With dual control, growth and division may still interact through cross-dependencies, but spatiotemporal regulation does not operate exclusively on one or the other.
To determine how a hypothesis based on dual control may account for all the observations, we used computational modelling. We focussed on the epidermal and subepidermal layers of the spch mutant, as these lack the complications of stomatal lineages. For simplicity and clarity, spatiotemporal control was channelled through a limited set of components for growth and division (Fig 8A). There were two components for growth under spatiotemporal control: specified growth rates parallel and perpendicular to a proximodistal polarity field (Kpar and Kper, respectively) [16]. Together with mechanical constraints of tissue connectivity, these specified growth components lead to a pattern of resultant growth and organ shape change [28]. There were two components for cell division under spatiotemporal control: competence to divide (CDIV), and a threshold area for division execution that varies around a mean (Ā). Controlling division execution by a threshold cell size (Ā) introduces a cross-dependency between growth and division, as cells need to grow to attain the local threshold size before they can divide. The cross-dependency is indicated by the cyan arrow in Fig 8A, feeding information back from cell size (which depends on both growth and division) to division. An alternative to using Ā as a component of division-control might be to use a mean cell cycle duration threshold. However, this would bring in an expected correlation between high growth rates and large cell sizes (for a given cell cycle duration, a faster-growing cell will become larger before cycle completion), which is the opposite trend of what is observed.
Spatiotemporal regulators of growth and division components can be of two types: those that become deformed together with the tissue as it grows (fixed to the tissue) and those that maintain their pattern to some extent despite deformation of the tissue by growth (requiring mobile or diffusible factors) [28]. In the previously published growth model, regulatory factors were assumed, for simplicity, to deform with the tissue as it grows [16]. These factors comprised a graded proximodistal factor (PGRAD), a mediolateral factor (MID), a factor distinguishing lamina from petiole (LAM), and a timing factor (LATE) (S15A and S15B Fig). However, such factors cannot readily account for domains with limits that remain at a constant distance from the petiole-lamina boundary, such as the observed corridors for division competence. This is because the boundary of a domain that is fixed to the tissue will extend with the tissue as it grows. We therefore introduced a mobile factor, proximal mobile factor (PMF), that was not fixed to the tissue to account for these behaviours. This motivation is similar to that employed by others [11–13]. PMF was generated at the petiole-lamina boundary and with appropriate diffusion and decay coefficients such that PMF initially filled the primordium and then showed a graded distribution as the primordium grew larger, maintaining a high concentration in the proximal region and decreasing towards the leaf tip (S15C and S15D Fig). This profile was maintained despite further growth, allowing thresholds to be used to define domains with relatively invariant distal limits. Further details of the growth model are given in Materials and methods, and the resultant growth rates are shown in S16 Fig (compare with Fig 1B and 1D).
Cells were incorporated by superimposing polygons on the initial tissue or canvas (S15A Fig, right). The sizes and geometries of these virtual cells (v-cells) were based on cells observed at corresponding stages in confocal images of leaf primordia [16]. The vertices of the v-cells were anchored to the canvas and displaced with it during growth. Cells divided according to Errera’s rule: the shortest wall passing through the centre of the v-cell [29], with noise in positioning of this wall incorporated to capture variability. V-cells were competent to divide if they expressed factor CDIV, and executed division when reaching a mean cell target area, Ā. As the observed area at time of division was not invariant (Fig 2F), we assumed the threshold area for division varied according to a standard deviation of σ = 0.2Ā around the mean. CDIV and Ā are the two core components of division that are under the control of spatiotemporal regulators in the model (Fig 8A, 8C and 8D). Variation between epidermal and subepidermal patterns reflects different interactions controlling cell division (interactions colour coded red and blue, respectively, in Fig 8C and 8D).
We first modelled cell divisions in the subepidermis, as this layer shows a more uniform pattern of cell sizes (Fig 3B and Fig 6A). Formation of intercellular spaces was simulated by replacing a random selection of cell vertices with small empty equilateral triangles, which grew at a rate of 2.5% h−1, an average estimated from the tracking data. To account for the distribution of divisions and cell sizes, we assumed that v-cells were competent to divide (express CDIV) where PMF was above a threshold value. This value resulted in the competence zone extending to a distal limit of about 400 μm. To account for the proximodistal pattern of cell areas in the lamina (Fig 3B and Fig 6A) and larger cells in the midline (Fig 3D and Fig 6A), we assumed that Ā was modulated by the levels of PMF, PGRAD, and MID (Fig 8D, black and blue). These interactions gave a pattern of average v-cell areas and division competence that broadly matched those observed (compare Fig 8E and 8F with Fig 6A and 6B, and Fig 3F and 3H with 3B and 3D, S3 Video).
For the epidermis, the zone of division competence was initially in the proximal region of the primordium and then extended with the tissue as it grew (Fig 1A). We therefore hypothesised that in addition to division being promoted by PMF, there was a further requirement for a proximal factor that extended with the tissue as it grew. We used PGRAD to achieve this additional level of control, assuming CDIV expression requires PGRAD to be above a threshold level (Fig 8C, red and black). V-cells with PGRAD below this threshold were not competent to divide, even in the presence of high PMF. Thus, at early stages, when PMF was high throughout the primordium, the PGRAD requirement restricted competence to the proximal region of the leaf (Fig 8H). At later stages, as the PGRAD domain above the threshold extended beyond 300 μm, PMF became limiting, preventing CDIV from extending beyond about 300 μm. To account for the earlier arrest of divisions in the midline region (Fig 1A), CDIV was inhibited by MID when LATE reached a threshold value (Fig 8C, red). As well as CDIV being regulated, the spatiotemporal pattern of Ā was modulated by factors MID and PMF (Fig 8D black).
With these assumptions, the resulting pattern of epidermal divisions and v-cell sizes broadly matched those observed experimentally for the epidermis (compare Fig 8G with Fig 1E, S4 Video). In particular, the model accounted for the observed increases in cells sizes with distance from the petiole-lamina boundary, which arise because of the proximal restrictions in competence (compare Fig 3E and 3G with Fig 3A and 3C). The model also accounted for the elongated cell shapes observed in the midline region, which arise through the arrest of division combined with low specified growth rate perpendicular to the polarity. Moreover, the negative correlations between growth rates and cell size, not used in developing the model, were similar to those observed experimentally (Fig 4B and 4D). These correlations arise because both growth and division are promoted in proximal regions.
We also measured the cell topology generated by the epidermal model. It has previously been shown that the frequency of six-sided neighbours observed experimentally for the spch leaf epidermis is very low compared with that for other plant and animal tissues and also with that generated by a previous implementation of Errera’s rule (S17 Fig) [30]. The topological distribution generated by the epidermal leaf model gave a six-sided frequency similar to that observed experimentally, falling two standard deviations away from the mean and thus close to a reasonable fit (S17 Fig). The increased similarity of the model output to the spch leaf epidermal topology, compared with a previous implementation of Errera’s rule [31], may reflect the incorporation of anisotropic growth in our model. If polarity is removed from our model to render specified growth as isotropic (while preserving local areal growth rates), the frequency of six-sided neighbours increases, becoming more like the empirical data for the shoot apical meristem (S17 Fig). A further likely contribution to the lowering of six-sided neighbour frequency generated by our model is the use of random noise to displace the positioning of new walls, rather than positioning them always to pass precisely through the cell centre. Thus, our analysis shows how incorporating more realistic growth patterns can be valuable in evaluating division rules.
Taken together, the simulations show that the pattern of growth and division can be broadly accounted for by factors modulating specified growth rates (Kpar and Kper) and cell division components (CDIV and Ā). Variation between epidermal and subepidermal patterns generated by the models reflects different interactions controlling cell division (Fig 8C and 8D).
Many mutants have been described that influence cell division and/or leaf size [32,33]. To gain a better understanding of such mutants, we explored how changes in key parameters in our model may alter leaf size, cell size, and cell number. As leaf size is normally measured at maturity, we first extended our analysis to later stages of development. Tracking spch to later stages of development showed that overall growth rates declined, on average, while remaining relatively high towards the proximal region of the lamina (S4B Fig), consistent with a previous study [18]. Cell divisions were not observed after the leaf reached a width of about 0.9 mm (S4A Fig, 96h). To capture arrest of division, we assumed that CDIV was switched off throughout the leaf after LATE reached a threshold value.
In the previously published growth model [16], the decline of growth rates with developmental time was captured through an inhibitory effect of LATE on growth. To extend the model to later stages and bring about eventual arrest of growth, we assumed that LATE increased exponentially after 189 h and inhibited both Kper and Kpar thereafter. Parameters for growth inhibition were adjusted to give a final leaf width of about 3 mm, which was the final size attained for leaf 1 in spch mutants in the bio-imaging chamber. The v-cell sizes generated by the model broadly matched the patterns observed (Fig 9A and 9B, S5 Video). As epidermal divisions have ceased by the time the spch leaf is about 1 mm wide, all the growth depicted in Fig 9A and 9B occurs in the absence of division (i.e., cell expansion). However, a notable discrepancy between the model output and the experimental data was the generation of distal v-cells that exceeded the values observed (about 20,000 μm2 compared with about 10,000 μm2). A similar result was obtained if the model was tuned to match not only the final leaf width but also the reduced growth rate of spch in the growth chamber at later stages (S14B and S14C Fig). A better fit was obtained by inhibiting specified growth rates in distal regions at later stages. This inhibition was implemented by introducing inhibitory factors with levels that increased distally. The result was that distal v-cells remained at or below about 10,000 μm2 (Fig 9C and S6 Video). We refer to this as the limit-free model. Another way of limiting the size of distal v-cells was to introduce feedback from cell size to growth, so that the specified growth rate decreased as v-cells approached upper size limits (Fig 9J and S7 Video). This feedback corresponds to introducing a further interaction in the regulatory pathway (Fig 8A, magenta). We refer to this as the limiting cell size model.
We varied parameters in both the limit-free model (Fig 9C) and the limiting cell size model (Fig 9J) to see how the parameters influence cell number, cell size, and final leaf size. Increasing Ā by a constant amount did not change leaf size with the limit-free model but resulted in fewer, larger v-cells (Fig 9D). Reducing Ā resulted in a leaf with more v-cells that were, on average, smaller but did not change leaf size (Fig 9E). With the limiting cell size model, increasing or decreasing Ā had similar effects as with the limit-free model but also slightly reduced or increased leaf size (Fig 9K and 9L). Thus, it is possible to affect cell number and size without a major effect on organ size or growth.
To investigate how changing growth parameters influences cell numbers and areas, we reduced the specified growth rates (values for Kpar and Kper) by 5%. For the limit-free model this resulted in a smaller leaf with both smaller and fewer v-cells (Fig 9F). There were fewer cells because they grew more slowly and thus took longer to reach Ā, and cells were smaller because they grew at a slower rate after they had ceased dividing. Conversely, increasing specified growth rate by 5% led to larger leaves, with more v-cells that were, on average, larger (Fig 9G). The model with limiting cell size gave similar results (Fig 9M and 9N). Thus, modulating growth rates has consequences on organ size, cell size, and cell number. This may account for why many mutants with smaller organs have both fewer cells and smaller cells [34].
To examine the effect of changes in developmental timing, we altered the onset of LATE. Moving the onset earlier for the limit-free model led to smaller leaves because of the earlier decline in growth rate (Fig 9H). There were fewer v-cells because of the earlier arrest of division, and there was also a slight reduction in v-cell size. Delaying the onset of LATE had the opposite effect of increasing leaf size, cell number, and cell size (Fig 9I). The limiting cell size model gave similar results (Fig 9O and 9P). Thus, changes in developmental timing affected organ size and cell number, with a lesser effect on cell size. This is because changing LATE shifts both the onset of the growth rate decline and the time of division arrest (inactivation of CDIV).
A further application of the model is to explore the effects of the environment on leaf growth and division. To illustrate this possibility, we analysed data for the spch mutant grown on plates, which exhibits a greatly reduced growth rate compared with growth in the chamber (S14A and S14B Fig). A prediction of the model is that cell divisions should cease when the leaf is at a smaller size (i.e., the leaf will have grown less by the time the threshold value of LATE for division arrest is reached). In addition, as spch plants grown on plates have impaired general physiology, the rate of developmental progression (physiological time) may also be slowed down. We simulated these effects by modifying the model parameters such that the overall growth rate was reduced by 40% and physiological time reduced by 45%. This gave a growth curve matching that observed for spch grown on plates (blue line, S14A Fig). As expected, this model takes longer to attain a given leaf width (e.g., 0.5 mm) than the original model. The resulting cell areas are larger at the 0.5-mm leaf-width stage, particularly in proximal regions, because divisions arrest when the leaf is at a smaller size, so all subsequent cell growth occurs in the absence of division (Fig 10A and 10B and S18 Fig).
To test this prediction of enlarged cell size, we compared leaves when they had attained a width of about 0.5 mm (Fig 10C and 10D), which is just before divisions cease for spch grown in the chamber (Fig 1). Cells in the proximal lamina of the chamber-grown leaves were relatively small (mean = 123.3 ± 6.4 μm2 for region shown in Fig 10I), typical of dividing cells (Fig 10C and 10G); whereas those of the plate-grown leaves were larger (mean = 199.8 ± 17.3 μm2 for region shown in Fig 10J), indicating division arrest (Fig 10D and 10H and S18 Fig). Proximal lamina cells in plate-grown leaves also showed greater shape complexity, typical of pavement cells that have ceased division (Fig 10K–10N and S18 Fig). These results suggest that cell divisions in much of the lamina cease when the leaf is smaller for plate-grown compared to chamber-grown leaves, as predicted by the model. The sizes of midline cells for plate-grown leaves predicted by the model are larger than those observed (compare Fig 10B with Fig 10H), indicating that withdrawal of competence from this region, as implemented in the model, may be activated too early. Conversely, the most proximal lamina cells in the plate-grown leaves (dark blue cells, Fig 10H) are smaller than predicted (Fig 10B), suggesting that the uniform arrest of division when LATE reaches a threshold value is an oversimplification.
Growth rates, cell division, and cell shapes and sizes in the growing first leaf of Arabidopsis exhibit complex spatiotemporal patterns. The main features observed in spch are (1) a proximal corridor of division competence with an approximately fixed distal limit; (2) the distal limit is greater for subepidermal compared to epidermal tissue; (3) a further proximal restriction of division competence in the epidermis at early stages that extends with growth until the distal limit is reached; (4) a proximodistal gradient in cell size in the epidermal lamina; (5) larger and narrower cells in the proximal midline region of the epidermis; (6) a negative correlation between cell size and growth rate that is stronger in the epidermis than in the subepidermis; (7) variation in both the size at which cells divide and cell cycle duration along both the proximodistal and mediolateral axes; (8) variation in growth rates parallel or perpendicular to the leaf midline. In wild-type plants these patterns are further modulated by expression of SPCH, which leads to division execution at smaller cell sizes and extension of competence, without affecting growth rates at early stages.
The observed varying relations between growth rates and division between tissue layers and genotypes argue against single-point-of-control models, in which spatiotemporal regulators act solely through either division or growth. Instead, they suggest dual control, in which spatiotemporal regulators act on both growth and division, with cross talk between them. We show that a model based on dual control can broadly account for the data. In this model, spatiotemporal control is channelled through two growth components (specified growth rates parallel and perpendicular to polarity) and two division components (competence and mean threshold size for division) (Fig 8A). The growth components reflect turgor and cell wall extensibility in different orientations, and the division components reflect regulatory mechanisms for partitioning cells. Orientation information is provided by a tissue-wide polarity field, for which direct evidence has recently been obtained in both wild-type and spch mutants [22,35]. The polarity field may be established through a biochemical mechanism as proposed here, likely involving tissue-level coordinated cell polarity [36]. Alternatively, information could be relayed through mechanical stresses [20,22,37]. The resulting patterns of growth and division determine the distribution of cell sizes and shapes and organ shape. The implications, limitations, and questions raised by this model are discussed below.
Execution of leaf cell division does not occur at an unvarying cell size, even within a given region and developmental stage. Similar variability has been observed for cell divisions in apical meristems [21,38]. Variability may reflect experimental errors in estimation of cell size, stochasticity in the process of division, and/or mechanisms other than geometric size sensing that influence division execution (e.g., factors such as vacuole size, which is not monitored in our analysis). We model such variability by explicitly adding variation around a mean threshold size needed for division, Ā. Controlling division execution by a threshold cell size (Ā) introduces a cross-dependency between growth and division, as cells need to grow to attain the local threshold size before they can divide. An alternative to using Ā would be to use a mean cell cycle duration threshold. However, this would bring in an expected correlation between high growth rates and large cell sizes (for a given cell cycle duration, faster growing cells will become larger before cycle completion), which is the opposite of the correlation observed.
In contrast to the epidermal layer, intercellular spaces are observed in the subepidermis of wild-type and spch from early stages. The spaces may originate, in part, from a reduction in adhesion between subepidermal cells, allowing cell walls to become detached from each other. In addition to reduced adhesion, a further requirement for intercellular spaces is that cells are not too tightly packed against each other. Packing may be reduced if subepidermal cells have lower specified growth rates than the epidermis. Subepidermal cells could move away or be pulled apart from each other, as epidermal growth creates more space than they can fill through their own expansive growth. According to this view, the epidermis rather than the subepidermis provides the expansive force driving planar growth, in contrast to what has been described for other tissues, such as the stem [39]. A primary role for the epidermis in driving planar growth is also consistent with the observed developmental effects of epidermal gene activity [40]. However, it is possible that the subepidermis provides a restraint on growth, which could account for the effect of subepidermal tissue on leaf shape in some chimeras [41].
Spatiotemporal control of growth and division in the model of spch is established through combinatorial interactions between five factors: PGRAD, MID, LAM, LATE, and a mobile factor that allows proximal corridors with fixed distal limits to be established (PMF). PMF is similar to the previously proposed mobile growth factor [11], except that the effect of PMF on division does not have a consequential effect on growth. To account for the difference in distal limits of the division corridor between cell layers, PMF action extends more distally in the subepidermis compared with the epidermis, either because the competence threshold requirement for PMF is lower in the subepidermal layer, or because PMF levels are higher. A candidate factor for coordinating proliferation between layers is the transcriptional coactivator ANGUSTIFOLIA3 [42,43]. Candidates for LAM are LEAFY PETIOLE [44] and members of the YABBY gene family [45], which are expressed in the lamina and promote lateral outgrowth.
A fixed corridor for division has also been described for other systems such as the root, where a division zone is maintiained at a distance of about 300–500 μm from the quiescent centre in Arabidopsis [46]. In contrast to the leaf, regions of highest growth rate in the root are outside the cell division zone, providing further support for a dual control mechanism. The spatial extent of the division zone in roots is maintained through auxin-cytokinin interactions [47]. Auxin-cytokinin interactions also influence leaf growth and division: temporal arrest of leaf growth depends on auxin-induced cytokinin breakdown [48]; increased cytokinin degradation in leaf primordia can accelerate termination of cell proliferation [49]; and accumulation of specific cytokinins may promote indeterminate leaf growth [50]. However, it is currently unclear whether auxin, cytokinin, and/or other molecular players underlie PMF.
A limitation of our model is that it does not consider modulation of growth or division near the leaf margin, creating serrations [51,52]. Serrations have previously been modelled by displacement of the leaf outline without modelling the tissue growth explicitly [52,53]. In terms of the modelling framework described here, they may reflect alterations in polarity and/or growth rates of tissue, and accounting for these behaviours would require the introduction of additional factors into the model, as illustrated by generation of winglike outgrowths in barley lemma mutants [54].
To account for the further proximal restriction of competence in the epidermis at early stages, PGRAD limits divisions in the epidermis until the distal limit set by PMF is reached. PMF also interacts with MID in the epidermis, accounting for larger cells in the midline region. The elongated shape of proximal midline cells is a result of early arrest of division combined with low specified growth rates perpendicular to the proximodistal polarity.
Divisions in the wild-type epidermis are also influenced by SPCH. We show that SPCH acts autonomously in the epidermis to confer competence, and has little impact in the proximal midline region, where its activity has previously been shown to be low [55]. The autonomous effect of SPCH on division competence contrasts with its nonautonomous effects at later stages of development, with regard to layer thickness and photosynthetic capacity [56]. This difference in autonomy may reflect primary and secondary consequences of SPCH activity. SPCH also promotes asymmetric divisions and divisions at smaller cell sizes or shorter cell cycle durations. The complex pattern of divisions in wild type epidermis observed here and elsewhere [9] would thus reflect the combined effect of PMF, PGRAD, MID, and SPCH, although the molecular basis of these interactions remains to be established.
In agreement with [24], we observed that mean cell cycle duration is relatively constant for wild type (about 20 h). However, cell cycle duration varies from 8 h to 50 h around the mean. Some of this variation depends on whether SPCH is active: epidermal cells that do not show high SPCH activity divide at a larger cell size and longer cell cycle duration. Moreover, the size at which cells with active SPCH divide is not fixed but becomes progressively smaller with successive divisions [23], indicating that cell cycle duration likely becomes shorter as well. Thus, the spatiotemporal variation in cell cycle duration may be the consequence of variation in growth rates (for a given threshold division size, cell cycle duration depends on growth rate) and/or direct control of cell cycle length.
Most small-leaf mutants have both fewer and smaller cells [34]. Such outcomes can be generated with the model by reducing specified growth rates. The leaves end up smaller because of the lower growth rate, cells are smaller because they grow less after divisions have arrested, and there are fewer cells because they grow more slowly and thus take longer to reach Ā. Thus, the observation that organ size, cell size, and cell number are commonly reduced together in mutants is a natural outcome of the model.
Change in developmental timing through factor LATE also leads to changes in leaf size, although this is mainly reflected in changes in cell number rather than cell size. This is because changing LATE shifts both the onset of growth rate decline and the time of division arrest (loss of division competence). Such variation in developmental timing could underlie mutants that change organ size with little or no effect on cell size, such as kluh and big brother [57,58].
Loss of expression of D-type cyclins leads to premature termination of cell division and fewer cells autonomously in each layer, without a major change in leaf size [59,60]. Such features can be captured by changing model parameters that are specific to cell division, such as the value of Ā, in one or more layers (Fig 9D, 9E, 9K and 9L). This situation corresponds to compensation [61–63], as change in cell number is counterbalanced by a change in cell area (organ size is preserved). However, no dedicated mechanism for counterbalancing is needed, as division is under separate spatiotemporal control from growth in our model.
Although execution of division does not have an immediate effect on growth rates in our model, we explore the possibility of feedback from division on growth at later developmental stages. If growth slows down when cells approach an upper size limit, then cell division could postpone the slowing down of growth by reducing cell size. Such a mechanism would lead to cell division extending the duration of growth, thus increasing leaf size. Mature leaves display an array of final cell sizes that correlate with levels of endoreduplication [64,65], suggesting that as cells approach a size limit, endocycles are induced that allow them to surpass the limit. If endoreduplication is impaired, these cell size limits may not be so easily overcome, leading to smaller leaves with smaller cells [66,67]. However, the extent to which endoreduplication is limited in wild type and thus may constrain final cell size and growth is unclear. Through modelling, we show that it is possible to account for the data with or without feedback from cell size on growth.
If endocycles are promoted as cells enlarge, then promoting division (e.g., by reducing Ā) should lead to lower levels of endoreduplication (as cells will be smaller). This prediction is in accord with the effect of overexpressing D-type cyclins, which leads to smaller cells with lower levels of endoreduplication [68]. Conversely, inhibiting cell division (e.g., by increasing Ā) should give larger cells and higher levels of endoreduplication, as observed with cycd3 mutants [7]. However, if both division and the ability to endoreduplicate are impaired, cell size may eventually feedback to inhibit growth rate, giving smaller leaves and perhaps accounting for the phenotype of ant mutants, which have smaller organs with larger cells that do not endoreduplicate more than wild type [7,69].
A further application of the model is to explore the effects of different environments on leaf growth and division. As an illustration, we compared leaves of spch mutants grown in a bio-imaging chamber (in which nutrients were continually circulated around the leaves) with those grown on agar plates (in which growth rate is greatly reduced). Cell divisions arrested when leaves were at a smaller size in the slow-growing conditions, as predicted by the model in which division arrest depends on a timing mechanism (LATE). However, the growth and cells sizes observed suggests that the timing mechanisms are not based on external time but passage of physiological time, which may also be affected by altered growth conditions.
The model presented here identifies core components of growth and division that may be regulated and interact to generate the spatiotemporal patterns observed. Further integrative studies on growth and division at the subcellular, cellular, and tissue level in different genotypes and environments should help provide a deeper understanding of the mechanisms by which regulatory factors are established and control these core components.
For tracking growth of the speechless mutant, we used the previously published Arabidopsis line, spch-1, containing a fluorescently labelled plasma membrane marker [70]. To more precisely determine division execution times, we crossed the spch mutant to an Arabidopsis line containing fluorescently labelled nuclei, HTA11-GFP [71], and PIN3:PIN3-GFP [72], which labels plasma membranes in the epidermal layer only. For tracking growth in the wild-type background and to distinguish cells in the stomatal lineage, we used the previously published Arabidopsis line containing pSPCH:SPCH-GFP and PIN3:PIN3-GFP [23]. For measuring leaf widths in the fama mutant we used the previously published line fama-1 (Ohashi-Ito and Bergmann, 2006).
Seeds were surface sterilised with 70% ethanol containing 0.05% Sodium Dodecyl Sulfate (SDS) for 10 min and then rinsed with 100% ethanol. Sterilised seeds were sown on petri dishes containing 25 mL of MS growth media {1× Murashige and Skoog salt mixture, 1% (w/v) sucrose, 100 mg/mL inositol, 1 mg/mL thiamine, 0.5 mg/mL pyridoxin, 0.5 mg/mL nicotinic acid, 0.5 mg/mL MES, 0.8% (w/v) agar, pH 5.7} and kept at 4 oC in the dark for 72 h (stratification). Plates were then transferred to a controlled environment room (CER) at 20 oC in long-day conditions (16-h light/8-h dark cycles) for 5–8 d.
At 5–8 d after stratification, seedlings were transferred under sterile conditions into an autoclaved optical live-imaging chamber [16,73] and continuously supplied with 1/4 strength MS liquid growth medium, including sucrose to support growth. Time-lapse imaging was carried out at regular intervals using a Leica SP5 Confocal microscope, a Zeiss LSM 5 EXCITER confocal microscope, or a Zeiss LSM 780 confocal microscope. For experiments imaged with a high temporal resolution (intervals of 1–2 h), the chamber remained mounted on the microscope stage for the duration of the experiment, with room temperature and photoperiod set to be similar to that of the CER in which seedlings were germinated. For experiments with a longer interval between imaging (12–24 h), the chamber was returned to the CER between confocal imaging. Experiments were carried out on leaf 1 within the range of 0.15–2.75 mm width. Seedlings were positioned in the chamber such that the abaxial epidermis of the leaf was oriented approximately parallel and adjacent to the coverslip, although it curved away to some extent at the leaf margins. This curvature affected the leaf outline produced when projected images were made from confocal image stacks. Leaf outlines (indicated by dotted lines in Fig 1, Fig 2, Fig 6, Fig 7, Fig 9, S2 Fig, S3 Fig, S4 Fig, S6 Fig, and S9 Fig) reflect projections onto the imaging plane rather than being corrected for curvature and thus convey a shape that appears narrower than the actual leaf outline. Some regions could not be tracked because of occlusion by overlapping leaves (at early developmental stages) or because movement in the z-dimension caused parts of the leaf to go out of focus. Thus, some cell lineages could not be traced all the way back to the initial time point. Images are available from https://figshare.com/s/b14c8e6cb1fc5135dd87.
To facilitate cell tracking, confocal image stacks were converted into 2D projections using either Volviewer [74] (http://cmpdartsvr3.cmp.uea.ac.uk/wiki/BanghamLab/index.php/VolViewer) or Fiji [75]. For early stages, when the leaf could be captured in a single scan, VolViewer was used to create a projection of the leaf surface. At later stages, when leaves were larger, multiple overlapping tiled scans were required to capture the entire leaf. In such cases, Fiji was used to create multiple 2D projections, which were merged together using Photoshop to create a single composite image. Leaf width was measured in 3D, when possible, using VolViewer. For later stages, leaf width was measured in 2D from merged projections using Fiji. Projections of the subepidermal layer were created in VolViewer using the ‘Depth-Peal Shader’ lighting editor. Several projections were created for each z-stack (using different parameters to reveal as many cells as possible in approximately the middle of the cell layer) and merged together using Photoshop to create a composite image.
Projected confocal images were used to calculate growth rates and cell areas and monitor cell division dynamics in 2D by placing points around the vertices of individual cells using PointTracker, as described in [16]. A toolset (Track ‘n’ R) was created for ImageJ (https://imagej.nih.gov/ij/) to facilitate access to ImageJ macros and offer improved visualisation of PointTracker data using R [76]. Track ‘n’ R was used to create leaf outlines, visualise the zone of cell division, and analyse cell lineages and to display cell cycle duration, cell area at division execution, and growth rates (source code and detailed instructions for Track’n’R and PointTracker have been deposited at https://github.com/fpantin/Track-n-R and https://figshare.com/s/b14c8e6cb1fc5135dd87 respectively).
Graphical outputs from Track ‘n’ R were reoriented so that the leaf tip pointed upwards. Cellular growth rates over a time interval t1–t2 were calculated according to ln(At2−At1)/(t2−t1), where At1 is cell area at t1 and At2 is cell area at t2. If a cell divided in this interval, At2 was the area of the clone it gave rise to at time t2.
For each tracking experiment, the first row of nondividing cells was identified in the first time point and coloured orange by hand using Photoshop. These cells were identified in each subsequent image and also coloured orange (Fig 1A and S3 Fig and S4 Fig). The approximate location of the petiole-lamina boundary was identified based on the shape of the leaf outline in the last image available for each dataset. A cell was identified in the midline of this image, in line with the base of the leaf lamina. This cell was then traced back to through each image to its earliest ancestor in the first time point, thus identifying the location of the petiole-lamina boundary even when the leaf shape was less developed. Cells were identified as part of the midline region based on appearance (shape and location) in the last image of each tracking experiment. The lineage of these cells were traced back to the beginning of the experiment (S2 Fig). Cells that did not form part of the midline region were classed as lamina cells.
For 3D segmentation and volume measurements, confocal image stacks were processed using Python scripts, as described [17], with additional scripts added to measure the external surfaces of epidermal cells in 3D and the corresponding 2D projections (source code and detailed instructions have been deposited at https://figshare.com/s/b14c8e6cb1fc5135dd87). Fiji macros [75] using the 3D Viewer and Point Picker plugins were used to visualise images and select cells during manual quality control.
For the epidermis, plotting projected segmentation-based area against vertex-based area gave a good linear fit (R2 = 0.87) with a gradient of about 1, showing that vertex-based cell area is a good proxy for projected cell area (S5A Fig). Areas extracted from the cell surface plotted against vertex-based cell area also gave a good linear fit (R2 = 0.77) with a gradient of about 1.2 (S5B Fig). The higher value for the gradient likely reflects curvature of the cell surface and the leaf, both of which increase area compared to projected values. Nevertheless, segmented surface area remains linearly related to vertex-based area. Plotting cell volume against segmentation-based cell surface area gave a linear fit, with R2 = 0.91 and a gradient suggesting an approximately constant cell thickness of about 9 μm (S5C Fig). Variation in cell thickness is displayed by plotting cell volume divided by surface area as a heat map. Although a slight increase in cell thickness was observed in the proximal midline (about 15 μm), cell thickness showed relatively little spatial variation for much of the lamina (S5D Fig), compared with the striking spatiotemporal variation in cell area (S5E Fig) and cell volume (S5F Fig). Thus, the major contribution to cell size variation derives from cell area rather than cell thickness. These results are also consistent with fixed leaf sections shown in [10], which have epidermal cells in the range of 8–15 μm thick (measured according to the scale in the published images).
For the subepidermal cell layer, fewer cells could be segmented in 3D because the bases of the cells were too deep within the tissue to be captured clearly by confocal imaging. However, around 13 cells could be segmented and plotting projected segmentation-based area against vertex-based cell area gave a good linear fit (R2 = 0.98) with a gradient of about 1 (S11A Fig). Plotting the volume of these cells against projected segmentation-based cell areas showed that they had similar thickness to epidermal cells of the same area, except for cells in the proximal midline region, where subepidermal cells have a greater volume because of increased thickness (S11B and S11C Fig). These results are also consistent with fixed leaf sections shown in [10], which have subepidermal cells in the range of 9–14 μm thick.
To quantify cell shape complexity, we employed Lobe-Contribution Elliptic Fourier Analysis (LOCO-EFA), a method to decompose the outline of each cell into a list of biologically meaningful descriptors [77]. The LOCO-EFA decomposition was used to estimate a measure of cell shape complexity, coined the ‘cumulative difference’ (CD), which is the integral over all LOCO-EFA modes larger than 2 of the mismatch (Exclusive OR or XOR) between the original and reconstituted shape, yielding a scalar value representing the degree of shape complexity of each cell [77]. We used this measure normalised to cell area; hence, a small or large cell with the same shape will yield the same cell complexity measure (CD). It ranged from zero (low complexity, which describes perfectly circular or elongated cells) to higher values, as more LOCO-EFA harmonics are required to accurately describe the shape. The computer code was written in C and is available on a remote repository (Git repository), which is publicly available on Bitbucket (https://bitbucket.org/mareelab/LOCO_EFA).
The expression pattern of pSPCH:SPCH-GFP was analysed from time-lapse images to distinguish cells in the stomatal lineages from non-stomatal lineages. For each cell division, the duration of SPCH expression was determined from the time when SPCH first became visible in the nucleus of the mother cell to when it could no longer be seen in each daughter cell. S1 Video shows an example of cell division in the stomatal lineage. S2 Video shows an example cell division in a non-stomatal lineage.
To facilitate imaging of the subepidermal cell layer in wild-type leaves, seedlings grown on plates were stained by the modified pseudo-Schiff propidium iodide (mPS-PI) method, as previously described [78]. After approximately 1 wk (for the mounting solution to set), leaf primordia were imaged using a Leica SP5 Confocal microscope. Projections of the subepidermal layer were created in VolViewer using the ‘Depth-Peal Shader’ lighting editor, as described above for spch in “Image processing”.
All models and GFT-box software used for modelling can be downloaded from http://cmpdartsvr3.cmp.uea.ac.uk/wiki/BanghamLab/index.php/Software.
Models are also downloadable from https://figshare.com/s/b14c8e6cb1fc5135dd87. To implement an integrated model of division and growth, we built on a previously published tissue-level model for wild-type leaf growth at early stages of development [16]. This model has two interconnected networks: the Polarity Regulatory Network specifies tissue polarity and hence specified orientations of growth, and the Growth Regulatory Network (KRN) determines how factors influence specified growth rates. Specified growth orientations are established in relation to a polarity field, determined by the local gradient of a factor determining polarity field (POL) that propagates through the tissue, termed the canvas. The resultant growth and shape depend on the specified growth rates parallel (Kpar) and perpendicular (Kper) to the polarity, and the mechanical constraints arising from the connectedness of the tissue.
In the equations, factors are denoted by i subscripted with the factor name. For instance, the factor PGRAD is described by ipgrad in the equations. Factors may promote growth rates through the linear function pro, defined as follows:
pro(pf,if)=1+pfif
where if is a factor, F, and pf is a promotion coefficient for that factor. Factors may inhibit growth through the function inh, defined as follows:
inh(hf,if)=1/(1+hfif)
where hf is a inhibition coefficient for factor F. All multiplications and divisions are elementwise.
The previously proposed tissue-level growth model [16] was based on tracking only a subset of cell vertices and therefore had a lower cellular resolution than the data presented in this paper. Based on the higher resolution the cell fate map of the midline region of wild type and spch (S2 Fig), we widened the initial MID domain (S15 Fig) so that it gave a better match to the cellular data. Running the model with this change produced a narrower leaf, as MID inhibits Kper. To compensate for this effect and to account for the regions with high growth rate perpendicular to the midline (Fig 1D and S1B Fig), we promoted Kper with PMF.
The initial starting canvas for all models consists of 3,000 finite elements, which are not subdivided during the simulations, and model time is aligned with days after initiation (DAI), which is defined based on growth curves of leaf width [16]. To give finer resolution, times are given in hours (hours after initiation [HAI]). A list of growth parameter values is given in Table 1.
Specified growth rates are modulated by a set of overlapping regional factors, PGRAD, MID, and LAM, the concentrations of which are fixed to the canvas and deform with it during growth (S15A Fig). PGRAD declines distally and accounts for the proximodistal variation in growth rate parallel to the polarity. LAM is expressed highest in the presumptive lamina and at lower levels in the distal regions that will form the petiole. LAM promotes growth perpendicular to polarity. MID is expressed in the midline region, as shown in S15A Fig, and inhibits growth perpendicular to the polarity. The maximum value of these factors is 1. Specified growth rates are also modulated by diffusible factor PMF, which is fixed to a value of 1 at the approximate position of the lamina-petiole boundary and allowed to diffuse through growth with a diffusion rate of dpmf and a decay rate of μpmf, giving the distribution shown in S15 Fig.
A temporally varying factor, LATE, is activated throughout the canvas to decrease growth at later stages. The value of LATE is initially 0 but rises linearly with time after 149 h:
ilate{0ift<148hglate(t−148h)ift≥148h
where glate defines the increase of LATE with time. LATE inhibits specified growth rates with an inhibition coefficient of hlate.
Polarity is established using factor PROXORG, which is set to 1 at the base of the canvas and 0 elsewhere (S15A Fig). The value of POL is fixed at a value of bpol, where PROXORG is greater than zero. POL diffuses throughout the canvas with a diffusion rate of Dpol and a decay rate of μpol. POL distribution is allowed to establish during the setup phase for 20 time steps before the commencement of growth. Polarity is initially proximodistal and then deforms with the canvas as it grows to its final shape.
This is a model for spch subepidermis during early stages of development.
To incorporate cell divisions within our tissue-level model, we superimposed polygons on the initial canvas to represent cells (S15A Fig, right). The sizes and geometries of these v-cells are based on cells observed at corresponding stages in confocal images of leaf primordia [16]. The vertices of the v-cells are anchored to the canvas and displaced with it during growth. New vertices are introduced as v-cells divide, according to the shortest wall passing through the centre of the v-cell [29]. Calling this the nominal new wall, the actual new wall is chosen to be parallel to this, through a point that is randomly displaced from the midpoint of the nominal new wall. The displacement is a vector chosen uniformly at random from a disc centred on the midpoint. The radius of this disc is 0.25 times the length of the nominal new wall. The length of the new wall is shortened slightly to give more realistic wall angles [79].
Cell divisions are determined through controlling competence and Ā.
This is a model for spch epidermis during early stages of development.
This model extends the Fig 8G and 8H—spch epidermis model to later stages of development.
A new factor, EARLYGROWTH, was introduced in the model setup and set to a value of 1 throughout the canvas. After 189 h, EARLYGROWTH decreases linearly by a value of 0.0417 h−1 until it reaches a minimum value of 0.
To arrest growth, we modified factor LATE to increase exponentially after 189 h:
ilate{0ift<148hglate(t−148h)if189h>t≥148hAeBtift≥189h
where A = glate (189–148) e−B 189 and B = 1 / (189–148). This ensures ilate evaluates to glate (t– 148) at 189 h.
This model inhibits distal growth during later stages in order to reduce the size of distal cells. It does this by using the existing factors PGRAD and LAM.
In this model, cell size can affect growth. This is an alternative way to limit the size of distal cells without having to use the factors PGRAD and LAM, as in the Fig 9C—Later stage spch limit-free epidermis model.
Cell division threshold models were developed for the Fig 9C—later stage spch limit-free epidermis model and Fig 9J—later stage spch limiting cell size epidermis model. Each of the cell division models was identical to its parent model, but the cell target area for division, Ā, was increased by a constant a' for t ≥ 114 h.
In Fig 9D and 9K, a' = 85 μm2, while in Fig 9E and 9L, a' = −85 μm2.
Growth rate mutant models were developed for the Fig 9C—later stage spch limit-free epidermis model and the Fig 9J—later stage spch limiting cell size epidermis model. Each of the growth rate mutant models was identical to its parent model, but Kper and Kpar were globally scaled by a factor k'. In Fig 9F and 9M, k' = 0.95, while in Fig 9G and 9N, k' = 1.05.
For the Fig 9C—later stage spch limit-free epidermis model:
Kpar=k'_.ppgrad¡pgrad.inh(hlate,¡late.inh(1.5,(1‑¡earlygrowth)).inh(2,(1‑¡lam).(1‑¡earlygrowth)).inh(4,(1‑¡pgrad).(1‑¡earlygrowth))
Kper=k'_.plam¡lam.inh(hmid,¡mid).pro(ppmf,¡pmftk).pro(plate,¡late.¡earlygrowth).inh(1.2,¡late.(1‑¡earlygrowth)).inh(4,(1‑¡pgrad).(1‑¡earlygrowth))
For the Fig 9J—later stage spch limiting cell size epidermis model:
Kpar=k'_.ω.Ppgrad¡pgrad.inh(hlate,¡late).inh(0.24,¡late.(1‑¡earlygrowth))
Kper=k'_.ω.Plam¡lam.inh(hmid,¡mid).pro(Ppmf,¡pmftk).pro(Plate,¡late.¡earlygrowth).inh(2.8,¡late.(1‑¡earlygrowth))
Changes to models relative to those used to generate Fig 9C and Fig 9J are shown underlined.
Models were developed for the Fig 9C—later stage spch limit-free epidermis model and Fig 9J—later stage spch limiting cell size epidermis model. Each of the LATE mutant models was identical to its parent model, but the activation of LATE and EARLYGROWTH was shifted by a constant number of hours, t′. In Fig 9H and 9O, t' = −6 h, while in Fig 9I and 9P, t' = 6 h.
ilate{0ift<148h+t′_glate(t−148h+t′_)if189h+t′_>t≥148h+t′_AeBtift≥189h+t′_
where A = glate. (189+t′ – 148+t′). e-B 189+t' and B = 1 / (189+t' – 148+t').
Changes to models relative to those used to generate Fig 9C and Fig 9J are shown underlined.
In the above models, t refers to actual time. We modified the late stage spch epidermis model (as used for Fig 9B) by setting physiological time to be a constant fraction (physiological ratio) of duration since the start of the simulation (when t = 87 h). Key transitions and growth rates were then set in relation to physiological time. Parameters describing physical processes such as diffusion were left unchanged. For the model for spch grown on plates, the physiological ratio was 0.55. In addition, the growth rates were globally scaled by a factor k' = 0.6 The net result of these two changes is that growth in actual time is slowed to 0.33 of normal. If this overall growth rate was matched purely by changing physiological time, the leaf on the plate would end up larger than observed at maturity. Conversely, if the growth rate was matched purely through changes in k', the leaf would end up much smaller than observed for spch on plates at maturity. Thus, changes in both physiological time and k' are needed to match the observed growth curve.
The only change in relation to the late stage spch epidermis model (as used for Fig 9B) was that the physiological ratio (as defined above) was set to 0.75.
|
10.1371/journal.pgen.1007443 | Convergent evolution of gene expression in two high-toothed stickleback populations | Changes in developmental gene regulatory networks enable evolved changes in morphology. These changes can be in cis regulatory elements that act in an allele-specific manner, or changes to the overall trans regulatory environment that interacts with cis regulatory sequences. Here we address several questions about the evolution of gene expression accompanying a convergently evolved constructive morphological trait, increases in tooth number in two independently derived freshwater populations of threespine stickleback fish (Gasterosteus aculeatus). Are convergently evolved cis and/or trans changes in gene expression associated with convergently evolved morphological evolution? Do cis or trans regulatory changes contribute more to gene expression changes accompanying an evolved morphological gain trait? Transcriptome data from dental tissue of ancestral low-toothed and two independently derived high-toothed stickleback populations revealed significantly shared gene expression changes that have convergently evolved in the two high-toothed populations. Comparing cis and trans regulatory changes using phased gene expression data from F1 hybrids, we found that trans regulatory changes were predominant and more likely to be shared among both high-toothed populations. In contrast, while cis regulatory changes have evolved in both high-toothed populations, overall these changes were distinct and not shared among high-toothed populations. Together these data suggest that a convergently evolved trait can occur through genetically distinct regulatory changes that converge on similar trans regulatory environments.
| Convergent evolution, where a similar trait evolves in different lineages, provides an opportunity to study the repeatability of evolution. Convergent morphological evolution has been well studied at multiple evolutionary time scales ranging from ancient, to recent, such as the gain in tooth number in freshwater stickleback fish. However, much less is known about the accompanying evolved changes in gene regulation during convergent evolution. Here we compared evolved changes in gene expression in dental tissue of ancestral low-toothed marine fish to fish from two independently derived high-toothed freshwater populations. We also partitioned gene expression changes into those affecting a gene’s regulatory elements (cis), and those affecting the overall regulatory environment (trans). Both freshwater populations have evolved similar gene expression changes, including a gain of expression of putative dental genes. These similar gene expression changes are due mainly to shared changes to the trans regulatory environment, while the cis changes are largely population specific. Thus, during convergent evolution, overall similar and perhaps predictable transcriptome changes can evolve despite largely different underlying genetic bases.
| Development is controlled by a complex series of interlocking gene regulatory networks. Much of this regulation occurs at the level of transcription initiation, where trans acting factors bind to cis regulatory elements to control their target gene’s expression [1,2]. Evolved changes in an organism's morphology are the result of changes in this developmental regulatory landscape. It has been proposed that the genetic bases of many of these evolved changes are mutations within the cis-regulatory elements of genes [3–5]. Indeed, recent work in evolutionary genetics suggests the molecular bases of a diverse array of traits from Drosophila wing spots [6] to mouse pigmentation [7] to stickleback armored plate number [8,9] and size [10] are changes in the activity of cis-regulatory elements.
Evolved changes in gene expression can be divided into two broad regulatory classes. Cis regulatory changes can occur within the proximal promoter [11], distal enhancer [12], or the gene body itself [13], and result in allele-specific gene expression differences in hybrid diploids [14]. Trans regulatory changes modify the overall regulatory environment [15,16], but are usually genetically unlinked to the expression change, and do not result in allele-specific expression in hybrid diploids. For any gene with an evolved expression difference, the total evolved gene expression difference can be partitioned into changes in cis and trans by quantifying expression differences between two populations and also testing for expression differences between alleles in F1 hybrids between the two populations [14]. As both alleles in F1 hybrids animals are exposed to the same regulatory environment, any difference in their expression must be due to a cis-regulatory change. Several studies have attempted to characterize evolved cis and trans-regulatory changes at a transcriptome-wide level [17–21]. Though the relative contribution of cis and trans regulatory changes varies extensively among studies, cis changes have been found to dominate [17,18,21] or at least be approximately equivalent to trans changes [19,20,22]. Additionally, compensatory changes (cis and trans changes in opposing directions) have been found to be enriched over neutral models [17,18], showing evidence for selection for stable gene expression levels. However, none of these studies examined contribution of cis and trans gene expression changes during convergent morphological evolution.
Populations evolve new traits following a shift to a novel environment, due to a mixture of drift and selection. Truly adaptive traits can often be repeatedly observed in multiple populations following a similar ecological shift. Threespine sticklebacks are an excellent system for the study of evolved changes in phenotypes, including gene expression [23–27]. Marine sticklebacks have repeatedly colonized freshwater lakes and streams along the coasts of the Northern hemisphere [28]. Each of these freshwater populations has independently adapted to its new environment; however, several morphological changes, including a loss in armored plates and a gain in tooth number, are shared among multiple newly derived populations [29,30]. The repeated evolution of lateral plate loss is due to repeated selection of a standing variant regulatory allele of the Eda gene within marine populations [8,9] and genome sequencing studies found over a hundred other shared standing variant alleles present in geographically diverse freshwater populations [31]. These studies suggest the genetic basis of freshwater adaptation might typically involve repeated reuse of the same standing variants to evolve the same adaptive freshwater phenotype.
However, more recent evidence has shown that similar traits have also evolved through different genetic means in freshwater stickleback populations. A recent study which mapped the genetic basis of a gain in pharyngeal tooth number in two independently derived freshwater populations showed a largely non-overlapping genetic architecture [30]. Another study using three different independently derived benthic (adapted to the bottom of a lake) populations showed that, even when adapting to geographically and ecologically similar environments, the genetic architecture of evolved traits is a mix of shared and unique changes [32]. Even in cases where the same gene is targeted by evolution in multiple populations (the loss of Pitx1 expression resulting in a reduction in pelvic spines), the individual mutations are often independently derived [33,34]. All of these genomic scale studies have looked at the genetic control of morphological changes, while the extent and nature of genome-wide gene expression changes has been less studied. It remains an open question as to whether similar gene expression patterns evolve during the convergent evolution of morphology, and if so, to what extent those potential shared gene expression changes are due to shared cis or trans changes.
Teeth belong to a class of vertebrate epithelial appendages (including mammalian hair) that develop from placodes, and have long served as a model system for studying organogenesis and epithelial-mesenchymal interactions in vertebrates [35]. Odontogenesis is initiated and controlled by complex interactions between epithelial and mesenchymal cell layers, and involves several deeply conserved signaling pathways [36–38]. Sticklebacks retain the ancestral jawed vertebrate condition of polyphyodonty, or continuous tooth replacement, and offer an emergent model system for studying tooth replacement. Previous work has supported the hypothesis that two independently derived freshwater stickleback populations have evolved an increase in tooth replacement rate, potentially mediated through differential odontogenic stem cell dynamics [30]. Recent studies have found teeth and taste bud development to be linked, with one study supporting a model where teeth and taste buds are copatterned from a shared oral epithelial source [39], and another study supporting a model where teeth and taste buds share a common progenitor stem cell pool [40].
We sought to examine the evolution of the regulatory landscape controlling stickleback tooth development and replacement. Using high-throughput RNA sequencing (RNA-seq) in parental non-hybrid fish, we found that two independently derived high-toothed freshwater populations display highly convergent gene expression changes, especially in orthologs of known tooth-expressed genes in other vertebrates, likely reflecting the convergently evolved tooth gain phenotype and the deep homology of teeth across all jawed vertebrates. We also quantitatively partitioned these evolved gene expression changes into cis and trans regulatory changes [14,19] in both populations at a transcriptome-wide level using RNA-seq on F1 marine-freshwater hybrids. We found that trans regulatory changes predominate evolved changes in gene expression in dental tissue. Additionally, we found that the trans regulatory changes are more likely to be shared between the freshwater populations than the cis regulatory changes. Thus, similar downstream transcription networks controlling tooth development and replacement have convergently evolved largely through different upstream genetic regulatory changes.
To test whether multiple freshwater populations have evolved increases in tooth number compared to multiple ancestral marine populations [30,41], we quantified total ventral pharyngeal tooth number of lab reared sticklebacks from four distinct populations: (1) a marine population from the Little Campbell river (LITCM) in British Columbia, Canada, (2) a second marine population from Rabbit Slough (RABSM) in Alaska, USA, (3) a benthic freshwater population from Paxton Lake (PAXBFW) in British Columbia, Canada, and (4) a second freshwater population from Cerrito Creek (CERCFW) in California, USA (Fig 1A and 1B). Freshwater fish from both populations had more pharyngeal teeth than marine fish at this 35-50mm standard length (SL) stage, consistent with previous findings [30,41] of increases in tooth number in freshwater sticklebacks (Fig 1B and 1C, S1 Table).
To estimate the genomic relatedness of these populations, we resequenced the genomes of three marine and six freshwater sticklebacks from the four different populations (S2 Table). We aligned the resulting reads (mean of ~53 million reads per sample, see Methods and S2 Table) to the stickleback reference genome [31] using Bowtie2 [42], and called 8.3 million (see Methods) variants using the Genome Analysis Toolkit (GATK) [43–45]. As it has been previously shown that Pacific marine stickleback populations are an outgroup to freshwater populations from Canada (PAXBFW) and California (CERCFW) [31], we hypothesized the two high-toothed populations would be more related to each other genomically than either marine population. A phylogeny constructed using a down-sampled set of 67.5 thousand genome-wide variants (see Methods) cleanly separated freshwater populations from each other and from marine fish (S1A Fig). Principal component analysis using 1.7 million filtered genome-wide variants (see Methods) revealed that the first principle component explains nearly half (41.4%) of the overall variance and separates PAXBFW sticklebacks from both CERCFW and marine fish (S1B Fig), representing the independent evolution of PAXBFW genomes. The second principal component separated both freshwater populations from marine populations, showing partially shared freshwater genome evolution. These results further support the model that populations of freshwater sticklebacks used a combination of shared and independent genetic changes [31,32] when evolving a set of similar morphological changes in response to a new environment.
As morphological changes are often the result of changes in gene expression patterns and levels, we sought to identify evolved changes in gene expression during tooth development at stages soon after the evolved differences emerge [41]. We quantified gene expression in ventral pharyngeal dental tissue for three females each from the two high-toothed freshwater (PAXBFW and CERCFW) and Alaskan (RABSM) low-toothed marine populations using RNA-seq (Fig 2A, S3 and S4 Tables). Principal component (PC) analysis of the resulting gene expression matrix showed a clustering of gene expression by population, with the first PC separating PAXBFW samples, and the second PC separating both PAXBFW and CERCFW samples from marine, similar to the PC analysis of the genome-wide variants (Fig 2B) [46].
Given the convergently evolved morphological change of increases in tooth number, we hypothesized that convergent evolution has occurred at the gene expression level in freshwater dental tissue. To test this hypothesis, we performed a differential expression analysis, defining evolved changes in gene expression as changes found to be significant in a differential expression analysis using cuffdiff2 [47]. We compared evolved change in gene expression in PAXBFW dental tissue (PAXBFW expression vs marine) to the evolved change in CERCFW dental tissue (CERCFW expression vs marine). We found 6,693 and 3,501 genes (out of a total of 22,442) with significant (as determined by cuffdiff2 [47], see Methods) evolved expression changes in PAXBFW and CERCFW respectively. Of these genes with evolved expression changes, 2,223 were called differentially expressed in both populations, with 1,898 (85%) showing expression changes in the same direction relative to marine.
At a genome-wide level, correlated changes in gene expression levels have evolved in the two high-toothed freshwater populations (Fig 2C, Spearman's r = 0.43). We next asked if orthologs of genes implicated in tooth development in other vertebrates showed an increase in correlated evolved expression changes. We compared the gene expression changes of stickleback orthologs of genes in the BiteIt (http://bite-it.helsinki.fi/) [48] or ToothCODE (http://compbio.med.harvard.edu/ToothCODE/) [36] databases (hereafter referred to as the “BiteCode” gene set, S5 Table), two databases of genes implicated in mammalian tooth development. Consistent with the conserved roles of gene regulatory networks regulating mammalian and fish teeth [49–52] and the major evolved increases in tooth number in both freshwater populations (Fig 1C), these predicted dental genes showed an increase in their correlated evolved gene expression change (Fig 2C red points, Spearman's r = 0.68), and tended to have an overall increase in gene expression (S2 Fig, P = 7.36e-6, GSEA, see methods). This correlation coefficient was higher than any observed in over 100,000 bootstrapped (sampled with replacement) gene sets of the same size from the same gene expression matrix. We also examined the expression levels of genes whose orthologs are annotated as being expressed in zebrafish pharyngeal teeth (www.zfin.org). Within this gene set, 27 of 40 genes were significantly more highly expressed in at least one freshwater population, with no genes expressed significantly higher (as determined by cuffdiff2 [47,53–55], see Materials and Methods) in marine samples than either freshwater population (Fig 2D).
Tooth development is controlled by several deeply conserved developmental signaling pathways [50,52]. To test whether expression changes in the components of specific developmental signaling pathways have evolved in the two high-toothed freshwater populations, we next analyzed the expression levels of stickleback orthologs of genes implicated in mammalian tooth development and annotated as components of different signaling pathways [36]. When comparing gene expression levels in freshwater dental tissue to marine dental tissue, genes annotated as part of the TGF-ß signaling pathway displayed significantly increased expression in freshwater dental tissue (S3A–S3F Fig).
Since these two freshwater populations have a largely different developmental genetic basis for their evolved tooth gain [30], we next asked whether any pathways were upregulated or downregulated specifically in one freshwater population. When comparing the expression of genes in PAXBFW dental tissue to expression in CERCFW or marine dental tissue, genes not only in the TGF-ß pathway, but also in the WNT signaling pathway, displayed significantly increased expression, consistent with the differing genetic basis of tooth gain in these populations (S3B Fig). In contrast, no significant pathway differences were found comparing CERCFW to PAXBFW or marine (S3C Fig).
We next asked whether any pathways, regardless of previous implication in tooth development, were significantly upregulated in either or both freshwater transcriptomes. Genes upregulated in freshwater dental tissue were enriched for Gene Ontology (GO) terms involved in anatomical structure development, signaling, and regulation of cell proliferation (S4A Fig, S6 Table). Genes upregulated in PAXBFW dental tissue over marine were enriched for GO terms involved in cell proliferation, division and cell cycle regulation, as well as DNA replication (S4B Fig, S7 Table), while genes upregulated in CERCFW over marine were enriched for GO terms involved in cell locomotion, movement, and response to lipids (S4C Fig, S8 Table). 204 of the 454 and 432 GO terms that were enriched in genes upregulated in PAXBFW and CERCFW relative to marine, respectively, were shared, further supporting the convergent gain of freshwater gene expression.
As teeth are constantly being replaced in polyphyodont adult fish, potentially due to the action of dental stem cells [40], we hypothesized that genes involved in stem cell maintenance have evolved increased expression in freshwater tooth plates, given the higher rate of newly forming teeth previously found in adults [30], and the possibly greater number of stem cell niches in high-toothed fish. We further hypothesized that since teeth are developmentally homologous to hair, perhaps an ancient genetic circuit regulating vertebrate placode replacement controls both fish tooth and mammalian hair replacement. For example, the Bmp6 gene, previously described as expressed in all stickleback teeth [41] was significantly upregulated in CERCFW fish, consistent with the evolved major increases in tooth number in this population (S4 Table). In contrast, no such significant upregulation was observed in the expression of PAXBFW Bmp6 (S4 Table), consistent with the observed evolved cis-regulatory decrease in PAXBFW Bmp6 expression [41]. Further supporting this hypothesis, the expression of the stickleback orthologs of a previously published set of mouse hair follicle stem cell (HFSC) signature genes [56] were significantly upregulated in freshwater dental tissue (S3A Fig), with 84 and 75 out of 254 genes displaying significant increases in expression in PAXBFW and CERCFW, respectively. CERCFW dental tissue displayed a small but significant increase in expression of this set of HFSC orthologs relative to both PAXBFW and marine samples (S3C Fig).
In cichlid fish, pharmacology experiments revealed that reductions in tooth density can be accompanied by concomitant increases or decreases in taste bud density [39]. To begin to test whether derived high-toothed stickleback populations have also evolved significantly altered levels of known taste bud marker gene expression, we examined the expression levels of known taste bud markers Calbindin2 and Phospholipase Beta 2 [57], as well as taste receptors such as Taste 1 Receptor Member 1, Taste 1 Receptor Member 3, and Polycystin 2 Like 1 [58]. Although four of these five genes had detectable significant expression changes between different populations, no consistent freshwater upregulation or downregulation of taste bud marker genes was seen (S5 Fig).
Evolved changes in gene expression are due to a combination of cis acting changes that are linked to the genes they act on, and trans acting changes which usually are genetically unlinked to the gene or genes they regulate. Since the genetic basis of freshwater tooth gain mapped to largely non-overlapping intervals in these two populations [30], we hypothesized that the observed shared freshwater gene expression changes were the result of a similar trans environment, but a largely different set of cis changes. To test this hypothesis, we measured evolved cis expression changes in marine-freshwater F1 hybrids, which have marine and freshwater alleles present in the same trans environment. We raised both CERCFW-marine and PAXBFW-marine F1 hybrids to the late juvenile stage, dissected their ventral pharyngeal tooth plates, then generated and sequenced five barcoded RNA-seq libraries per population (10 total). We then quantified the cis expression change as the ratio of the number of reads mapping uniquely to the freshwater allele of a gene to the number of uniquely mapping marine reads (Fig 3A, S9–S11 Tables). Trans expression changes were calculated by factoring the cis change out from the overall parental expression change [19].
We found 11,832 and 8,990 genes in PAXBFW and CERCFW F1 hybrids, respectively, that had a fixed marine-freshwater sequence difference which had more than 20 total reads mapping to it. We observed no significant bias towards either the marine or freshwater allele in either set of F1 hybrids (Fig 3B). We next classified genes into one of four categories (cis change only, trans change only, concordant cis and trans changes, discordant cis and trans changes). We found 1640 and 1116 PAXBFW (Fig 3C) and CERCFW (Fig 3D) genes, respectively, with only significant cis changes, and 1873 and 1048 genes, respectively, with only significant trans changes. We also found 478 and 359 genes with significant cis and trans changes in the same direction, which we term concordant changes in gene expression. Conversely, we found 772 and 607 genes with significant cis and trans changes in opposing directions, which we termed discordant changes. Discordant cis and trans changes were more common in both populations, suggesting selection for stable levels of gene expression.
We next wanted to determine the relative contribution of cis and trans gene expression changes to evolved changes in gene expression. We restricted our analysis to differentially expressed genes (as determined by cuffdiff2 [47]) to examine only genes with a significant evolved difference in gene expression and quantifiable (i.e. genes with transcripts containing a polymorphic variant covered by at least 20 reads) cis and trans expression changes. When evolving a change in gene expression, the cis and trans regulatory basis for this change can be concordant (cis and trans effects both increase or decrease expression) or discordant (cis effects increase and trans decrease or vice versa). We hypothesized that genes would tend to display more discordant expression changes, as stabilizing selection has been found to buffer gene expression levels [17,22,59]. To test this hypothesis, we binned differentially expressed genes into a 2x2 contingency table, with genes classified as cis or trans based on which effect controlled the majority of the evolved expression change, and discordant or concordant based on the direction of the cis and trans changes (Fig 4A and 4B). In the CERCFW population, significantly more discordant changes than expected by a neutral model (P = 1.35e-7, binomial test) have evolved. In both populations, we found increased discordant changes when the trans effect is larger than the cis effect (P = 1.29e-7, 1.44e-13, PAXBFW and CERCFW respectively, binomial test). In both populations, we observe the opposite (an enrichment of concordant changes) when the cis effect is stronger, relative to the ratio when the trans effect is dominant (P = 1.34e-36, 8.2e-11 PAXBFW and CERCFW respectively, binomial test). When considering all (not just differentially expressed) genes with quantifiable cis and trans expression changes, discordant changes dominated regardless of the relative strength of the cis effect (S6 Fig).
If all gene expression changes were due to changes only in cis, we would expect to see the measured cis ratios in the hybrids match the parental expression ratios. Instead, in both cases of evolved change, we saw parental expression ratios of a greater magnitude than F1 hybrid ratios, indicating a stronger contribution of trans changes to overall gene expression changes (Fig 3C and 3D). Indeed, when we examined the overall percentage of expression changes of differentially expressed genes that were due to changes in cis, we observed median per gene values of only 25.2% and 32.5% of PAXBFW and CERCFW gene expression changes, respectively (Fig 4C). Comparing the expression levels of orthologs of known dentally expressed genes from the BiteIt [48] and ToothCODE [36] databases revealed a similarly small number of gene expression changes explained by changes in cis, relative to the genome-wide average (Fig 4D). Evolved changes in CERCFW gene expression were more due to changes in cis than PAXBFW genes (Fig 4D, P = 1.25e-22, Mann-Whitney U test). Thus, trans effects on gene expression dominate the evolved freshwater gene expression changes.
We next wanted to test the hypothesis that the shared freshwater gene expression changes were primarily due to shared trans changes, rather than shared cis changes. We first compared the overall expression levels of genes called differentially expressed between PAXBFW and marine as well as CERCFW and marine. We restricted our analysis to differentially expressed genes whose cis-regulatory change we were able to measure in our F1 hybrids, including genes without a significant cis change. Similar to the genome-wide comparison, we found a highly significant non-parametric correlation coefficient (Spearman's r = 0.62, P = 1.2e-132) for the expression change of these shared differentially expressed genes (Fig 5A). When comparing the PAXBFW cis changes of these genes to the CERCFW cis changes, however, we found a much lower (though still significant) correlation coefficient (Spearman's r = 0.13, P = 5.1e-6) (Fig 5B). We calculated trans changes for each of these differentially expressed genes, defined as the difference between the expression change in the freshwater parent relative to marine and the freshwater allele relative to the marine in the F1 hybrid [18,19,60]. When comparing the calculated trans changes for these shared differentially expressed genes, we observed much higher correlation coefficient (Spearman's r = 0.51, P = 1.2e-80) (Fig 5C). When comparing all, not just differentially expressed, genes, trans changes are still likely to be more shared than cis (S7 Fig). Additionally, 35/38 of the shared differentially expressed putative dental genes have shared regulatory increases or decreases in both freshwater populations relative to marine in overall expression difference. 32/38 of these gene show regulatory changes in the same direction in trans, but only 25/38 in cis (Fig 5G–5I). Thus, the trans effects on evolved gene expression are more likely to be shared by both freshwater populations than the cis changes.
We sought to test the relative contribution of cis and trans gene regulatory changes during convergent evolution of tooth gain, as well as to ask whether the same or different regulatory changes underlie evolved changes in gene expression during this case of convergent evolution. We quantified the overall regulatory divergence, as well as the specific contribution of cis and trans changes, between ancestral low-toothed marine and two different independently derived populations of high-toothed freshwater sticklebacks. Similar overall changes in gene expression have evolved in both freshwater populations, especially in orthologs of known dental regulators in mammals. In this system, trans-regulatory changes play a larger role than cis changes in both populations. Furthermore, trans acting changes were much more likely to be shared between freshwater populations than cis changes, suggesting the two high-toothed populations evolved their similar gene expression patterns through independent genetic changes.
Convergent evolution at the gene expression level occurs when similar gene expression levels evolve in different populations. Both the PAXBFW and CERCFW stickleback populations have adapted from an ancestral marine form to their current freshwater environments. The genomic nature of their derived changes appears largely divergent, with major axis of variation separating PAXBFW genomes from the geographically proximal marine populations (LITCM), as well as the more distant marine (RABSM) and CERCFW populations. However, when looking at the gene expression basis of their convergently evolved gain in tooth number, orthologs of genes implicated in mammalian dental development showed strong correlated freshwater gains in expression. This correlation suggests both that sticklebacks deploy conserved genetic circuits regulating tooth formation during tooth replacement, but also that both populations have convergently evolved changes to similar downstream transcriptional circuits resulting in a gain of tooth number.
Though both freshwater populations showed strongly correlated changes in evolved gene expression at the trans regulatory level, the cis changes were largely not shared across populations. This was especially true for putative dentally expressed genes with evolved expression changes–the vast majority of the trans but not cis expression changes were shared between both freshwater populations. This suggests that the similar freshwater gene expression patterns evolved through independent genetic changes. It is possible that the small number of shared cis changes are sufficient to drive the observed changes to the overall trans regulatory environments. However previous work has shown that the genetic basis of tooth gain in these two populations is largely distinct [30], and it seems parsimonious that the genetic basis of a gain in dental gene expression is also mostly independent. Thus, convergent freshwater gene expression changes appear to be largely due to distinct, independent population-specific regulatory changes. This finding suggests that there are many regulatory alleles that are accessible during the evolution of an adaptive trait.
Other studies have used RNA-seq to compare the relative contribution of cis and trans-regulatory changes in the evolution of gene expression in a multitude of species and tissues. In mice, evolved gene expression changes in the liver [18] and the retina [61] were driven primarily by cis-regulatory changes. In Drosophila, work on organismal-wide evolved gene expression changes on the genome-wide level has shown the opposite, with trans-regulatory effects playing a larger role in the evolution of gene expression [19,22]. Other studies have found trans effects contribute more to intraspecific comparisons, while cis effects contribute more to interspecific comparisons [17,20,60]. Consistent with this, we observe trans effects dominating in both of our intraspecific comparisons.
Another key distinction could be that cis-regulatory effects dominate when looking at more cellularly homogenous tissues, while trans-regulatory effects dominate when looking at more heterogeneous tissues. Stickleback tooth plates likely fall into an intermediate category, less heterogenous in cell type composition than a full adult fly or fly head, but more heterogeneous than a specialized tissue such as the mouse retina. Overall, freshwater tooth plates are more morphologically similar to each other than marine, with freshwater tooth plates possessing a larger area, increased tooth number, and decreased intertooth spacing [30,41]. Freshwater tooth plates likely have more similar cell type abundances and compositions (e.g. more developing tooth germs with inner and outer dental epithelia, and odontogenic mesenchyme) compared to each other than to marine tooth plates. Similar cell types tend to have similar gene expression patterns, even when compared across different species [62]. Much of the shared freshwater increase in dental gene expression could be due to an increase in dental cell types in both freshwater populations. As other evolved changes to stickleback morphology have been shown to be due to cis regulatory changes to key developmental regulatory genes [8,33,41,63], this trans regulatory increase in cell type abundance could be due to a small number of cis regulatory changes. These initially evolved developmental regulatory changes could result in similar downstream changes in the developmental landscape, resulting in the shared increase in dental cell types. Consistent with this interpretation, stickleback orthologs of genes known to be expressed during mammalian tooth development were found here to have a much greater incidence of convergently evolved increase in trans regulatory gene expression.
Previous studies [17,18] have shown compensatory cis and trans changes are essential for the evolution of gene expression. These findings are consistent with the idea that the main driving force in the evolution of gene expression is stabilizing selection [59] where compensatory changes to regulatory elements are selected for to maintain optimal gene expression levels. In both PAXBFW and CERCFW dental tissue, when considering all genes with a quantifiable (i.e. polymorphic and covered by ~20 reads, see Methods) cis effects, discordant compensatory cis and trans changes were far more common than concordant ones. This trend could be driven by some initial selection on pleiotropic trans changes, followed by selection for compensatory cis changes to restore optimal gene expression levels [17,18,22]. However, the trans, but not the cis, evolved changes in gene expression were highly shared among the two freshwater populations. Thus, collectively our data support a model where two independently derived populations have convergently evolved both similar genome-wide expression levels as well as ecologically relevant morphological changes through different genetic means.
PAXBFW and CERCFW sticklebacks have an increased rate of new tooth formation in adults relative to their marine ancestors [30]. In constantly replacing polyphyodonts, it has been proposed that teeth are replaced through a dental stem cell intermediate [37,38]. A strong candidate gene underlying a large effect PAXBFW tooth quantitative trait locus (QTL) is the secreted ligand Bone Morphogenetic Protein 6 (Bmp6) [41], which is also a key regulator of stem cells in the mouse hair follicle [56]. Freshwater dental tissue displayed significantly increased expression of known signature genes of mouse hair follicle stem cells, perhaps reflecting more stem cell niches supporting the higher tooth numbers in freshwater fish. Genes upregulated in freshwater dental tissue also were significantly enriched for GO terms involved in the cell cycle and cell proliferation. Together these findings suggest that both freshwater populations have evolved an increased tooth replacement rate through an increased activity or abundance of their dental stem cells, and also suggest the genetic circuitry regulating mammalian hair and fish tooth replacement might share an ancient, underlying core gene regulatory network.
Experiments were approved by the Institutional Animal Care and Use Committee of the University of California-Berkeley (protocol # R330).
Fish from all populations were raised in 110L aquaria in brackish water (3.5g/L Instant Ocean salt, 0.217mL/L 10% sodium bicarbonate) at 18°C in 8 hours of light per day. Young fry [standard length (SL) < 10 millimeters (mm)] were fed a diet of live Artemia, early juveniles (SL ~10–20 mm) a combination of live Artemia and frozen Daphnia, and older juveniles (SL > ~20 mm) and adults a combination of frozen bloodworms and Mysis shrimp.
Sticklebacks were fixed in 10% neutral buffered formalin overnight at 4°C. Fish were washed once with water and then stained in 1% KOH, 0.008% Alizarin Red for 24 hours. Following a water rinse, fish were cleared in 0.25% KOH, 50% glycerol for 2–3 weeks. Branchial skeletons were dissected as previously described [64]. Pharyngeal teeth were quantified with fluorescent illumination using a TX2 filter on a Leica DM2500 microscope. Representative tooth plates were created using montage z-stacks on a Leica M165 FC using the RhodB filter. Adult fish were imaged using a Canon Powershot S95. Some tooth count data from the CERCFW, RABSM, and PAXBFW populations; n = 11, 13, 29, respectively, (see S1 Table) have been previously published [30].
Caudal fin tissue was placed into 600μl tail digestion buffer [10mM Tris pH 8.0, 100mM NaCl, 10mM EDTA, 0.05% SDS, 2.5μl ProK (Ambion AM2546)] for 12 hours at 55°C. Following addition of 600 μl of 1:1 phenol:chloroform solution and an aqueous extraction, DNA was precipitated with the addition of 1ml 100% ethanol, centrifuged, washed with 75% ethanol, and resuspended in water. 50ng of purified genomic DNA was used as input for the Nextera Library prep kit (Illumina FC-121-1031), and barcoded libraries were constructed following the manufacturer’s instructions. Library quality was verified using an Agilent Bioanalyzer. Libraries were pooled and sequenced on an Illumina HiSeq 2000 (see S2 Table for details), resulting in a mean of 52.8 million reads per sample, with a max of 70.3 million reads and a minimum of 39 million reads (S2 Table).
Late juvenile stage female sticklebacks (SL ~40mm) were euthanized in 0.04% Tricaine. Dissected [64] bilateral ventral pharyngeal tooth plates were placed into 500μl TRI reagent, then incubated at room temperature for 5 minutes. Following addition of 100μl of chloroform, a further 10 minute incubation and centrifugation, the aqueous layer was extracted. Following addition of 250μl isopropyl alcohol and 10 minute incubation, RNA was precipitated by centrifugation, washed with 75% EtOH, and dissolved in 30ul of DEPC-treated water. RNA integrity was assayed by an Agilent Bioanalyzer. 500ng of RNA from each fish was used as input to the Illumina stranded TruSeq polyA RNA kit (Illumina RS-122-2001), and libraries were constructed following the manufacturer’s instructions. Library quality was analyzed on an Agilent Bioanalyzer, and libraries were pooled and sequenced on an Illumina HiSeq2000 (see S3 Table). We obtained a mean of 84.1 million reads among the parental samples, with a max of 91.0 million and a minimum of 78.6 million (S3 Table).
RNA-seq reads were mapped to the stickleback reference genome [31] using the STAR aligner [65] (version 2.3, parameters = —alignIntronMax 100000—alignMatesGapMax 200000—outFilterMultimapNmax 20—outFilterMismatchNmax 999—outFilterMismatchNoverLmax 0.04—outFilterType BySJout), using ENSEMBL genes release 85 as a reference transcriptome. The resulting SAM files were sorted and indexed using Samtools version 0.1.18 [66], PCR duplicates were removed, read groups added and mate pair information fixed using Picard tools (version 1.51) (http://broadinstitute.github.io/picard/) with default settings. Gene expression was quantified with the Cufflinks suite (v 2.2.1) [47,53–55] using ENSEMBL genes as a reference transcriptome, with gene expression quantified with cuffquant (-u—library-type fr-firststrand) and normalized with cuffnorm. Differentially expressed genes were found using cuffdiff2, with parameters (-u—FDR .1—library-type fr-firststrand, using the reference genome for bias correction). Genes with a mean expression less than 0.1 FPKM were filtered from further analysis.
The BiteCode gene set was generated by combining all genes in the BiteIt (http://bite-it.helsinki.fi/) or ToothCODE (http://compbio.med.harvard.edu/ToothCODE/) [36] databases. Stickleback orthologs or co-orthologs were found using the annotated names of ENSEMBL stickleback genes. Gene set expression change statistical enrichment was done as previously described [67]. Briefly, a t-test was performed for each gene to test for a difference in mean expression between the two treatments. The resulting t-values were subject to a 1-sample t-test, with the null model that the mean of the t-values was 0. Cutoffs were validated using 10,000 bootstrapped replicate gene sets drawn from the same gene expression matrix. Stickleback orthologs of mouse or human genes were determined using annotated ENSEMBL orthologs. Sorted lists of genes, ranked by log2 expression change in PAXBFW dental tissue relative to marine, CERCFW relative to marine, or the mean of CERCFW and PAXBFW relative to marine, were generated using the measured gene expression data. Gene Ontology enrichment was done using Gorilla [68,69], and results were visualized using REVIGO [70].
Genomic resequencing reads were aligned to the stickleback reference genome [31] using the bwa aln and bwa sampe modules of the Burrows-Wheeler Alignment tool (v 0.6.0-r85) [71]. Resulting SAM files were converted to BAM files, sorted and indexed by Samtools version 0.1.18 [66], with PCR duplicates removed by Picard tools. GATK's (v3.2–2) IndelRealigner (parameter: '-LOD 0.4'), BaseRecalibrator, and PrintReads were used on the resulting BAM files. BAM files from the above RNA-seq alignment were readied for genotype calling using GATK's SplitNCigarReads, BaseRecalibrator, and PrintReads. Finally, the UnifiedGenotyper was used to call variants from the RNA-seq and DNA-seq BAM files, with parameters (-stand_call_conf 30 -stand_emit_conf 30 -U ALLOW_N_CIGAR_READS—genotype_likelihoods_model BOTH) [43,45]. This analysis identified a set of 8,341,326 variants.
Principal components analysis of the genome-wide set of variants was performed by first filtering all multiallelic variants or variants with a missing genotype, resulting in a set of 1,690,729 variants. PCA was performed using FactoMiner [46] and a set of custom R scripts. Phylogenetic trees were constructed using the set of variants, downsampled to 67,507 SNPs (no indels) for use with BEAST and SNAPP [72,73]. We constructed phylogenies using SNAPP, estimating substitution rate and proportion invariant from the data, and ran 1 million generations of MCMC simulations. The best tree was picked with TreeAnnotator and visualized with FigTree.
To accurately phase RNA-seq data from F1 hybrids, pseudo-transcriptomes were created for each hybrid. The pseudo-transcriptomes consist of the predicted sequence for each allele within an F1 hybrid, with all predicted splicing variants of a gene collapsed to a single transcript. A variant was added to the pseudo-transcriptome if and only if it was homozygous in the sequenced parents (or parent’s sibling in the case of the RABSM parent of the CERCFW x RABSM F1 hybrids) and called heterozygous in the F1 hybrid.
RNA-seq reads from F1 hybrid sticklebacks were aligned to the individual’s pseudo-transcriptome using STAR (v 2.3) with the parameters:—outFilterMultimapNmax 1 and—outFilterMultimapScoreRange 1. By only looking at uniquely aligning reads, we ensured we only considered reads which overlapped a heterozygous variant site. Counting these unique reads minimizes double counting a single read that supports two different variant positions. Total cis divergence in each F1 hybrid was quantified by comparing the number of reads mapping uniquely to each allele in the pseudo-transcriptome.
Following cis divergence quantification in all F1 hybrids, we considered the overall cis change in the different freshwater populations. Genes which only had 20 or fewer uniquely mapping reads across all replicates were filtered from further analysis. We filtered 28 genes that had >32 fold expression changes that included genes that either had zero reads from one allele and thus infinite expression differences (20 genes), were highly repetitive (2 genes), or mitochondrial (2 genes). Reported cis ratios were calculated by comparing the ratio of uniquely mapped freshwater reads to uniquely mapped marine reads. Evolved trans changes were quantified as the difference between the log of the overall gene expression change between the freshwater and marine parents and the log of measured cis freshwater expression change. Percent cis change was calculated as the absolute value of the log of the cis change divided by the sum of the absolute value of the log of the cis change and the absolute value of the log of the trans change. Statistical significance of cis changes was determined by a binomial test comparing overall reads mapping to the freshwater allele to a null model of no cis divergence, with a false discovery rate of 1% applied using the Benjamini-Hochberg method. Statistical significance of trans changes was determined by a G-test, comparing the expected (based on the measured cis change) and observed ratios of marine and freshwater, with a 1% false discovery rate.
|
10.1371/journal.pgen.1003028 | Genomic Study of RNA Polymerase II and III SNAPc-Bound Promoters Reveals a Gene Transcribed by Both Enzymes and a Broad Use of Common Activators | SNAPc is one of a few basal transcription factors used by both RNA polymerase (pol) II and pol III. To define the set of active SNAPc-dependent promoters in human cells, we have localized genome-wide four SNAPc subunits, GTF2B (TFIIB), BRF2, pol II, and pol III. Among some seventy loci occupied by SNAPc and other factors, including pol II snRNA genes, pol III genes with type 3 promoters, and a few un-annotated loci, most are primarily occupied by either pol II and GTF2B, or pol III and BRF2. A notable exception is the RPPH1 gene, which is occupied by significant amounts of both polymerases. We show that the large majority of SNAPc-dependent promoters recruit POU2F1 and/or ZNF143 on their enhancer region, and a subset also recruits GABP, a factor newly implicated in SNAPc-dependent transcription. These activators associate with pol II and III promoters in G1 slightly before the polymerase, and ZNF143 is required for efficient transcription initiation complex assembly. The results characterize a set of genes with unique properties and establish that polymerase specificity is not absolute in vivo.
| SNAPc-dependent promoters are unique among cellular promoters in being very similar to each other, even though some of them recruit RNA polymerase II and others RNA polymerase III. We have examined all SNAPc-bound promoters present in the human genome. We find a surprisingly small number of them, some 70 promoters. Among these, the large majority is bound by either RNA polymerase II or RNA polymerase III, as expected, but one gene hitherto considered an RNA polymerase III gene is also occupied by significant levels of RNA polymerase II. Both RNA polymerase II and RNA polymerase III SNAPc-dependent promoters use a largely overlapping set of a few transcription activators, including GABP, a novel factor implicated in snRNA gene transcription.
| The human pol II snRNA genes and type 3 pol III genes have the particularity of containing highly similar promoters, composed of a distal sequence element (DSE) that enhances transcription and a proximal sequence element (PSE) required for basal transcription. In pol II snRNA promoters, the PSE is the sole essential core promoter element whereas in type 3 pol III promoters, there is in addition a TATA box, which determines RNA pol III specificity [1], [2]. The PSE recruits the five-subunit complex SNAPc, one of the few basal factors involved in both pol II and pol III transcription. Basal transcription from pol II snRNA promoters requires, in addition, TBP, TFIIA, GTF2B (TFIIB), TFIIF, and TFIIE, and from pol III type 3 promoters TBP, BDP1, and a specialized GTF2B-related factor known as BRF2 [3], [4], [5]. The DSE is often composed of an octamer and a ZNF143 motif (Z-motif) that recruit the factors POU2F1 (Oct-1) and ZNF143 (hStaf), respectively [1], [2]. POU2F1 activates transcription in part by binding cooperatively with SNAPc and thus stabilizing the transcription initiation complex on the DNA (see [6], and references therein).
In addition to requiring some different basal transcription factors for transcription initiation, pol II and pol III transcription at SNAPc-recruiting promoters differ in the way transcription terminates. In pol III genes, there are runs of T residues at various distances downstream of the RNA-coding sequence, which direct transcription termination ([7] and references therein). In pol II snRNA genes, a “3′ box” starting generally 5–20 base pairs downstream of the RNA coding sequence directs processing of the RNA, with transcription termination reported to occur either just downstream of the 3′ box [8], or over a region of several hundreds of base pairs [9].
Although model snRNA promoters have been extensively studied, it is unclear how broadly SNAPc is used, and to what extent the highly similar pol II and pol III PSE-containing promoters are selective in their recruitment of the polymerase. It is also unclear how generally the use of the basal factor SNAPc is coupled to that of the activators POU2F1 and ZNF143, and by which mechanisms ZNF143 activates transcription. To address these questions, we performed genome-wide immunoprecipitations followed by deep sequencing (ChIP-seq) to localize four of the five SNAPc subunits, GTF2B, BRF2, and a subunit of each pol II and pol III. These studies define a set of SNAPc-dependent transcription units and show that although most loci are primarily bound by one or the other polymerase, the RPPH1 (RNase P RNA) gene is occupied by both enzymes. Pol II is detectable up to 1.2 kb downstream of the end of the RNA-coding regions of pol II snRNA genes, thus defining a broad region of transcription termination. Localization of POU2F1 and ZNF143 shows widespread usage of these activators by PSE-containing promoters, and we find that several of these promoters also bind the activator GABP [10], which has not been implicated in snRNA gene transcription before. Activators are recruited before the polymerase in G1, and this process is less efficient when ZNF143 levels are decreased by RNAi.
We performed ChIP-seq with antibodies against SNAPC4 (SNAPC190), the largest SNAPc subunit, SNAPC1 (SNAP43), and SNAPC5 (SNAP19) in IMR90Tert cells. To localize SNAPC2 (SNAP45), we used an IMR90Tert cell line expressing both biotin ligase and SNAPC2 tagged with the biotin acceptor domain for chromatin affinity purification (ChAP)-seq (see [11]). We also used antibodies against GTF2B, which should mark pol II snRNA promoters, BRF2, which should mark type 3 pol III promoters, and POLR2B (RPB2), the second largest subunit of pol II. We used POLR3D (RPC4) ChIP-seq data [11] to localize pol III.
Most of the human pol II snRNA and type 3 pol III genes are repeated and/or have given rise to large amounts of related sequences within the genome. We therefore aligned tags as described before [11], excluding tags aligning with one or more mismatches but including tags with several perfect matches in the genome (see Methods). We selected regions containing at least two SNAPc subunits and either BRF2 and pol III, or GTF2B and pol II, as described in Methods. We obtained loci encompassing all known type 3 pol III genes as well as most annotated pol II snRNA genes. In addition, we obtained a few novel loci occupied by SNAPc and pol II. Table S1 shows these loci as well as the annotated snRNA genes that did not display any tags, namely four RNU1 and one RNU2 snRNA genes (in red in the first column). It also shows, in grey, RNU2 genes that are still in the “chr17_random” file of the human assembly and were thus not in the reference genome used for tag alignment.
In some cases, we noticed adjacent POLR2B peaks separated by only one or a few nucleotides, which often corresponded to annotated SNP positions. Inclusion of tags aligned with ELAND, which allows for some mismatches, often resulted in the fusion of adjacent peaks, as for the SNORD13 gene shown in Figure S1A (compare upper and lower panels). Such loci are likely to be occupied by POLR2B –indeed their promoter regions are occupied by significant amounts of GTF2B and SNAPc subunits– and they are labeled in yellow in the first column of Table S1. In a few cases, however, this did not result in fusions of adjacent peaks, as shown in Figure S1B for a RNU1 gene (U1-12). Such peaks probably result from attribution of tags with multiple genomic matches to an incorrect genomic location and are thus likely to be artifacts. Consistent with this possibility, U1-11, U1-12, U1-like-8, U3-2, U3-2b, U3-4, and U3-3, all labeled in orange in Table S1, had POLR2B, GTF2B, and SNAPc subunits scores with either 0% or, in the cases of U3-4, less than 15%, unique tags. We consider these loci unlikely to be occupied by pol II in vivo. In contrast, the POLR2B peak on the RNU2 snRNA gene on chromosome (chr) 11, even though interrupted about 500 base pairs downstream of the snRNA coding region, is constituted mostly of unique tags, as are the GTF2B and SNAPc subunit peaks. This gene is likely, therefore, to be indeed occupied by pol II and other factors, and is labeled in striped yellow in the first column (Table S1).
We calculated occupancy scores for all loci by adding tags covering peak regions, as described in Methods (see legend to Table S1 for exact regions). We first examined the POLR2B, POLR3D, GTF2B, and BRF2 scores. For most genes there was a clear dominance of either POLR2B and GTF2B or POLR3D and BRF2 (Figure 1A). Further, there was a good correlation between POLR2B and GTF2B (0.89) or POLR3D and BRF2 (0.80) scores, but not between POLR2B and BRF2 (0.075), or POLR3D and GTF2B (0.22) (Figure S2). This is consistent with GTF2B and BRF2 being specifically dedicated to recruitment of pol II and pol III, respectively, and indicates that most SNAPc-occupied genes are transcribed primarily by a single polymerase.
Strikingly, among SNAPc-occupied promoters, only thirteen loci were occupied primarily by BRF2 and pol III (listed on top of Table S1), corresponding to the known type 3 genes previously shown to be occupied by pol III in IMR90hTert and other cell lines [11], [12], [13], [14]. We identified a larger number of SNAPc-bound loci occupied primarily by GTF2B and pol II. They included genes coding for the U1, U2, U4 and U5 snRNAs, all involved in splicing of pre-mRNAs; U11, U12, and U4atac snRNAs, which have similar functions as U1, U2, and U4 but participate in the removal of a smaller class of introns referred to as AT-AC introns; U7 snRNA, involved in the maturation of histone pre-mRNAs; U3, U8, and U13 small nucleolar RNAs (snoRNAs), involved in the maturation of pre-ribosomal RNA, as well as snRNA-derived sequences. The relationship of these loci with previously described snRNAs and snoRNA genes is described in the Results section of Text S1. We also uncovered a few non-annotated loci harboring SNAPc subunits, as well as GTF2B and POLR2B, peaks constituted by at least 20% of unique tags and, therefore, likely to correspond to new actively transcribed regions. These are labeled Unknown-1 to 7 (rows 76–82 in Table S1). As described below, these sequences harbor a PSE as well as some other sequence elements typical of pol II snRNA promoters, and contain similarities to the 3′ box.
Although most genes were occupied mostly by either BRF2 and POLR3D, or GTF2B, and POLR2B, there were a few exceptions. The most notable was the RPPH1 gene, which is considered a type 3 pol III gene [15] but was in fact occupied not only by BRF2 and POLR3D but also by significant amounts of POLR2B and GTF2B, comparable to those found on the RNU4 snRNA genes (Figure 1A and 1B). This suggested that this gene could be transcribed in vivo by either of two RNA polymerases, pol II or pol III. To explore this possibility further, we treated cells with a concentration of α-amanitin known to inhibit pol II but not pol III transcription [16]. As expected, this treatment reduced the POLR2B signal of the pol II RNU2 gene but not the POLR3D signal on the pol III hsa-mi-886 gene (Figure 1C, upper panels). To determine the effects of α-amanitin for the RPPH1 gene and the U6-2 gene, which also displayed some POLR2B signal in addition to the expected POLR3D signal (see Figure 1A), we set the POLR2B and POLR3D signals obtained in the absence of α-amanitin at 1. In each case, addition of α-amanitin to the medium reduced the POLR2B but not the POLR3D signal (Figure 1C, lower panels). Thus, the RPPH1 gene can be transcribed either by pol II or pol III in vivo.
One of the criteria used to select the genes in Table S1 was the presence of at least two of the four SNAPc subunits examined. We obtained a good correlation between scores for the four SNAPc subunits tested (Figure S3), consistent with SNAPc binding as a single complex to snRNA promoters [17]. Figure 2A shows the peaks obtained for the SNAPc subunits, BRF2, GTF2B, POLR3D, and POLR2B on the pol III TRNAU1 gene and the pol II RNU4ATAC gene, and Figure 2B shows two non-annotated genomic loci occupied by POLR2B, GTF2B, and SNAPc subunits. Whereas the polymerase subunits were detected over the entire RNA coding sequence of the corresponding genes (and further downstream in the case of POLR2B), the other factors were located within the 5′ flanking region, with GTF2B and BRF2 close to, or overlapping, the TSS. Although peaks were sometimes constituted of too few tags to allow an unambiguous determination of the peak summit location (see for example the SNAPC4 peak in Figure 2A), we could nevertheless detect clear trends. The GTF2B or BRF2 peaks were generally the closest to the TSS, the SNAPC4, SNAPC1, and SNAPC5 peaks were within the PSE sequence, and the SNAPC2 peak was upstream of the PSE (Figure 2C).
Figure S4 shows an alignment of the PSEs and TATA boxes of the 14 pol III type 3 promoters (including the RPPH1 gene), and Figure S5 an alignment of the PSEs of all pol II loci listed in Table S1. The non-annotated loci occupied by POLR2B and factors contain clear PSEs. Moreover, as noted previously [1], [2], the PSE is located further upstream of the TSS in pol III than in pol II snRNA genes. The corresponding LOGOs revealed similar but not identical consensus sequences for the PSEs of pol II and pol III genes (Figure 2D); for example, adenines were favored in positions 11 and 12 of pol III, but not pol II, PSEs. Thus, although the TATA box is the dominant element specifying RNA polymerase specificity –indeed the U2 and U6 PSEs can be interchanged with no effect on RNA polymerase recruitment specificity [16]– the exact PSE sequence may also contribute to specific recruitment, for example in the context of a weak TATA box.
The U1 and U2 snRNA genes are followed by a processing signal known as the 3′ box [18], [19], which is also found downstream of several other pol II snRNA genes [1]. We could identify 3′ boxes in most of the pol II genes in Table S1. An alignment of these motifs allowed us to generate a matrix with GLAM2 [20], which we then used to search for 3′ boxes in all pol II with GLAM2SCAN [20]. As shown in Figure S6, we could identify putative 3′ boxes downstream of all annotated pol II genes in Table S1 (except for the non-expressed RNU1 (U1-9) and RNU1 (U1-13) genes), as well as for the non-annotated genes. For the RPPH1 gene, the best match to a 3′ box was located within the RNA coding sequence, from −73 to −61 relative to the end of the RNA coding sequence (Figure S6). The resulting 3′ box LOGO derived from all sequences aligned in Figure S6 is shown in Figure 3A.
Pol II transcription termination has been reported to occur either shortly after, or several hundred base pairs downstream of, the 3′ box [8], [9]. Our POLR2B ChIP-seq data reveal the extent of pol II occupancy downstream of the RNA coding region. Whereas on average, the POLR3D ChIP-seq signal dropped quite abruptly downstream of the RNA coding region of pol III genes (see [7]), POLR2B could be detected as far as about 1200 base pairs past the RNA coding region of pol II snRNA genes (Figure 3B). Moreover, examination of the POLR2B peak downstream of individual pol II genes revealed a gradual decrease of tag counts over regions of 500 or more base pairs (see for example Figure 2A and 2B, and Figure 4A below). Thus, transcription termination occurs well downstream of the 3′ box and over a broad region.
snRNA promoters are characterized by an enhancer element (DSE) typically containing an octamer motif and a ZNF143 binding site (Z-motif), which in some specific genes has been shown to recruit, respectively, the POU domain protein POU2F1 and the zinc finger protein ZNF143 (see [1], [2] and references therein). To determine how general the binding of POU2F1 and ZNF143 is among SNAPc-binding promoters, we localized POU2F1 by ChIP-seq in HeLa cells and we analyzed ChIP-seq data obtained by others in HeLa cells (JM, VP, and Winship Herr, personal communication) for ZNF143 and, as ZNF143 was found to bind often together with GABP (JM, VP, and Winship Herr, personal communication), for the α subunit of GABP (GABPA). The scores for all genes are listed in Table S1 and, in a summarized form, in Table S2. The pol III genes in Table S1, which were all occupied by basal factors (see above), were each occupied by at least one activator. Among pol II genes, those not occupied by basal factors (labeled in red in the first column of Tables S1 and S2) did not display peaks for any of the activators, and those with interrupted POLR2B peaks (orange in the first column) had peaks composed solely of tags with multiple matches in the genome, consistent with the possibility raised above that these genes are, in fact, not occupied by factors.
Of the genes clearly occupied by basal factors, all displayed peaks for at least one activator with three exceptions, U1-like-11, unknown-2, and unknown-3; these last three loci had basal factor peaks with relatively low scores and thus may bind some of these activators at levels too low to be detectable in our analysis. Most genes had a POU2F1 peak (93%), a large majority had a ZNF143peak (81%), and about half had a GABPA peak (45%). Interestingly, some genes had specific combinations of activators; for example the RNU5 and U5-like genes as well as most pol III genes had peaks for both POU2F1 and ZNF143 but not for GABPA. In contrast RNU6ATAC, SNORD13, and RNU3 genes had POU2F1 and GABPA peaks but no ZNF143 peak. Only few genes had only one activator (RMRP, RNY4, RNU2-2, U3b2-like, RNU7, and Unknown-5) suggesting that most snRNA genes require some combination of the three activators tested for efficient transcription. Indeed, altogether 23 genes had peaks for all three factors and 23 had peaks for both ZNF143 and POU2F1 but not GABPA. Thus, the very large majority (79%) of SNAPc-binding genes bound both POU2F1 and ZNF143. The scores for the various activators were surprisingly correlated (see Figure S7), perhaps indicating that these factors bind to snRNA promoters interdependently. Figure 4A shows two examples (RNU4ATAC and U1-like-5) with the three factors present, and two examples (Unknown-6 and tRNAU1) with only POU2F1 and ZNF143. In all cases, the factors bound upstream of the PSE with GABP, when present, generally binding the furthest upstream.
We analyzed 5′ flanking sequences for motifs and identified POU2F1 (octamer, see [21]), ZNF143 [22], [23], and GABP [24], [25], [26] binding sites (Figure 4B, Figure S8A and S8B). This analysis revealed a high concordance between occupancy as determined by ChIP-seq and presence of the corresponding motif, with only a few cases (GABP and ZNF143 for U1-like-10, and GABP for U5E-like, U4-1, and unknown-7 genes) where no convincing motif could be identified. We then aligned all occupied motifs (see Figures S9, S10, and S11) to generate the LOGOs shown in Figure 4C, which thus reflect the ZNF143, POU2F1, and GABP binding sites in SNAPc-recruiting genes.
Transcription of RNU6 and probably RNU1 and RNU2 is known to be low during mitosis and to increase as cells cycle through the G1 phase [27], [28], [29], [30], [31], hence we measured the levels of U1, U2, and U6 snRNA during mitosis and at several times after entry into G1. Since snRNA transcripts are very stable, making it difficult to measure transcription variability, we generated HeLa cell lines containing RNU1 or RNU6 reporter construct expressing unstable transcripts whose levels therefore better reflect ongoing transcription. For U2 snRNA, we measured its precursor, which has a short half-life [16]. Cells were blocked in prometaphase with Nocodazole and released with fresh medium. RNA levels were low during mitosis and, in the case of the U1 reporter RNA and pre-U2 RNA, increased to a maximum 6–7 h after release, around the middle of the G1 phase (as determined by FACS analysis, see Methods). For the U6 reporter RNA, RNA levels reached a maximum 3 h after release, at the beginning of the G1 phase (Figure 5A). POLR2B occupancy was apparent 4 h after the mitosis release and peaked after 6 h, as measured by ChIP-qPCR analysis of both RNU1 and RNU2 loci (Figure 5B). This was specific, as no significant amounts of POL2RB were detected on the control region. In comparison, increased POLR3D occupancy of RNU6 (but not the control region) was apparent 3 h after release and peaked after 6 h, consistent with the accumulation of U6 RNA earlier in G1 than U1 and U2 RNA.
We then examined promoter occupancy by transcription activators (Figure 5B). ZNF143 occupancy increased over time on both the RNU1 and RNU6 promoters, becoming clearly detectable at 3 h and reaching a maximum at 6 h for RNU1 and 4 h for RNU6. In contrast, ZNF143 was undetectable on the RNU2 promoters. POU2F became detectable at 3 h on the RNU1, RNU2, and RNU6 promoters and then remained at a more or less constant level. GABP was detected only on the RNU1 promoters and was recruited early, starting 2 h after the release and reaching a maximum at 5 h. Thus, activators were recruited on the promoters expected from the ChIP-seq data above, with kinetics slightly faster than the polymerase. Among activators, GABP was recruited the earliest, followed by concomitant recruitment of ZNF143 and POU2F1.
Some basal transcription factors such as TBP are thought to remain bound to chromatin, and hence probably promoters, during mitosis [32], [33]. To explore whether this is the case for SNAPc, GTF2B, and BRF2, we monitored occupancy by these factors at mitosis (1 h after release) and in mid-G1 (7 h after release). On the pol II RNU1 snRNA promoter, we observed enrichment of GTF2B and SNAPc subunits, as well as the pol II subunit POLR2B, the activators ZNF143, POU2F1, and GABP, and H3 acetylated on lysine 18 (H3K18Ac) at mid-G1 compared to mitosis (Figure 5C, upper panel). This was specific as the pol III subunit POLR3D was not enriched. On the pol III RNU6 promoter, we observed enrichment of POLR3D, BRF2, SNAPc subunits, ZNF143, POU2F1 and H3K18Ac, but not POLR2B nor GABP, as expected (Figure 5C, lower panel). This suggests that at snRNA promoters, both basal transcription factors and activators are removed from promoter DNA during mitosis and are recruited de novo upon transcription activation in G1.
To explore the role of ZNF143 in transcription factor recruitment, we targeted endogenous ZNF143 by siRNA and synchronized the cells as above. Total protein levels measured both at mitosis and in mid-G1 were reduced by more than 70% (Figure 6A), and in mid-G1, ZNF143 bound to the U1 promoter was decreased by 50% (Figure 5B). Under these conditions, binding of the activators POU2F1 and GABP, the basal transcription factors GTF2B and SNAPC1, and POL2RB were reduced by 40 to 70%. In contrast, the H3K18Ac levels were not reduced (Figure 6B). Thus, ZNF143 contributes to efficient recruitment of other activators, basal transcription factors, and the RNA polymerase, but not to H3K18 acetylation, at the pol II U1 promoter.
Using stringent criteria of co-occupancy by two SNAPc subunits and either GTF2B and pol II, or BRF2 and pol III, we identified a surprisingly small number of SNAPc-occupied promoters comprising the 14 known type 3 pol III promoters, some 40 pol II snRNA genes, and 7 novel pol II-occupied loci. It seems, therefore, that in cultured cells, SNAPc is a very specialized factor participating in the assembly of transcription initiation complexes at fewer than 100 promoters. We have not explored, however, the possibility that some of the SNAPc subunits participate in transcription of other genes or in other functions as part of complexes other than SNAPc. Indeed, in a previous localization of SNAPc subunits on genomic sites also binding TBP, a correlation analysis on non-CpG islands split the SNAPc subunits into two subgroups, one containing SNAPC1 and SNAPC5 and the other SNAPC2, SNAPC3, and SNAPC4 [34], consistent with the possibility that other SNAP -subunit-containing complexes exist.
A peculiarity of SNAPc is its involvement in transcription from both pol II and pol III promoters, promoters that differ from each other mainly by the presence or absence of a TATA box. We found that most SNAPc-occupied promoters were predominantly occupied by either pol II or pol III with two exceptions, the U6-2 and most notably the RPPH1 genes, which were occupied not only by BRF2 and pol III, as expected, but also by levels of GTF2B and pol II comparable, in the second case, to those found on some pol II snRNA genes. We showed that pol II occupancy of the RPPH1 gene was obliterated by levels of α-amanitin shown before to inhibit pol II transcription in cultured cells [16]. Previous experiments comparing the 3′ ends of pol II and pol III transcripts derived from wild-type and mutated versions of the human RNU2 and RNU6 promoters have shown that pol II-synthesized transcripts end downstream of a signal referred to as the “3′ box” whereas pol III-synthesized transcripts are not processed at such boxes and instead end at runs of T residues [16]. The best similarity to a 3′ box lies within the RPPH1 RNA coding region. However, we detect only one type of transcript, terminated at the run of T residues downstream of the RPPH1 gene, in endogenous RNA from proliferating IMR90Tert cells (data not shown), suggesting that the transcript synthesized by pol II is highly unstable, at least under the conditions tested. It is conceivable that the ratio of RPPH1 genes transcribed by pol II and pol III, as well as the ratio of stable pol II and pol III RNA products, change in different cell types or under different conditions. The observation that a gene can be transcribed by two different polymerase in vivo thus raises the possibility of an added layer of complexity in the regulation of gene expression. It is not clear why the U6-2 and RPPH1 promoters are capable of recruiting significant levels of pol II. The RPPH1 promoter has a short TATA box, but the U6-7 and U6-8 promoters have the same TATA box and are not promiscuous. An intriguing possibility is that the presence of a 3′ box at a correct distance downstream of the TSS, together with a weak TATA box, allow pol II recruitment.
The locations of the occupancy peaks for the four SNAPc subunits we tested are remarkably consistent with what is known about the architecture and DNA binding of SNAPc. SNAPC4, the largest SNAPc subunit and the backbone of the complex, binds directly to the PSE through Myb repeats located in the N-terminal half of the protein [35]. SNAPC1 and SNAPC5 associate directly with SNAPC4, N-terminal of the Myb repeats (aa 84–133, see [36]). Consistent with this architecture, we find that SNAPC4, SNAPC1, and SNAPC5 generally peak very close to each other within the PSE. In contrast, SNAPC2, which associates with the C-terminal part of SNAPC4 (aa 1281–1393, see [36]), peaks upstream of the PSE. This suggests that the N-terminus of SNAPC4 is oriented facing the transcription start site whereas the C-terminal part is oriented towards the upstream promoter region. This is consistent with the orientation of D. melanogaster SNAPC4 [37] on the U1 and U6 D. melanogaster snRNA promoters as determined by elegant studies combining site-specific protein-DNA crosslinking with site-specific chemical protein cleavage ([38], see also [39] and references therein).
The 3′ end of pol II snRNAs is generated by processing at a sequence called the 3′ box [2], [40]. The 3′ box is efficiently used only by transcription complexes derived from snRNA promoters, suggesting that the polymerase II recruited on these promoters is somehow different from that recruited on mRNA promoters. Indeed, the C-terminal domain of pol II associated with snRNA genes carries a unique serine 7 phosphorylation mark, which recruits RPAP2, a serine 5 phosphatase, as well as the integrator complex, both of which are required for processing ([41] and references therein; [42], [43]). Moreover, pol II transcription of snRNA genes requires a specialized elongation complex known as the Little Elongation Complex (LEC) [44]. It has been unclear, however, how far downstream of the 3′ box processing signal transcription continues, with one report indicating a very sharp drop in transcription within 60 base pairs past the U1 3′ box [8] and another reporting continued transcription for several hundreds of base pairs downstream of the U2 3′ box [9]. Our ChIP-seq data indicate that pol II can be found associated with the template more than 1 Kb downstream of the 3′ box, for both the RNU1 and RNU2 genes as well as all other pol II snRNA genes. This suggests that transcription termination downstream of snRNA gene 3′ boxes does not occur at a precise location but rather over a broad 1.2 Kb region, and is triggered by passage of the polymerase through the processing signal, reminiscent of transcription termination downstream of the poly A signal, in this case in a region of several Kbs [45].
Activation of several SNAPc-dependent promoters has been shown to depend on a DSE and on the binding of POU2F1 and ZNF143 (see [1], [2] and references therein, [23]). Our ChIP-seq analyses show that POU2F1 and ZNF143 are associated with the large majority of SNAPc-dependent promoters and identify GABP as a new factor binding to a subset of these promoters. During transcription activation in G1, we observed binding of ZNF143 and POU2F1 preceding binding of RNA pol II and pol III, consistent with the possibility that binding of these activators prepares the promoters for polymerase recruitment. Indeed, lowering the amount of ZNF143 by siRNA strongly affected recruitment of POU2F1, GABPA, basal factors, and the polymerase itself on the U1 promoter. Thus, ZNF143 could either recruit and stabilize POU2F1 by direct protein-protein contact, or affect chromatin structure to allow recruitment of POU2F1, or both. In support of the first hypothesis, ZFP143, the mouse homolog of ZNF143, recruits another POU-domain protein, Oct4 (the mouse homolog of POU5F1) by direct association [46]. On the other hand, ZNF143 and POU2F1 do not bind cooperatively to the human U6-1 promoter [47], but then U6-1 is weakly POLR3D-occupied compared to other human RNU6 genes [11]. In support of the second possibility, we have shown before that ZNF143 can bind to an snRNA promoter, in this case the pol III U6 snRNA promoter, preassembled into chromatin [48], suggesting that it is an early player in the establishment of a transcription initiation complex. However, promoter H3K18 acetylation, which is low just after mitosis and increases during G1, was unaffected. This suggests that SNAPc-dependent promoters are targeted very early in G1 by as yet unidentified factors that lead to histone modifications, in particular H3K18 acetylation. It will be interesting to determine how this modification combines with the H3K4me3 mark observed on pol III promoters, including type 3 pol III promoters [12], [13], [14], [49].
ChIPs were performed as described [11]. The antibodies used (rabbit polyclonal antibodies except where indicated) were as follows: POLR3D, CS682, directed against the C-terminal 14 aa [50]; POLR2B, H-201 from Santa Cruz Biotechnology; BRF2, 940.505 #74; GTF2B, CS369 #10, 11; SNAPC4, CS696 #4,5; SNAPC5, CS539 #7,8; SNAPC1, CS47 #7,8; GABP, sc-22810 X from Santa Cruz Biotechnology; POU2F1, mix of YL8 and YL15 [51], [52] or mix of two polyclonal antibodies (A310-610A from Bethyl Laboratories); ZNF143, antibody 19164 raised against ZNF143 aa 623–638, [48]. The ChAPs have been described [11].
The sequence tags obtained after ultra-high throughput sequencing were mapped onto the UCSC genome version Hg18, corresponding to NCBI 36.2, as before [11] except that we included tags mapping to up to 500 rather than 1000 different locations in the genome. Table S3 shows the total number of tags sequenced for each ChIP and the percentages of tags mapped onto the genome. In all cases, 75.5% or more of the total tags mapped onto the genome had unique genomic matches.
Peaks were detected with sissrs (www.rajajothi.com/sissrs/) [53] with a false discovery rate set at 0.001%, as previously described [11]. We identified 77312 POLR2B, 4838 GTF2B, 1366 POLR3D, and 2526 BRF2 peaks. We then selected the POLR2B peaks within 100 base pairs of a GTF2B peak (3878 peaks), and the POLR3D peaks within 100 base pairs of a BRF2 peak (125 peaks). The ChIPs with the anti-SNAPc subunit antibodies gave relatively weak signals. We therefore divided the genome into 200 nucleotide bins, counted tags obtained for each of the four SNAPc subunits analyzed, and retained only bins displaying an enrichment for at least two of the SNAPc subunits. Bins were considered positive only if the tag number in bin reached at least the minimum tag count determined by sissrs for enriched regions with a 0.001 false discovery rate as the one used in sissrs set at the default parameters. We then considered genomic regions containing POLR2B and GTF2B, or POLR3D and BRF2, sissrs peaks as well as a bin positive for two SNAPc subunits within 100 nucleotides of the polymerase sissrs peak. We obtained 157 and 58 loci for the POLR2B and POLR3D lists, respectively, which were all visually inspected. We eliminated peaks in regions of high background, with shapes never found in known snRNA genes (for example peaks with rectangular shapes resulting from artefactual accumulation of tags), or with identical shape and location in all samples. The most convincingly occupied loci are listed in Table S1, which also shows all annotated pol II snRNA genes, whether or not they were found occupied by POLR2B, GTF2B, and SNAPc subunits. Scores were calculated as described in [49] and contained a component consisting of the sum of tags with unique matches in the genome and another representing tags with multiple matches in the genome: such tags were attributed a weight corresponding to the number of times they were sequenced divided by the number of matches in the genome, with a maximum weight set at 1. In Table S1, the score percentage contributed by unique tags is indicated in separate columns. Scores and peak shapes are more reliable for scores consisting mostly of unique tags, as in these cases there is no ambiguity as to where in the genome tags should be aligned.
For the SNAPc subunits, we confirmed the results of the first analysis by performing a second analysis in which we counted tags in 200 nucleotide bins as before, then fitted a normal distribution to the data, and used the normal distribution's standard deviation and mean to attribute a P-value for each SNAPc subunit to each genomic bin. We then adjusted it with Benjamini & Hochberg (BH) correction and kept the bins with an adjusted P-value under 0.005 that were located within a 100 nucleotides of either a RPB2 and TF2B positive region, or a RPC4 and BRF2 positive region (as defined by sissrs). We then applied a second filter to keep only the bins containing at least two (of the four mapped) SNAPc subunits. This gave us a total of 275 bins, which contained all the genes listed in Table S1 except for 10 loci. Of these 10 loci, 5 of them are flagged Table S1 as being not occupied (U1-7, U1-9, U1-10, U1-13, U2-1). The remaining five (U1-like-1, U1-like-11, RNU5 (U5F), UNKNOWN-2, and RNU6-7 (U6-7)) have low scores. The additional regions with positive bins (93 regions) corresponded to regions of high background and were eliminated after visual inspection.
To measure RPPH1-dependent transcription in vivo, 1.2×106 HeLa cells were transiently transfected (48 hours) with pU6/Hae/RA.2 [16] or derivatives containing the wild-type RPPH1 promoter, or the RPPH1 promoter harboring a mutation in the TATA box (TTATAA changed to TCGAGA), as well as the RPPH1 3′ flanking region. To specifically inhibit POLR2B transcription, the cells were treated with 50 µg/ml of α-amanitin (Santa Cruz Biotechnology, sc-202440) for two or six hours before harvesting.
Clonal cell lines expressing U1 or U6-promoter-directed unstable RNA were established by transfection of HeLa cells with plasmid derivatives of pU6/RA.2+U6end-Dsred [48] (see Methods section of Text S1 for details). Individual clones were expanded and tested for expression of the U1 or U6 construct. HeLa cell lines were synchronized as described [54]. Briefly, cells were first incubated for 24 h with 2 mM of Thymidine, then 3 h with normal medium, then 14 h with 0.1 mg/ml of Nocodazole. Cells were then harvested (M phase) or transferred to normal medium and harvested at different time points. The cell cycle stage of each sample was determined by flow cytometry analysis with the UV precise T kit (Partec, Germany), which involves isolation of nuclei followed by DAPI staining.
RNA was extracted from HeLa cells with TRIzol reagent (Invitrogen) according to the manufacturer's protocol and analyzed by RNase T1 protection as before (see Methods section of Text S1 for details). To reduce levels of endogenous ZNF143, a siRNA duplex was generated (Microsynth) to target the ATAAGCTGTGGTACCATCTTCCAGCTG region of the ZNF143 gene. HeLa cells were seeded at 2×106 cells per 10 cm plate the day before transfection. Thirty µl of INTERFERin transfection reagent (Polyplus) was added to 1 ml of DMEM serum-free medium containing 60 nM of siRNA duplex, incubated for 15 minutes, and added to the 10 cm plate containing 10 ml of medium. As negative control, we used a siRNA directed against the firefly luciferase [55] (Dharmacon). Two other siRNA treatments were performed 12 and 24 h after the first transfection. Thirty hours after the 1st transfection, the cells were synchronized as described above.
The data can be accessed at NCBI Gene expression Omnibus (http://www.ncbi.nlm.nih.gov/geo) under accession number GSE38303.
|
10.1371/journal.pcbi.1004245 | Geometrical and Mechanical Properties Control Actin Filament Organization | The different actin structures governing eukaryotic cell shape and movement are not only determined by the properties of the actin filaments and associated proteins, but also by geometrical constraints. We recently demonstrated that limiting nucleation to specific regions was sufficient to obtain actin networks with different organization. To further investigate how spatially constrained actin nucleation determines the emergent actin organization, we performed detailed simulations of the actin filament system using Cytosim. We first calibrated the steric interaction between filaments, by matching, in simulations and experiments, the bundled actin organization observed with a rectangular bar of nucleating factor. We then studied the overall organization of actin filaments generated by more complex pattern geometries used experimentally. We found that the fraction of parallel versus antiparallel bundles is determined by the mechanical properties of actin filament or bundles and the efficiency of nucleation. Thus nucleation geometry, actin filaments local interactions, bundle rigidity, and nucleation efficiency are the key parameters controlling the emergent actin architecture. We finally simulated more complex nucleation patterns and performed the corresponding experiments to confirm the predictive capabilities of the model.
| Many cellular processes or morphological changes, for example division and migration, strongly depend on the cytoskeleton. For successful completion of these processes in higher cells, thousands of cytoskeletal filaments need to be organized into precise higher-order structures. Identifying the most important parameters governing the emergent organization of large systems of filaments is thus a crucial step in understanding cytoskeletal systems. However, due to the complexity of these systems and the limits of imaging, this has proven to be a difficult task. To overcome this limitation we previously established an in vitro assay for actin filament assembly that allows controlling both the chemical components and geometrical boundaries. Here we used detailed computer simulations to rationalize the experimental observations and identify some of the physical principles governing actin organization. Such approach will ultimately allow one to predict actin network formation under various constraints.
| Actin assembles to form higher order structures [1] that are essential to cell morphogenesis, adhesion and motility [2]. A single filament can either resist or generate forces according to its local environment [3,4], but most physiological processes require the assembly of a higher ordered network [5,6]. Therefore, one needs to study the collective behavior of system composed of many thousands of actin filaments to understand their physiological functions. Indeed, whereas one actin filament could not sustain forces higher than a few pN [4], a bundle of actin filaments can resist hundreds of pN [7] and larger structures are able to bear even higher forces (nN range) [8]. In addition to the number of filaments, the architecture of the network is also adapted to achieve different cellular functions. At the edge of the cell, in the lamellipodium, actin filaments form a dense branched network [9–11], that seems optimal to push the membrane forward during actin-based motility [12,13]. Actin filaments are oriented with their elongating ends near the membrane, at an optimum angle of ± 35° with respect to the membrane [14, 15, 16], and being present in high density close to the membrane [17], they can efficiently sustain the protrusive force [13]. Actin filaments can also form bundles of parallel filaments creating finger-like protrusions in the membrane called filopodia [18] that explore the extracellular matrix [19]. Actin bundles can be also used as tracks for protein or cargo transport [20,21]. These bundles can be formed by branched organization that merged into elongated parallel actin filament [22,23]. Finally, actin filaments can form aligned structures of anti-parallel filaments in stress fibers or transverse arcs, that are site of active contraction driven by myosin motors [24,25] and are responsible for the cellular mechanical response [26]. Overall, the architecture of an actin network is expected to be directly related to its physiological function. The binding partners guide the organization of actin filaments, and conversely the binding of actin-associated proteins is sensitive to the architecture of the actin network [27,28]. Deciphering the principles governing the assembly of the different actin structural architectures is an important step towards a better understanding of the variety of cellular processes. Many excellent studies have focused on identifying the biochemical composition of the different actin organizations [6], but the physical and geometrical laws governing their architecture are still largely unknown.
In a previous work, we developed an in vitro method to control actin filaments assembly with a designed pattern composed of areas with different surface properties [1]. In these experiments, the geometry of the nucleation zone dictated the collective behavior of the actin filaments with some patterns resembling in vivo like structures [1]. The system was built from a minimal set of purified proteins, and avoiding the biochemical complexity of an in vivo system, it provided a well-controlled and reproducible assay to study the assembly of actin filaments in a variety of structures. Although, our study revealed how geometrical constraints affect actin assembly, the key components responsible for the formation of defined structural organizations remained poorly defined. The stochastic model developed in this study [1] helped to describe the observation but did not bring additional information on the properties of the system dictating the collective behavior. To better identify these parameters and ultimately understand the higher order architecture, we need to be able to predict the emergent actin organization, based on the microscopic properties such as the rates of actin assembly, the mechanical properties of actin filaments, the geometry of the nucleation region, and the biochemical composition of the experimental system.
Different types of modeling are available for this purpose [29]. Collective cytoskeleton behavior has been studied at macroscopic scale with ordinary or partial differential equations [30–32]. Stochastic methods (Monte Carlo simulations) can take into account the variability arising from the intrinsic randomness of the microscopic processes [15,33–35]. The potential drawback of such detailed modeling is that they can be computationally expensive, but with modern methods [29], one can simulate systems containing thousands of filaments over hundreds of seconds. We used Cytosim [36], a versatile cytoskeleton simulation software, which can be used for a diverse range of cytoskeleton simulations. Actin filaments can be simulated with different growth and shrinkage rates and bending elasticities. Associated proteins can be added (e.g. molecular motors, crosslinkers, severing proteins, capping proteins, nucleators), and environmental constraints can be imposed (confinement, asters, solid objects, laser cutting, flow…). Cytosim polymer has been used to study microubule systems: self-organization [37,38], effect of confinement [39], spindles [40], asters positioning [41], nuclear positioning [42] but also to predict experimental design [43]. Nonetheless, Cytosim had never been used to study actin assembly under geometrical constraints.
Numerical experiments based on similar modeling have already been used to study the parameters controlling the global organization of cytoskeleton components. Recently, it was shown how different modes and efficiency of actin filament crosslinking would affect the self-organization of actin structures such as the contractile ring and actin cables [44, 45]. Similarly, simulations of taxane-stabilized microtubules [46] showed that geometrical constraints (cell confinement) combined with the bending properties of the filament was sufficient to create the bundling effect observed in vitro.
In this work, we first designed a simulation of actin organization from a nucleation region with a defined geometry, and reproduced the in vitro collective actin assembly behavior, with a minimum set of parameters. We then used this system to investigate how these parameters control the network organization in a range of experimental conditions. This study showed that we have identified the key parameters that define geometrically-controlled actin assembly and are able to reproduce a variety of actin organization in silico. Thus we believe that we have in hand a powerful predicting tool, and that numerical simulations in combination with in vitro experiments will help understand more complex actin processes.
http://www.ncbi.nlm.nih.gov/pubmed/23388829
To study geometrically-controlled actin assembly, we previously developed an assay where actin nucleation is triggered from a micropatterned surface coated with an Nucleation Promoting Factor (NPF) (Fig 1A and [1]). The assay includes crowding agents constraining the filaments parallel to the glass. In this assay, actin nucleation occurs near the glass surface, and the growing actin structures never exceed 7μm in height, as revealed by confocal imaging (S1 Fig). The highest rise occurs immediately above the nucleating regions, while most of the network that extends away from them seems to be within 200 nm of the glass surface. This is confirmed by the fact that confocal and TIRF images of the system are comparable. The 2D model (S2 Fig) developed in Cytosim seems therefore appropriate to simulate the layer of actin assembly that is directly in contact with the glass, which is the part of the network that is most interesting in our experimental condition.
To simulate a geometrically constrained actin nucleation as used in experiments [1] we randomly disposed non-diffusible Arp2/3-like entities within the desired area (see S2, S3 Figs and Material and Methods). These Arp2/3 complexes are able to nucleate a new (daughter) filament, but only if they are already bound to an existing (mother) filament. Because they do not move by diffusion, and thus remain in the nucleating region, the Arp2/3 complexes can generate branches within the nucleation area, but not outside. This mimics the activation of the Arp2/3 complex by the pWA-peptide coated on the pattern (S2 Fig). The Arp2/3 complex acts as a mechanical link between mother and daughter filaments, constraining them relative to each other in position and in orientation, such that the pointed end of the daughter filament remains on the side of the mother filament, and the daughter barbed end grows out making a ~70° angle with the mother filament [47, 48]. In the same area, we also added fixed nucleators, which initiate the process by generating the primer filaments, and fixed binders, which may bind to any actin filament within their range, and are anchored at a fixed position with a Hookean spring. These “fixed binders” have a non-zero unbinding rate and thus they effectively create some friction acting on any filament entering the micro patterned region. This friction accounts for the constrained nucleation processes and is necessary to maintain a dense patterned area (S2 Fig).
The simulated filaments are not meant to only represent individual actin filaments, but may also represent bundles of several crosslinked filaments. We will thus refer to them as fibers. Similarly, the ‘steric’ interaction between fibers is meant to be effective, representing the different forces that may act between neighboring F-actins. These parameters can be varied experimentally by modifying the type and concentration of actin binding proteins, the buffer solution or fiber confinement, conditions that have been shown to induce different actin organizations [1]. Our first task was to calibrate the effective parameters of the simulation, by comparing simulations and experiments obtained with a simple pattern.
Input of Cytosim is a configuration file specifying the values of all parameters of the model: some represent physical quantities (e.g. temperature, viscosity), some are associated with the algorithm of the simulation (e.g. time step, segmentation) and were set to get an appropriate precision. Yet, most parameters are characteristics of the real components of the system (e.g. F-actin growing speed and bending rigidity). We adopted measured values when possible, using for example the measured persistence length of 15 μm of actin filaments [6] and a rate of elongation of 0.033 μm/s (actin concentration of 1–2 μM [49]). Finally, a few parameters are associated with effective interactions for which a molecular implementation is not necessarily feasible. This is the case in particular for filament-filament interactions, which are essential to capture the collective behavior of growing filaments [50,51]. Actin filaments are physically not able to interpenetrate each other and within distances equal to their diameter must experience strong repulsive force. Moreover, although similarly charged, actin filaments attract each other at short range, due to the presence of counterions in the solution [50,51]. Moreover, the presence of chemically neutral polymers in the solution creates a depletion effect that also induces neighboring filaments to pack together [51, 52]. This explains how we could observe bundles in vitro without adding any actin filament crosslinkers to the solution. It also means that changing the experimental conditions (pH of buffer, ionic strength, concentration of polymers, etc.) may affect how actin filaments interact. Rather than trying to simulate these processes in details, we have used an ad-hoc steric interaction between simulated filaments, and calibrated the parameters of this interaction to best reproduce the behavior of the in vitro system. In Cytosim, a ‘steric’ filament-filament interaction can include both attractive and repulsive forces (Figs 1B and S2):
Fs=k(d−d0),withk={Kpushifd<d0Kpullifd0≤d≤dm0otherwise
(1)
where d is the distance between the two interacting elements, d0 is the distance at which the fibers are at equilibrium (this is an effective fiber diameter, which is larger than the real diameter), dm is the maximal interaction distance between fibers (the choice of these values is discussed in Material and Methods). To fix the two stiffness values, Kpush and Kpull (Fig 1B), we tested a range of values and compared systematically the simulation outputs with experimental results at first using a simple pattern configuration: a horizontal bar (Fig 1A, 1C and 1D). On this pattern actin filaments assembled into a dense network on the patterned bar, with their barbed ends growing away from the pattern, filaments align parallel to each other and generate bundles growing out perpendicularly to the bar (Fig 1A and [1]). In the simulations we found that a balance between repulsing and attracting steric interactions strongly affects the organization of actin (Fig 1D). When the attractiveness was similar or higher than repulsion (Fig 1D panel b, c, f) actin fibers collapsed together. On the contrary, when the repulsion dominates (Kpush = 75 pN/μm, Fig 1D, panel g, h, i), actin fibers are disorganized and do not generate bundle-like structure. For a quantitative comparison of simulations and experiments we computed the average local intensity in the corners of the patterned bar normalized by the total average intensity (Ic, Fig 1C). This quantity gives a measure of the spatial distribution of the filaments: when organized in bundle-like structures, fibers are less dense in the corners of the pattern thus Ic is lower than 1. A bar plot of Ic in experiment (N = 10) and for the 9 simulation scenarios (N = 20) is shown in Fig 1D. We found that the corner intensities corresponding to panels a, b, c, f, g, h, i (Fig 1D) were significantly different from experimental ones (p-value < 0.005, Kolomogorov-Smirnov test), whereas the ones corresponding to panels d and e were not (p-value>0.1). To further compare simulations and experiments, we also measured the variation of intensity along the perimeter of an ellipse around the pattern (Id). This gives a measure of the bundling of the filaments: a high value indicates the presence of intense bundles, whereas a low value accounts for spread filaments all over the pattern. With this measurement, we found that the deviation of intensity corresponding to panels a, b, d, f, g, h, i were significantly different from experimental ones. Thus, panel e (Kpush = 7.5 pN/μm, Kpull = 0.5 pN/μm) quantitatively reproduced the in vitro actin filaments behavior most accurately. Indeed, the actin network growing from the micro-pattern in simulation (Fig 1E and S1 Movie) was fully consistent with our experimental observations [1]. All following simulations thus used these calibrated parameters.
When more than one area of nucleation is present on the pattern, the distance between them, and their relative orientation have a major impact on the emergent organization [1]. Indeed, actin filaments coming from two different actin networks may change their initial direction when contacting each other. A typical example illustrating this behavior is the global organization of actin filaments initiated by V-shaped branches of an eight-fold radial array (Fig 2A). Growing actin filaments elongating outward from each ray formed parallel bundles on the bisecting line between adjacent rays (Fig 2A, left panel). These parallel bundles originate from a transition point that separates the assembly of antiparallel bundles in the proximal part of the rays and the assembly of parallel bundles in their distal part (Fig 2A and cartoon Fig 2B). Depending on the angle θ between two patterns creating a V-shape motif (Fig 2B), actin filaments contact each other differently [1]. We also noticed that addition of a crosslinker (fascin) from the beginning of the experiment changed the final organization of the filaments thereby increasing the proportion of anti-parallel structures. A likely explanation for this is that crosslinked actin filaments generated by fascin behave differently than isolated actin filament (Fig 2A, right panel). One property affected by crosslinking is the stiffness of the resulting bundle. We therefore simulated V-shape patterns with different angles and varied the persistence length Lp of the simulated filaments, a quantity proportional to the bending stiffness K (K = kBT Lp) (the results of varying the bending rigidity are further discussed in Material and Methods).
For a native persistence length of Lp = 15 μm (Fig 2C, middle panels and S2 Movie), fibers meeting with a shallow angle usually grow past each other and form anti-parallel structures (θ = 0°, meeting angles > 160°). On the contrary, fibers growing from two bars disposed at right angle (θ = 90°) tend to bend as they meet, resulting in parallel structures (meeting angles < 20°). For an intermediate angle (θ = 45°), the two types of arrangement coexist: as in experiment (Fig 2A) antiparallel structures are found in the bottom part between the two bars, and parallel structures in the top part. We found that the rigidity of the fibers (relative to the filament length) controls the proportion of parallel and anti-parallel structures. At high persistence length (Lp = 1000 μm, Fig 2C bottom panels and S4 Movie) corresponding to bundles of 8–10 F-actins [53], anti-parallel structures are more prominent, compared to the structures simulated with a native persistence length Lp = 15 μm. This behavior is reminiscent of the effect of fascin (Fig 2A, right panel). At small persistence length (Lp = 2 μm), corresponding to actin filaments decorated by ADF/cofilin [54], the distance between the bars is much greater than the persistence length and thus the fiber brushes deform upon interaction (Fig 2C top panels and S3 Movie).
We next analyzed the local orientation of actin fibers along the bisecting lines when we varied both the angles between the two bars and the polymer persistence length. We classified fibers whose orientation compared to the x-axis was around 90° ± 20° as parallel filaments, and those whose orientation was of 0° ± 20° as anti-parallel. The graphs (Fig 2C, right panel) show the proportion of different actin fibers orientations as function of the angle θ between the patterns (for an angle varied from 0° to 120°) for the 3 different persistence lengths. This analysis demonstrated the continuous dependence of the structures on the pattern angle.
These results can be rationalized as follows: since filaments initially grow primarily perpendicular to the bar, they need to bend by an angle θ /2 to enter an antiparallel structure, and by 90°—θ /2 to enter a parallel structure. For θ < 90°, the bending angle required to form an antiparallel structure is smaller than the one required for a parallel structure, which makes it unfavorable for stiff filaments to form a parallel structure, all the more if they are short. This is because given a certain force F applied transversally at the growing tip of the filament, its bending angle will scale as θ ~ FL2/K, where K is the filament bending stiffness (K = kBT Lp) and L the length of the actin filament. At the bottom of the pattern, actin filaments meet while they are short, the bending required to enter a parallel structure is consequently unfavorable, and anti-parallel structures are created instead. On the top of the pattern however, actin filaments are longer as they contact opposite filaments; and consequently are easier to bend leading to parallel bundles (Fig 2B and 2C).
To further confirm this analysis, we repeated these simulations with fibers of 15 μm persistence length and varied the length of the fibers (and thus the distance between the patterns and the simulated time as well). The analysis of the presence of structures (S4 Fig) revealed that indeed the ratio between the polymer persistence length and their length will determine their collective behavior. Indeed, short actin fibers with a persistence length of 15 μm will have the same behavior than shorter fibers with 2 μm persistence length or longer fibers with 1000 μm persistence length (S4 Fig). The definition of the persistence length makes this relation between filament length and rigidity obvious for isolated filaments, but was necessary to demonstrate in the context of collective filaments behavior.
Thus, both the geometrical conditions of nucleation and the mechanical properties of the simulated filaments are key parameters controlling the formation of anti-parallel or parallel structures.
We next simulated a pattern containing two parallel pWA (Arp2/3 activator) bars of width 3 μm to study the “primer effect”. This effect is based on the ability of a drifting actin filament to contact the nucleation area and trigger actin assembly (S2 and S3 Figs). The in vitro experiment was constructed such that a growing actin filament coming into contact with a virgin patterned region will nucleate new actin filament on its side [55]. However, when two or more patterned regions are in close proximity, a region of pWA may already be covered by an actin network (Fig 3A). We therefore have two possible scenarios for the fate of a filament reaching a pattern of pWA on which a branched network is already present (Fig 3A): (i) the filament gets entangled in this network and stop growing, (ii) the filament can nucleate new filaments as if the pattern was virgin (and get entangled or not). To find which scenario is correct, we compared simulations and in vitro experiments by systematically varying the distance between the two bars. As control condition, i.e. non-interaction between the two patterns, we selected cases where the distance d between adjacent bars is above (d∞ = 80 μm). In this case filaments from one bar do not reach the other bar.
To assess the contribution of a neighboring pattern on the overall actin network we computed:
ρext(d)=I(d,tf)I(d,ti)−I(d∞,tf)I(d∞,ti),
which measures to what extent the outer (away from the neighboring bar) intensity is made different by the presence of the second bar (Fig 3A). For this we calculated the intensity of the network (≈density of filaments) close to the pattern (7 μm away from the pattern center; a pattern is 3 μm wide), at two different times ti and tf = ti + 1h (Fig 3B). If the filaments cannot cross the pattern and get entangled, the variation of intensity at the outer sides of the pattern should be independent of the distance between the bars (Fig 3A) and thus similar to the control case (ρext(d) ≈ 0). On the other hand in case the filaments cross and/or nucleate on the other bar we expect the outer intensity to be higher than in the control case (ρext(d) > 0) and ρext(d) would be a decreasing function of d. Because the computational costs were too high, we simulated shorter and closer pattern bars than experimental ones. To allow comparison of the results, we looked at normalized distances (d/d∞).
We performed 100 simulations with the two different hypotheses, using a random distance d for each simulation (Fig 3B, and S5 Movie). We calculated ρext(d) for each simulation and analyzed its dependency on the normalized distance (Fig 3C) for the assumptions of entanglement (H1 solid red line) and nucleation (H2 dashed blue line). We then analyzed the results of four in vitro experiments (Fig 3B and 3C). The results are significantly positive for distances smaller than d∞ (with 95% confidence with Wilcoxon rank test for comparison of experimental results with null distribution), and decreasing with d (p-value < 10–12 with Spearman's rho test). We therefore found that the second hypothesis (H2, nucleation) best matches the experimental behavior. However, a simulated nucleation efficiency of 100% was too high according to the experimental data (Fig 3C). Thus, we decreased the efficiency down to 2,5% (Fig 3C, solid blue line). This predicted low value for “primer activation” is consistent with experimental quantification of the nucleation in presence of the Arp2/3 complex [48]. We therefore can conclude that in vitro, actin filaments may cross over neighboring bars, nucleate new filaments, and thus influence the network on the distal side of another patterned region.
To unveil how actin filaments can be organized into higher structures, a lot of efforts went into analyzing the role of actin-associated proteins [2,6,56]. While largely justified by the importance of protein effectors, this was also due to the lack of tools that allowed one to dissect how geometrical or physical parameters affect the overall actin architecture. However, recent technological developments such as microfluidics [57], micropatterning [1,58,59] or cytoskeleton growth into defined volumes [39] have highlighted some of the key features of biochemical-independent parameters in controlling the cytoskeleton architecture. Examples of applications include showing how chromatin shapes the mitotic spindles organization [60] or how actin nucleators of the formin family respond to mechanical stress [61]. Using a micro-printing technique, it was also possible to demonstrate that geometry is a key parameter in controlling the macroscopic architecture of actin filaments during assembly [1]. This made it possible to document the emergent behavior by which actin assembly organizes at higher scale, depending on the initial localization of the actin nucleation-promoting factors.
Using Cytosim, we developed a model of a geometrically-constrained actin assembly, in which actin filaments/fibers are initiated from defined regions where branched nucleation occurs. This framework allowed us to study how the biophysical properties of actin filaments and their environment determine the organization of the final network. It is a tool with which one may study actin dynamics in more complicated systems. We were able to reproduce a diversity of actin organizations obtained from an initial geometry of nucleating promoting factor (Figs 1–3). We identified a combination of three essential components that determines the actin organization (Fig 4A and 4B and S6 Movie). The first one is the steric interaction between filaments, this is essential to obtain an aligned distribution of actin filaments growing away from the nucleating pattern (Figs 1 and 4B). In absence of steric interaction actin filaments have a tendency to buckle (Fig 4B1) preventing them from extending away from the nucleating region, and thus limits the contact between filaments nucleated from two different areas (Fig 4B1). These interactions are modulated in vitro by the chemical properties of the media. The second key component is the bending elasticity of actin bundles [62]. A persistence length of Lp ≈ 15 μm, close to the experimentally reported persistence length for single actin filaments [62], resulted in patterns with parallel and antiparallel actin organizations similar to the patterns observed in vitro (Figs 2 and 4B3). The model predicted the effect of regulatory proteins, such as ADF/cofilin [54] or crosslinkers [53] that modify the persistence length, on the observed patterns. Actin filaments decorated by ADF/cofilin are softer (Lp of 2 μm [54]) and will buckle under small force and we predicted that this would preclude the efficient formation of parallel or antiparallel organizations. On the contrary, a high persistence length (Lp = 1000 μm) induced for example by bundling of actin filaments in the presence of crosslinkers, limited the deformation of the growing polymers. In this case we predicted that the original geometry imposed by the micropattern is maintained, indicating that this is the main parameter that defines the final macroscopic organization during actin assembly. The effect of adding fascin in our system could be simulated by adapting only the fiber persistence length, without having to change the fiber steric interactions. This suggests that the electrostatic properties of the interaction could be negligible compared to the change in elastic property of the fibers, which arise from the capacity of fascin to generate stiff bundles [63]. For the native Lp ≈ 15 μm, the situation is more complex and the resulting actin macroscopic organization resulted from a combination of the boundary imposed by the geometry of the micropattern and the deformation induced on growing actin filament when contacting each other (Figs 2C, S4 and 4B3). Finally, a third component that is important in defining the overall actin macroscopic organization is the ability of growing actin fibers to contact the nucleating pattern and therefore trigger additional actin assembly following the ‘primer’ activation [55]. By doing so, the density of the actin organization will be modulated according to the efficiency of the ‘primer’ activation (Figs 3 and 4B4). The accessibility of the nucleating region is therefore a key parameter in controlling the overall actin organization. As demonstrated in Fig 4, the combination of these three components is sufficient to capture most of the actin organization observed in vitro, with the geometrically constrained actin nucleation assay.
To further test the power of our simulation tool, we predicted the overall collective behavior of actin initiated from more complex pattern geometry (Fig 4C). We simulated the actin filaments organization obtained from square and diamond dotted patterns (Fig 4C). We then constructed the same geometries experimentally (Fig 4C). In both simulations and experiments the network were organized similarly (See Fig 4C zoomed panel). This shows that our simulation framework is predictive. In its present form our simulations with Cytosim have a number of limitations. Firstly, the high computational demands (S2 Fig), lead us to simulate the model in 2D using a reduced number of effective filaments. This likely explains why the model could not fully reproduce the formation of very thick parallel actin bundles at the bisecting line between each ray of the 8 branched radial arrays (Fig 4). A 3D model, possibly reduced in size, but one in which single F-actin are simulated, could be used in the future to explore this further. However, even with these limitations, our current model is already useful to test a variety of geometrical configurations and predict the macroscopic organization of actin filaments. Cytosim was used here to mimic branched actin nucleation via Arp2/3, with minor modifications we could study other modalities such as formin dependent nucleation [64], the role of processive elongation induced by Vasp [65], or predict actin organization obtained by mixing several nucleation mechanisms. In the future we could investigate how myosin motors deform specific actin architectures [27] or study the movement of single myosin or myosin-driven cargo on established networks. Thus the present study sets the foundation of future research, where one will attempt to realistically reproduce in silico the emergent actin organization observed in vitro.
Actin filaments are simulated with Cytosim, following a Brownian dynamics approach [36]. Filaments may bend following linear elasticity, and are surrounded by an immobile viscous fluid; they grow and interact with each other (see Figs 1 and S2). Considering the low Reynolds number of a filament at this scale, inertia can be neglected and the mass of the objects is not a parameter that appears in the equations [37]. The parameters of the model were chosen to reflect the characteristics of in vitro actin filaments (see S1 Table). Notably, actin growth is force dependent, and the assembly speed is reduced exponentially by any antagonistic force [66] present at the barbed end:
v=v0exp(f⋅tfs),iff⋅t<0andv=v0otherwise,
with fs = 0.8 pN ([67]). f and t are respectively the force vector and the normalized tangent vector at the barbed end. Moreover, considering the extensive supply of monomers in the system, depletions effects were neglected, and it was not necessary to simulate actin monomers to correctly model the increase of “polymer mass” in the system.
To simulate the formation of patterns, three entities were used: nucleator objects, Arp2/3-like complexes, and binder objects (see S2 and S3 Fig). First, we placed randomly few nucleators on the pattern area, which will nucleate new ‘primer’ filaments. These nucleators can trigger the nucleation of one filament each, and have a fixed position so that they remain on the pattern. To simulate the branched network generated by activated Arp2/3 on the pattern, we placed Arp2/3-like objects in the pattern, which bind to any filament that is sufficiently close. Bound Arp2/3 object will then nucleate a new filament with an angle of 70° with respect to the other filament on which it is bound. These complexes are placed in a fixed position on the pattern, but will move with the filaments after binding, without inducing any constraint. The link between the mother and daughter filament is modeled as a Hookean spring, with a torque to constrain the angle between the branches. To add some friction on the pattern, we also placed on the pattern fixed binders that may bind and unbind to the filaments, thus restraining their motion. These binders are also modeled as Hookean springs between the fixed anchoring point, and the attachment location on the fiber.
S1 Table shows the important parameter values used in the simulations. The complete configuration file used to simulate a horizontal pattern (see Fig 1) is given in Supplementary Data.
The interaction between filaments is assumed to come from both depletion and electrostatic interactions. These are modeled in Cytosim as Hookean springs acting between the modeled segments, with an equilibrium distance d0 corresponding to the diameter of a fiber. Because of excessive computational costs (S2 Fig), we could not simulate all the filaments present in the system, and reduced their number by lowering the density by roughly a factor 10. For this, we used an effective diameter d0 = 100 nm (while the diameter of one actin filament is around 7 nm), and a maximum interaction range of dm = 200 nm (while interactions between actin filaments mediated by electrostatic or depletion forces would be limited to tens of nm). Choosing d0 and dm, as well as the scale of the simulated system (length of the filaments, simulated time…) resulted from an empirical tradeoff between computational expenses and accuracy. With the chosen values, one filament usually interacts with its first and 2nd neighbors, and any modification of this aspect of the model can strongly affect the emergent organization. For example changing the value of dr (defined from dm = d0 + 2 dr, see S5A and S5B Fig) indicated a strong effect on filament bundling. Thus the adjustment between the filament radius, interaction range, and steric coefficient values is delicate, and the calibration procedure is essential to find appropriate values of those parameters. Note that with our parameter set, each simulated filament may represent ~10 neighboring actin filaments, and one should be careful while interpreting the results at a molecular scale.
We computed fibers of 3 different persistence lengths (2 μm, 15 μm, 1000 μm) to account for the effect of actin binding proteins. We calibrated the steric parameters for filaments having a native persistence length (15 μm), and kept these values fixed for the other persistence lengths. Indeed, we considered that the repulsive force, due to hard-core repulsion, should not be strongly affected by the addition of other proteins, so Kpush could be kept constant for all rigidities. The electrostatic interaction term, Kpull, could be affected by the addition of actin binding proteins if they do change the electronic charge of the filaments, or the depletion force. However, we do not have any measure of this potential effect to recalibrate this value, and keeping the same value was sufficient to explain the experimental observations in the presence of fascin. Moreover, we found that the range of steric parameters giving patterns similar to the one observed in vitro was minimally affected by the value of the persistence length (S6A and S6B Fig).
The increased persistence length due to fascin addition could have been modeled by true addition of crosslinker in our simulation, or by increasing the value of Kpull. However, to be able to compare the behavior of fibers with different persistence lengths, we preferred to keep as many parameters as possible constant, and to vary instead the persistence length.
In all simulations we kept the same density of fibers. Therefore we do not explicitly model the decrease in fiber number that is expected from bundling. This choice was motivated by the computational costs of simulating large number of fibers. For native persistence lengths we simulated 300 to 1700 fibers, a 10 times decrease in case of bundling, i.e. 30 to 170 fibers, will not allow for comparison of the actin organizations. Instead we compared the behavior of single fibers against bundle of fibers for a same density of fibers/bundle of fibers.
We also tested how sensitive the simulations were to the chosen value of the filament persistence length, by varying it between 2 μm and 1000 μm (S7 Fig). We noticed that above a threshold, around 10 μm for the persistence length, the collective behavior of actin filaments was mostly similar with the same tendency of forming bundles (S7 Fig). Thus the choice of a precise value of 15 μm while actin persistence length is evaluated to be between 10 and 20 μm is acceptable.
To simulate the entanglement effect on the pattern and its effect on actin elongation (Fig 3), we defined two different fiber types in Cytosim. Their parameters are identical, but this allowed us to define an entity in the pattern that could bind at the barbed end of actin filaments coming only from the other pattern. When a filament has one or more of these entities bound close to its end, it will stop growing. For the second scenario (primer effect), we added the possibility for these entities to nucleate a new filament when bound. By mixing the two types of entities (capping one and nucleating one), we could control the nucleation efficiency.
|
10.1371/journal.pntd.0005345 | Exploring local knowledge and perceptions on zoonoses among pastoralists in northern and eastern Tanzania | Zoonoses account for the most commonly reported emerging and re-emerging infectious diseases in Sub-Saharan Africa. However, there is limited knowledge on how pastoral communities perceive zoonoses in relation to their livelihoods, culture and their wider ecology. This study was carried out to explore local knowledge and perceptions on zoonoses among pastoralists in Tanzania.
This study involved pastoralists in Ngorongoro district in northern Tanzania and Kibaha and Bagamoyo districts in eastern Tanzania. Qualitative methods of focus group discussions, participatory epidemiology and interviews were used. A total of 223 people were involved in the study. Among the pastoralists, there was no specific term in their local language that describes zoonosis. Pastoralists from northern Tanzania possessed a higher understanding on the existence of a number of zoonoses than their eastern districts' counterparts. Understanding of zoonoses could be categorized into two broad groups: a local syndromic framework, whereby specific symptoms of a particular illness in humans concurred with symptoms in animals, and the biomedical framework, where a case definition is supported by diagnostic tests. Some pastoralists understand the possibility of some infections that could cross over to humans from animals but harm from these are generally tolerated and are not considered as threats. A number of social and cultural practices aimed at maintaining specific cultural functions including social cohesion and rites of passage involve animal products, which present zoonotic risk.
These findings show how zoonoses are locally understood, and how epidemiology and biomedicine are shaping pastoralists perceptions to zoonoses. Evidence is needed to understand better the true burden and impact of zoonoses in these communities. More studies are needed that seek to clarify the common understanding of zoonoses that could be used to guide effective and locally relevant interventions. Such studies should consider in their approaches the pastoralists’ wider social, cultural and economic set up.
| Zoonoses are diseases transmissible between animals and humans. Risk factors include animal slaughter, the handling and preparing food of animal origin and particularly the consumption of such food when raw or undercooked. Pastoralists are daily in contact with their livestock and are likely to be more frequently exposed to zoonotic pathogens. However, very few studies have focused on the understanding of zoonoses and related risk factors among pastoralists of Tanzania. This was striking considering pastoralists’ close bond with and reliance on livestock for their nutritional needs and livelihoods. Our study was implemented in pastoral communities located in two different ecological zones of Tanzania. In all ten communities visited no local term for zoonoses existed, yet participants recognised there are some symptoms that appear in both animals and humans. However, these were not thought to be harmful to humans and people did not perceive the animal products from which they originate to be dangerous. Other zoonoses were known because community members received a diagnosis from a health facility. We also learnt that cultural practices relating to the preparation and consumption of food of animal origin were considered more important to the community than the risk of infection. The findings of this study highlight the gaps in knowledge of zoonoses among pastoralists in Tanzania and the importance of social and cultural practices. Therefore, we propose that interventions to control zoonoses should be informed by research of peoples’ behaviours related to handling of animals and consumption of animal source foods.
| Zoonoses account for the most commonly reported emerging and re-emerging infectious diseases (EIDs) in Sub-Saharan Africa [1]. There is a large number of human diseases that originate from both domestic and wild animals (900 +), and a smaller number of zoonoses (250+), which are diseases transmitted between animals and humans [2]. The frequency of occurrence of zoonoses is relatively small in the overall burden of human disease but zoonotic EIDs are of major concern globally [3]. Although zoonoses have been reported in Africa, the continent is the least capacitated in terms of technological advancement and capacity for the early detection and control of infectious diseases in both human and animal populations [4–6].
Control of zoonoses is beneficial for developing economies and human health, as it reduces morbidity and mortality, saves costs for disease management and increases productivity, health and well-being [7]. Livestock keepers in Africa in general and Tanzania in particular, face a potential double burden of animal and human diseases as they fend for their livelihoods. They are a group at risk of zoonosis due to livelihood and livestock keeping practices, which puts them in direct contact with their livestock and wildlife and therefore at high transmission risk [8–11]. Another important transmission pathway is related to the consumption of animal source foods, in particular specific local practices like consumption of raw or undercooked meat, and drinking raw blood/or milk. The practice of sharing community water sources with livestock constitutes another potential source of zoonotic infection [12,13]. In Tanzania, a number of studies have documented the presence of infections that are potentially zoonotic. They include bovine tuberculosis [14,15], rabies [16,17], brucellosis [18,19], anthrax [20,21]and Rift Valley fever (RVF) [22–24]. Most of these diseases affect people’s capacity to effectively manage their livestock and abate food insecurity.
Pastoralists have lived close to their animals for millennia and possess a wide repertoire of local traditional knowledge systems in the identification and addressing of both human [25] and animal afflictions [26]. Some scholars have gone as far as comparing pastoralists diagnostic skills to that of the modern medicine [27]. Thus, early detection and response to illnesses are important steps towards effective interventions. Such knowledge is important to facilitate communication between pastoralists on the one hand, and animal and human health experts on the other [28,29]. In spite of the socio-political changes affecting pastoralists in sub-Saharan Africa, they still practice a traditional way of life that is rooted in a cultural system, which guides all spheres of life, including human health. Health and illnesses are approached both using naturalistic and ritual activities which inform their causes and the subsequent interventions. The Maasai pastoralists, like many other African traditional societies are said not to distinguish between religious beliefs and empirical knowledge when it comes to seek healing [25,30] but prescription of the most relevant treatment has to come from understanding the source of the illness since they make a distinction between natural and supernatural caused illnesses.
For many illnesses pastoralists rely on a number of options, which include local healers, spiritual diviners, midwives and modern medical care. Illnesses that are believed to naturally occur are addressed using medicinal plants and/or modern medical care, and those that are believed to arise from misfortune or unusual causes are addressed through consulting a traditional diviner [30]. The breadth of pastoralists’ knowledge and usage of medicinal plants demonstrates the extent of treatment, which relies on local knowledge [31]. For example, Fratkins’ account of the Samburu pastoralists in northern Kenya lists more than 90 medicinal plants that are believed to treat more than 50 different ailments and conditions [25]. Similarly, a number of scholarly works have documented medicinal plants and corresponding treatment for the Maasai of Eastern Africa [31,32], Dinka of Sudan [33] Fulbe and Arabs pastoralist groups of Northern Cameroon [29] and pastoralists from Ethiopia [34]. Some of the ailments covered include fever, malaria, chest congestion, wounds, burns, women's stomach pain, hepatitis, snakebites, stomach problems, colds, polio, gonorrhea, arthritis and abortions among others. There are also additional human afflictions and misfortunes that are believed to come from not following relevant culturally informed diets, a process that is said to ‘pollute’ the body and block internal circulations. For example, the consumption of wild animals is prohibited among the Maasai [25].
Qualitative studies on zoonoses, which attempt to assess knowledge and perceptions of zoonoses and risky behavioural practices towards transmission in Tanzania are few and are still rather descriptive [28,35] There have been a few studies on knowledge, attitudes and practices among pastoralists, but their designs have important limitations [36]. For example, Shirima and colleagues [8] noted that Maasai pastoralists named malaria, east coast fever (ECF), mastitis, allergies, typhoid fever and cancer as zoonotic. The authors suggested that pastoralists fail to correctly recognize the animal sources of certain infections. Therefore, there is a need for well-planned qualitative analytical studies among pastoralists. Such studies are likely to support the integration of traditional knowledge systems into modern health approaches that can improve zoonosis management. This study was carried out to understand perceptions and knowledge on zoonoses among pastoralists in Ngorongoro, Kibaha and Bagamoyo districts of Tanzania. Understanding people’s perceptions about zoonoses and other infectious diseases, within their social and cultural context is important in order to provide relevant evidence for planning appropriate interventions [37].
The Medical Research Coordinating Committee of the National Institute for Medical research approved the study (Ref. NIMR/HQ/R.8a/Vol. IX/1649). Prior to all interviews and discussions, information about the study, which included purpose of the research, confidentiality and uses of data, was read out to the participants and a local translator translated into the local language of the participant where the National Kiswahili language was not popular. Verbal informed consents to participate and allow recording of the interviews were obtained from all participants before beginning.
The study was carried out as part of a larger programme on infectious disease control among Maasai pastoralists of northern and eastern Tanzania with the aim of formulating participatory disease control interventions through research. This research was implemented between May 2014 and May 2015, in Ngorongoro (Northern Tanzania), and Kibaha and Bagamoyo (Eastern Tanzania) districts (Fig 1). A total of ten (Ngorongoro = 6; Kibaha/Bagamoyo = 4) villages were included (Table 1).
The study villages were therefore selected from these three program districts. The selection of the villages for this study was purposefully done by taking into account the contrasting landscape within each district. These included the farming system practiced [pastoralism or pastoralism and crop cultivation] altitude, and location [located within a conservation or not] and to a lesser extent, ethnic groups.
The Ngorongoro District is comprised of arid and semi-arid lands, more favorable to nomadic livestock keeping than crop agriculture [38]. Ngorongoro District is home to a number of pastoral communities, typical of sub- Saharan Africa, where traditional transhumance pastoralism, wildlife and crop production exist side by side. The district is inhabited mainly by the Maasai, a semi-nomadic ethnic group, who make up approximately 85% of the population [39]. The district is also divided into areas determined by two main socio-economic activities. The Ngorongoro Conservation Area Authority (NCAA) in Ngorongoro Division, where pastoralists reside with wildlife but cultivation is strictly disallowed by regulation. Pastoralists in the NCAA receive subsidized maize parcels once in a while to compensate for non- cultivation. Beyond the NCAA, in the Loliondo and Sale divisions to the north, a combination of wild life game controlled areas and cultivation exist side by side.
Kibaha and Bagamoyo districts are humid savannah peri-urban sites situated along the eastern coastal hills with limited human-wildlife-domestic animal interactions. The sites comprised mainly of Maasai pastoralists who share their borders with crop farmers. Other pastoral groups such as the Sukuma and Datooga have recently migrated to Kibaha district in search of pasture, water and markets [40,41]. In our study Sukuma pastoralists were included as participants from Kwala division as they are the largest and settled pastoralist group in the area compared to the Datooga and Maasai. While our main focus was on Kibaha pastoralists, the inclusion of a village situated in Bagamoyo district was necessary due, not only to the shared boundaries with neighbouring pastoralists in Kibaha but also shared resources such as pasture, water and markets.
Multiple data collection techniques were used to gather data for this study. Focus Group discussions (FGDs) were mainly used, which were combined with interviews and observations, and participatory epidemiology. The observations were combined with interviews and were conducted while participating in events of interest in the community [42]. These events included attending local markets, slaughter of animals, grazing and herding livestock, milking cattle and identification of livestock diseases. In addition, participatory epidemiology was employed through asking group participants to rank and weigh illnesses in the order of importance. During discussions communities were involved in defining and prioritizing veterinary-related problems and solutions [43].
All data collection activities were conducted by trained researchers supported by a translator who was conversant in the participants’ local language and the national, Kiswahili language, which was used by the facilitators and the interviewers. One of the research team member (MN) is a Maasai and spoke the Maa language fluently. He assisted in most of the interviews and FGDs. Some participants, especially the Maasai from the coastal districts of Kibaha and Bagamoyo, spoke Kiswahili more fluently compared with participants from Ngorongoro.
FGDs using a question guide were conducted in separate groups consisting of men or women. The guide included topics on human and animal health, health-seeking practices for humans and animals and livelihoods and food security. Most discussions lasted about one hour. In addition, interviews were done with village leaders, local chiefs, medical and local livestock officers, and district officials. Some of these were informal while others were prearranged [42]. The informal interviews did not necessarily last long as they were conducted on the spot to clarify aspects of the study observed in the community or during the FGDs. Furthermore, in-depth interviews were conducted with livestock keepers during dry season migrations, market days, at dipping facility sites or at local slaughter slabs.
Both in-depth interviews (IDIs) and FGDs were conducted in a flexible and iterative manner that adapted to the topics emerging during data collection. As new themes emerged and were considered important, additional questions were formulated for follow up probes with other interviewees. For example, it became apparent to the team that to better understand how people perceive the potentiality of zoonoses, it was necessary to conduct observations and interviews at the point of meat inspection.
Ranking of important perceived zoonotic illnesses were conducted by first asking to list the commonly occurring perceived illnesses and then proceed to rank based on their discussions. A simple ordinal scheme was used whereby the illness perceived by the participants as having greatest impact on livelihoods and affecting their daily engagement in production activities was assigned a value of 1, the illness seen as the second got a value of 2, and so on. During the ranking process it was important to obtain a common agreement of the common illnesses and their perceived status. To do so, participants in groups were encouraged to discuss among themselves and agree on whatever they thought were the major syndromes encountered in their village.
Discussions from the digital recorders were transcribed aided by the F4 transcription software. The transcripts that were in Kiswahili were translated to English for analysis by the research team. The Maa speaking researcher, directly translated transcripts that were in Maa, into English. The results obtained from each illness were analyzed guided by the content analysis deductive method [44] based on a categorization matrix. Data was coded according to the categories preferred under corresponding main themes. Initial analysis commenced while still in the field through expanding the notes after the interviews and FGDs were conducted [42]. This also included cross-checking the described symptoms with human and animal health professionals in the field. Coding initially was done using Nvivo10 (QSR International) by identifying themes and sub themes related to the topic of interest, such as the perceptions and local understanding of the meanings of zoonosis. The transcripts were read several times to get an overall understanding of the messages in the text and the codes were refined, added or removed depending on their significance to the developing data. Hand written notes from numerous follow-up short and long interviews were manually incorporated into the related themes. In each theme and sub theme of interest, discussions were built around them and links between were identified from which the main conclusions were reached. Illness rankings were analysed using a simple descriptive ranking with the illness receiving the highest rank toping the list followed by the rest.
Demographic data of participants are summarised in Tables 2 and 3. Overall a total of 203 individuals (females = 110, males = 93) participated in the discussions and the illness ranking exercises. Of the participants, 61 and 142 people represented Kibaha/Bagamoyo and Ngorongoro districts, respectively. Most of the participants were married (88%). About half (45.8%) of the respondents had no formal education; 49.8% had completed primary education while 4.5% had above primary education. In addition, a total of 20 people were participated in face-to-face interviews (Table 4). Participants owned different types of animals, which in the order of importance included cattle, goats, sheep, donkeys, dogs, chickens and cats. However, only participants from the coastal district sites owned chickens and cats. Summaries of the main thematic analytic results are presented on Table 5.
To understand how participants perceived zoonosis, it was important to first focus on the etymology of the terms of illness used in the local language. Since the term “zoonosis” did not exist in the local language as a single word, posing the question demanded a careful but broad interpretation of the term. In initial discussions, the study participants mentioned two perceived illnesses that they categorized as ‘zoonoses’. Particularly with participants of Kibaha and Bagamoyo districts the idea of the existence of ‘zoonoses’, especially from infections as a result of consuming meat/milk/blood from a sick animal, was met with rather a long pause and counter questions as to whether there is such a possibility of humans getting infection from afflictions that affect their livestock. Some questioned repeatedly if humans can get infected from consuming meat. In all groups, it was notable that participants mentioned fewer syndromes transmitted from livestock to humans as compared to naming illnesses and syndromes that affected their livestock.
There was a sharp contrast in the way pastoralist groups from the eastern and northern districts talked about illnesses that are potentially zoonotic. In Ngorongoro District some participants recognized that some illnesses were shared between livestock and humans, although there were also individuals who were skeptical. On the other hand, participants from coastal districts were generally skeptical about the existence of zoonoses. Zoonotic illnesses were generally not perceived as a threat or harmful among those who said they were aware of infections contracted from animals. In the coastal districts however, the notion of zoonosis was reported as uncommon. In some groups people were made aware of this possibility for the first time through our discussions. Some of the responses below attest to this observation:
Although some participants reported to have heard of zoonotic illnesses such as tuberculosis and anthrax, this was not well received by the rest of the discussants who refuted such a possibility. In the male FGD at Magindu, a long debate among the participants took place following the moderator’s posing of the question about the presence of zoonosis. One member responded to the existence of emboroto (mostly referring to anthrax by most pastoralists in northern Tanzania) in the community. Signs mentioned included swellings on some parts of human body. One middle- aged participant insisted it came from eating meat but most of the members seemed to disagree. Although they knew the syndrome, they had never heard that it was acquired from consuming meat or milk:
The discussion from this group showed that there was unclear understanding on the source of this illness and they thought to be something else, as anthrax has not been reported in the community before.
The discussion guide therefore had to be expanded with a number of follow up questions and probes. For example, we asked what animal diseases or health problems they think and believe they could acquire or become infected with. Another follow-up question was what [animal] syndromes they can acquire from staying close or living with their animals. Further, in many instances participants were asked what syndromes they can acquire from consuming animal products such as meat, blood or milk. These questions resulted in broad responses and more discussions, which were guided more by the participant’s perspectives. This approach allowed participants to be as detailed as possible as to what they could group as ‘zoonotic’ and this led to an interesting listing of perceived zoonoses (Table 6).
Throughout the discussions and interviews, participants differentiated zoonoses from two main angles. These were illnesses that they believed to have been around for many years; and illnesses handled at a health facility and from receiving a diagnosis. For the latter, which included tuberculosis and brucellosis, they were informed by health workers at health facilities that the infection is acquired from consuming animal products such as milk, meat and blood from infected animal. The participants in the FGDs referred to these illnesses as ‘hospital diseases’. The group of syndromes that have always been there included anthrax (known as emboroto) and foot and mouth disease (known as oloirobi in the Maa language).
Syndromes in humans attributed to zoonoses were identified by correlating syndromes as they presented in suspected animals. For example, the correlating syndromes in humans for oloirobi were blisters around the mouth, flu, nasal discharges, sneezing, fever, cough and diarrhea. Participants in one group reported that the syndrome (oloirobi) could also affect pregnant women, in which the illness could be transmitted to the unborn child and also from one human to another through the air. The syndrome was described to be self-limiting with spontaneous recovery without treatment. Some people would use analgesics such as paracetamol to relieve the associated pain. Furthermore, the illness occurrence in human is seasonal and coincides with the occurrence of foot and mouth symptoms in cattle. It was perceived to come from consumption of milk during the peak of rainy season (April-July) when milk is plenty. However, many participants in the Ngorongoro believed this seasonal occurrence of the illnesses has also changed and now it may occur at any period during the year (appearing at least three times).
Anthrax (emboroto) was perceived locally to have been present for generations. Livestock keepers reported that mostly signs were only recognized after the animal had died although some were apparently health before. The syndrome is characterized by swellings, and oozing of blood from the nostrils and/or ears. When dissected some parts of the carcass are noted to be black in colour. In humans, recognised symptoms include swellings, general body malaise, fatigue and loss of appetite. The swellings, later turn in to black spots during healing. In the local Maa language this black spot is called emboroto. It is also the blackness that is recognized in a suspected case in an animal, which is observed only after death.
On the other hand, ‘hospital diseases’ syndromes were said to be diseases that the health care providers will diagnose in people. Interestingly, in most of the groups in Ngorongoro, participants referred to brusela (brucellosis) as a new illness. Upon further probing the illness was considered new because they had only been told about it by health workers from year 2007, although a few cited it from the early 2000. In Osinoni (Ngorongoro) it was given a local name nangidaeton- meaning discomfort when moving, the feeling of general body malaise. But they insisted it was not there before 2007. This suggests that it was perceived as an emerging illness.
Among the participants in eastern Tanzania, tuberculosis (TB) was perceived as somewhat uncommon among pastoralists and its link with milk seen as untrue. For example, it received some resistance among the Magindu male groups for its presence among pastoralists. Many discussants reported that they only hear about the TB when a relative is taken to hospital and diagnosed with the disease to be told by health care providers that it could be because of drinking unpasteurized milk. Participants strongly disagreed that they were in danger from drinking milk and directed that notion to other social or cultural factors. Those who had heard of the possibility of milk being a source of TB did not directly relate the TB to the milk itself but to something else:
Other perceptions of possible syndromes locally identified by the participants demand further scientific investigation if they actually exist or do fall within the category of zoonotic illnesses. For example mapafu ya kikohozi ya mbuzi -a syndrome that was said to attack only goats with prominent coughing symptoms was mentioned as an infection that later occurs in humans, manifesting with more or less similar symptoms. This syndrome is acquired possibly through inhalation when staying close to goats or breathing the same air with goats. Equally, a condition locally called ndororo, which was thought to be a fungal infection in humans was described to occur on the feet of people who walk barefoot in cattle slurry or step on raw livestock remains. However, these symptoms were considered not to be harmful to people’s health, as they could be managed with syrups and ointments from drug shops.
Interviews with Ngorongoro District officials and local health personnel revealed that there have been a number of health activities targeting zoonoses. In particular, project activities by a number of academic and research institutions have been taking place in recent years targeting zoonoses especially human brucellosis, through diagnosis and treatment. The health officer believed this has raised the awareness about the disease and others zoonoses among the people who seek medical care at their facilities. The capacity to conduct disease diagnosis in Tanzania is available at the district hospitals. The health officials for Ngorongoro district reported that to a large extent, interventions in terms of both community and hospital based clinical studies, whose results have been published [19,45], have facilitated and built local capacity to readily diagnose and treat brucellosis, which in turn has raised awareness in the community. In the process the awareness of other zoonoses was also raised since the experts do provide health education on the existence of other zoonoses and their pathways while focusing on brucellosis. On the other hand, all health facilities found in villages and neighbouring towns in Kibaha and Bagamoyo district, which are frequently used by the pastoralists from the area, reported not to have the capacity to perform tests for zoonoses such as brucellosis.
Given that the understanding of the concept of a zoonosis was lacking by the pastoralists, it prompted the need to further understand their general perceptions of illnesses and the links to food sources. It was important to understand perceptions of livestock disease to shed light on how they might perceive zoonotic disease possibilities. In all group discussions in both sites and in individual interviews with elders, respondents claimed that they did not fear any disease that affected animals:
None of the locally known livestock health problems were perceived as seriously harmful to humans. A common practice in Maasai tradition for example is to consume the meat from dead or ill livestock. It is believed that no matter what killed the animal in the first place, that the affected animal cannot harm again, [in Maa: meyha engeeya nabo iltung’anak aare]. In other words, the effect of the poison or illness ends in the affected animal and will not go to the person who eats meat of that animal.
This was emphasised when talking about anthrax in cattle. Although it was perceived as a feared illness by a good number of participants, reported practices related to cadavers intensifies the notion of the lack of fear towards meat from dead animals. A local livestock field officer reported to have seen cases of meat been sliced off for consumption from an animal believed to have died of anthrax.
Interviews with livestock keepers revealed that during the dry season migrations in the search of water and pasture in the Ngorongoro District, the consumption of meat from suspected anthrax infected cattle, knowingly or unknowingly, was common. They said that they would dissect the animal after observing that it has not completely decomposed and that they will only consume the meat of body parts that appear to be safe. One elder reported that they can curb the effects of anthrax by consuming medicinal herbs known locally as emporokway ekop by boiling its roots and drinking the juice. They believe it totally eliminates any possibility of contracting anthrax. Another approach, reported to limit effects of suspected anthrax, is to dip a finger in the blood of the animal and touch the finger of blood on the upper part of the mouth. This was normally done by an elder. The act symbolizes stopping the spread of the infection to other people who will consume the meat.
The lack of fear of infected livestock was not limited to the Maasai pastoralists alone in this study. Among the Sukuma pastoralists in Kwala, the researchers observed the consumption of meat from a dead cow confirmed to have contagious bovine pleuropneumonia [CBPP]. The meat was well roasted after it was butchered and then later consumed. Members from neighbouring bomas were invited to partake in the feast including the researchers. When asked about practices of consuming that particular animal and other infected animals they too expressed fearlessness of consuming meat from infected animals.
In contrast to the Maasai, the Batemi FGD participants reported that they are more careful with what to eat although they also asserted that they do not fear to consume meat from suspected animal illness and they thought that an illness transmitted from cattle to cattle can never be harmful to humans. They said they would not eat meat of a dead animal not knowing what killed it in the first place or if it was from a known anthrax case. They also reported that they would not eat body parts that have been affected, although the lack of post slaughter meat inspection services is hindering them to practice this behavior:
While most of the discussions and interviews suggested fearlessness in the consumption of meat from known sick animals, it seems there was a limit to that fearlessness. Fear of eating a dead or sick animal was reported to come from two aspects. First, it may come from the severity of the sickness. It was stated that should an unknown health problem affect and decimate a large number of animals then the perception that it may be harmful or dangerous to humans is agreed. However, if the sickness kills only a few, then there is no fear in consuming the dead animal. To the pastoralists, the death caused unto only few animals suggests that the illness may not be harmful.
Secondly, some Maasai pastoralists reported that they would consider the degree of decomposition of a dead animal before considering it unfit for human consumption. Reporting their experiences, local meat inspectors said that it takes a lot of effort to convince most of the pastoralists that meat from dead or sick animals is not fit for human consumption and needs to be disposed off according to stipulated government regulations. A Maasai pastoralist will mostly agree that a carcass is unfit for consumption if it contains multiple abscess in the liver or ribs:
The pastoralists interviewed reported consumption of raw blood for various purposes such as rites of passage. For the Maasai, they reported that there were different occasions when taking raw blood is practiced. Young initiates consume raw blood after returning from circumcision ceremony. The main belief behind this custom was that the blood replenishes nutrients lost during circumcision. Equally, raw blood is an important meal for women who have given birth and in particular those who are thought to have lost blood due to bleeding. Raw blood is usually mixed with milk and given to the woman. The mixture provides important nutrients in terms of protein and fat that they consider necessary for the body in terms of energy production. Internal organs that can be consumed raw include kidneys, intestines and omasum. However, the custom in the consumption of raw meat was also strongly associated with a common belief about what is fresh and not fresh and the interpretation of freshness considers the time passed since the death of the animal. Blood that is cooked or cooled loses its freshness with time. Freshness of the blood means that the ingredients in the blood are still alive and will be immediately useful to the body. To the Maasai blood is also used as a drink when milk is not available especially in rare occasions like during distant grazing by young warriors known as morani.
The local understanding behind the value of rawness of meat and other internal organs is equally applicable to local knowledge about taking raw milk, a practice well defended despite known possibilities of sources of infections such as TB. Raw milk was reported in all groups as the main staple of the Maasai’s diet. It was also mentioned as important among Batemi and Sukuma groups interviewed but not to the extent emphasized by the Maasai. Raw milk is also the main meal for young boys (Nyangulo) and warriors known as Morani and Koriangas, who partake in migratory seasonal transhumance in search for pasture. It is stored in long gourds and sometimes mixed with special tree herbs known as olirin, oseki, osinoni and ormisigiyo to preserve the aroma of the milk and prevent the milk from contamination:
This was evident during fieldwork with migrating herders in search of pasture and their livestock around the border of NCAA with Serengeti National Park during dry season pasture search. Raw milk is believed to contain important nutrients necessary for survival in these harsh environments and during long trips in herding cattle. The fat in the milk is believed to keep the body warm, the benefits of which are welcomed during the cold seasons. Furthermore, raw milk is used to defeat hunger and believed to stay for a longer time in the stomach than when milk is boiled. Boiling milk is believed contribute to fat loss and making the milk soft. Practically, milk is easy to obtain and can be taken straight from the cow and is perceived as safer than when it is handled and stored in different containers.
An additional angle to understanding perceptions people hold about zoonoses was obtained at the auction markets and slaughter places visited. On the spot interviews of cattle and goat owners, including butcher-men suggested that the requirement of meat inspection is, by and large, perceived as the imposition of authority, conforming to the rules and regulations surrounding meat inspection, rather than a supportive service to prevent harm, such as zoonosis, to the consumer. When the question was asked “Why do you think meat inspection is done?”, a respondent in Chamakweza in Kibaha said it was because the government wanted revenue. This sentiment was echoed at auction markets in Endulen, Wasso, and Loliondo in the Ngorongoro district. Complaints by local animal health officials on the unwillingness on the part of pastoralists for meat inspection before local ceremonies were very common across the sites.
The extent to the practice of ignoring official orders goes very far with some people who will consume meat from animals declared unfit for human consumption. Some customs were reported to be so spontaneous to the extent that seeking services of a meat inspector was unfeasible. For example, slaughtering a goat to visitors, meat feasts by the young morani (warriors) out in the bush and the distance from human settlements. From the discussions and informal interviews, it was obvious that the perceptions of possible infectious diseases or illnesses a person could acquire from consuming meat from infected animals as a threat were not tagged as something serious among the pastoralists. What was important were meanings attached to the practices themselves. An educated local elder and chief said the reason for Maasai not believing the threat of illnesses from animal products, was because meat and milk form their main diet, and in different times [especially during migratory grazing periods] are their only meals. This has been the case for many generations and it was unthinkable to believe that their main and only meal could be harmful to them:
The findings of this study suggest that perceptions of zoonosis are still developing, as there is no agreed conceptualization. There is a widespread absence of the notion of a zoonotic illness and a disbelief that a zoonosis can be transmitted through consumption of animal products, in particular meat and milk, especially among the participants from the two eastern districts. Pastoralists in Ngorongoro were somehow more articulate about the existence of zoonosis than those from Kibaha and Bagamoyo, possibly implying differences in belief systems or awareness about illnesses across the different settings. This understanding has been influenced by people’s own local knowledge and by experiences about livestock and human illnesses, but also their interaction with medical and animal health services.
Knowledge of zoonoses in these communities, as framed in biomedicine, is at best still evolving and is confirmed by pastoralists’ frames of references of what constitutes an infectious disease caused by livestock pathogens crossing into the humans. Pastoralists participating in this study, no doubt believe that there are similarities between some syndromes that occur between animals and humans, but whether these constitute a zoonosis as defined and understood in biomedical sciences remains unclear. This calls for more research, and thereafter communication of findings back to the community. Findings from a previous study conducted in northern Tanzania showed that participants confused zoonotic conditions with other non-zoonotic ailments [8]. This suggests our results may demonstrate the likelihood of increased knowledge within the area. A number of reasons could be attributed to this improved knowledge. This includes the availability of tests especially for brucellosis in the local hospitals, but also research and community studies on zoonoses in the region [19,45].
Despite threats posed by zoonoses to human health, participants expressed little concern. To understand pastoralists’ indifference to animal diseases, especially as displayed by the Maasai, one cannot but embrace their conception of health, illness and causality, which is rooted in their cosmology. As Arhem [46] and Westerland [30] contends, Maasai do not conceptually distinguish between “supernatural” and “natural” illnesses and their concurrent treatments in herbals and ritual medicine, which derive their power from God [47]. Maasai’s knowledge on what causes disease is also different since they do not necessarily employ conventional understanding to categorize illness causative agents into viruses, bacteria, parasites and fungi [27]. Locally, studies about causative agents of tuberculosis among pastoralists and agro-pastoralists in northern Tanzania did not mention microbes as the suspected cause in humans [8,48–50]. The absence of local knowledge on microbial and parasitic causative agents of disease that can survive in both the humans and animals may explain the low understanding of zoonoses in this study including their unrelenting risky practices. This understanding is not limited to pastoralists in Tanzania alone; pastoralists elsewhere do not believe that milk consumption is a significant transmission pathway for pathogens as they too lack knowledge of contagious microbial agents [29].
The results further showed that pastoralists do not conceive the threats of zoonosis unless they have been able to link the biomedical realities, for example, from hospital or veterinary diagnosis with the visible or actual disease. The perceptions may also be muted by indigenous practices, which render the food safe, such as the use of herbs during cooking and eating of meat and milk. Furthermore, populations who experience high exposure levels over long periods, to pathogens in undercooked meat and milk may develop immunity over time. Risk or danger [expressed here in terms of fear] is actually expressed by the desire among pastoralists of the need for physical verification that the animal is in fact infected. It appears that perceived susceptibility to zoonoses becomes a reality only when they can visibly confirm the symptoms of a condition in the animal and the extent of decomposition in the carcass. This conviction calls for a culturally adapted but relevant communication, education and information strategy. Such a strategy can make an impact if it is rolled out over a period of time, and if it is a strategy engaging multiple stakeholders from the regional, district down to the village level.
The understanding of the possibility of an infectious disease being transmitted from animals to human vis-à-vis their known risky practices needs to be approached carefully. For example, previous studies on pastoralist local knowledge of disease transmission have demonstrated how the South Sudan Dinka cattle keepers who did not regard anthrax as transmittable to humans and will therefore dissect a dead carcass, cook and eat it [33]. This practice is done to all animals that die. All the social groups in our study affirmed this practice although the Maasai recognize human infection and a subsequent death will not always be caused by the infection in the animal, which is why they will also consume meat from an infected animal. This practice among the Maasai seems to be a long and ongoing one as it has been documented elsewhere [51].
To appreciate the local knowledge about known and propagated practices of the pastoralists that puts them at risk of acquiring infectious diseases, such as drinking of raw milk, and subsequent known alternatives, also needs further scrutiny. For instance, human subsistence of Maasai pastoralists has mostly been measured through milk availability [52]. The benefits of raw milk are undeniably both culturally and scientifically correct, but the presence of pathogens through unsafe milk preparation or from infected livestock renders it unsafe [53]. However, campaigns on preventing disease transmission such as boiling milk rarely consider these long known benefits of raw milk, which the pastoralists hold dearly and defend. Further, such solutions ignore the cultural and socio-political context upon which pastoralists live [54]. Put differently, if raw milk is considered as food [or a main staple food] by some groups of people, then milk pasteurization may imply going hungry. In his study about Maasai food as symbolism, Arhem [55] portrays how milk, like other foods, is as much culturally constructed as materially produced. Innovative alternatives to milk pasteurization may be needed while preserving its benefits as held in the local context. Although Maasai are changing their diet to consume more grain, milk will remain an important part [56]. This is likely to contribute to continued risk exposure by pastoralists even though they are becoming aware of zoonotic conditions and their transmission pathways. Interestingly, some studies do emphasize the benefits of raw milk in asthmatic persons and as an allergy prophylaxis [57][58]. Therefore, introducing solutions to raw milk in pastoral communities may need to consider such developments since changing behaviors in complex and marginal environments demand sustained efforts over time.
The consumption of undercooked meat and drinking raw blood reported by pastoralists show how just both cultural and social factors in environmental context considerably mediate the impact of infectious agents on humans. The social practice of eating undercooked meat and the drinking of raw blood from goats and cattle have the potential of causing a number of zoonoses. These include tuberculosis, anthrax, and brucellosis. However, these seemingly harmful practices are performed within an established system of life whose function goes beyond a mere consumption of these products in the manner that they do. They maintain order and social cohesion, instill bravery and help build respect between and among different age set groups [55,56]. A discrepancy between government control measures and local realities may hinder long-term community engagement and uptake of health officials’ directives. Although the regular consumption of blood has drastically waned due to, among others, reduction of livestock numbers, drinking raw milk is still part of their main diet. The negative attitude towards the local health officials can perhaps be viewed in part as a resistance to suppression of pastoralist’s way of life. From a different point of view, these actions may be viewed as the government’s way of utilizing disease and disease control as a political and economic vehicle through which tariffs and other taxes or sanction can be applied [59]. But the extent to whether pastoralists currently, will intentionally conduct harmful practices even after some have acquired modern medical and veterinary knowledge will demand further investigation.
The resistance towards associating tuberculosis among pastoralist in coastal districts of Tanzania could be because bovine tuberculosis [the zoonotic TB] was not commonly known about. What seems to be popular was the knowledge that the human form of TB was not transmitted through the consumption of raw milk from cattle. This could also be explained by some studies conducted on in pastoral communities in northern Tanzania, which showed the prevalence of bovine tuberculosis to be extremely low [60]. When put within the pastoralists’ social, economic and cultural context, the resistance could further be explained by the fact that milk is also an important source of income and nutrition and in the hands of the women. Associating it with TB would signal it was not fit and will hinder its sales.
There might be reasons to believe that Maasai pastoralists refused to accept diseases such as TB [especially human TB] because of the stigma it posed in some Maasai communities, where is sometimes associated with HIV/AIDS. A number of recent studies from similar groups and socio-ecological studies documented a gap in knowledge on the perceived causative agents of tuberculosis [48,49,61]. In their study about perceptions of tuberculosis among Maasai of Simanjiro, Haasnoot and colleagues [48] documented multiple reported causes of tuberculosis such as staying for long periods in strong sun, excessive exercise, smoking, promiscuity, breathing in dust and most reported that it was hereditary. It was also noted that God brought TB as a punishment to sinners. In light of this study future research on local knowledge about zoonotic TB should therefore strive to conceptually differentiate on the onset what version of TB people in that community recognize.
The study has shown that the knowledge and perceptions about zoonoses by pastoralists, as understood in the biomedical field, is at best still evolving. Using their own indigenous knowledge frameworks pastoralists understand the possibility of a few infections that could cross over to humans from animals but harm from these are generally tolerated and are not considered as threats. It also showed that perceived risks of zoonoses were overshadowed by local knowledge in cultural practices and the value of animal source food. There are contradictions in perceptions and perhaps knowledge of zoonoses in these settings between local people and professionals. Therefore, to better understand perceptions about zoonoses it is equally pertinent to understand how pastoralists perceive animal diseases and meaning of illness and health management within their socio-cultural context. Evidence is needed to better understand the true burden and impact of zoonoses in these communities and to clearly categorize other risk factors if disease burden is confirmed. Once there is more clarity and common understanding on zoonoses, then interventions will be more effective and accepted. A synergy of scientific work around zoonoses and peoples’ social and cultural practices needs to be carefully tailored to allow for infusion of scientific understanding by pastoralists.
|
10.1371/journal.pcbi.1003206 | Spike Triggered Covariance in Strongly Correlated Gaussian Stimuli | Many biological systems perform computations on inputs that have very large dimensionality. Determining the relevant input combinations for a particular computation is often key to understanding its function. A common way to find the relevant input dimensions is to examine the difference in variance between the input distribution and the distribution of inputs associated with certain outputs. In systems neuroscience, the corresponding method is known as spike-triggered covariance (STC). This method has been highly successful in characterizing relevant input dimensions for neurons in a variety of sensory systems. So far, most studies used the STC method with weakly correlated Gaussian inputs. However, it is also important to use this method with inputs that have long range correlations typical of the natural sensory environment. In such cases, the stimulus covariance matrix has one (or more) outstanding eigenvalues that cannot be easily equalized because of sampling variability. Such outstanding modes interfere with analyses of statistical significance of candidate input dimensions that modulate neuronal outputs. In many cases, these modes obscure the significant dimensions. We show that the sensitivity of the STC method in the regime of strongly correlated inputs can be improved by an order of magnitude or more. This can be done by evaluating the significance of dimensions in the subspace orthogonal to the outstanding mode(s). Analyzing the responses of retinal ganglion cells probed with Gaussian noise, we find that taking into account outstanding modes is crucial for recovering relevant input dimensions for these neurons.
| In many areas of computational biology, including the analyses of genetic mutations, protein stability and neural coding, as well as in economics, one of the most basic and important steps of data analysis is to find the relevant input dimensions for a particular task. In neural coding problems, the spike-triggered covariance (STC) method identifies relevant input dimensions by comparing the variance of the input distribution along different dimensions to the variance of inputs that elicited a neural response. While in theory the method can be applied to Gaussian stimuli with or without correlations, it has so far been used in studies with only weakly correlated stimuli. Here we show that to use STC with strongly correlated, -type inputs, one has to take into account that the covariance matrix of random samples from this distribution has a complex structure, with one or more outstanding modes. We use simulations on model neurons as well as an analysis of the responses of retinal neurons to demonstrate that taking the presence of these outstanding modes into account improves the sensitivity of the STC method by more than an order of magnitude.
| How do neurons encode sensory stimuli? One of the primary difficulties in answering this long-standing problem is the fact that sensory stimuli have high dimensionality. For example, responses of many visual neurons are affected by image patterns that require at least a pixel grid for their description as well as a temporal history spanning multiple time bins or basis functions. Determining what input combinations affect the neural responses is a key step in characterizing the neural computation. Broadly speaking, to detect the presence of certain features in the environment over a range of distances and light conditions, one needs to disambiguate the presence of this feature at a weak contrast from the presence of a similar, but different feature presented at a higher contrast. This can only be achieved with nonlinear functions that depend on multiple input components, such as the presence of an edge of correct orientation and the absence of the edge orthogonal to it [1]. In support of these arguments, the responses of neurons in different sensory modalities are found to be sensitive to multiple input combinations. Examples include vision [2]–[7], audition [8]–[10], olfaction [11], somatosensation [12] and mechanosensation [13]. Neurons respond with all-or-none responses termed spikes. The goal of different methods for characterizing neural feature selectivity is to determine how the probability of eliciting a spike from a neuron depends on its inputs. The underlying assumption is that this dependence of spike probability on input parameters will have a low-dimensional structure. Finding either the linear input dimensions that modulate the spike probability (we will refer to these dimensions as relevant) or quadratic forms of inputs [14]–[16] is the focus of much of the current research in the field.
Much of the analysis of neural selectivity for multiple input combinations has been carried out using uncorrelated (“white noise”) or weakly correlated inputs. With such inputs, the relevant input dimensions can be found using a computationally inexpensive method known as spike-triggered covariance (STC) [6], [7], [17]–[22]. The STC method works by comparing the change in variance along different dimensions in the input space across all stimuli and across stimuli that elicited a spike. The dimensions along which the variance is found to be significantly different represent the relevant input dimensions for the response of a particular neuron. The method is not limited to strictly Gaussian inputs provided that the inputs are still circularly symmetric [23], which is another example of an input distribution without correlations.
In principle the STC method can also be used with correlated Gaussian stimuli [7], [20]. The case of correlated stimuli - especially with strong correlations, where the second moment of the covariance spectrum may be infinite - is important for neural coding. This is because signals in the sensory environment possess such correlations in both the second and higher orders [24]–[30]. Because the properties of a cell's relevant subspace may change depending on the stimulus statistics as a result of adaptation [31], [32], it may not be sufficient to study neural coding using uncorrelated stimuli. Here we show that with strongly correlated inputs, the significance analysis for determining which of the dimensions obtained by the STC method are relevant for neural spiking needs to be modified to take into account a rather complicated covariance structure of randomly selected inputs drawn from such input ensembles. The nonuniform covariance structure, which has properties akin to the graph laplacian in small-world networks [33], breaks the symmetry in the input space, and therefore may obscure many significant dimensions.
The most prominent aspect of the natural scenes covariance structure is the presence of the so-called “coherent” mode [34]. This stimulus dimension approximately corresponds to the zero frequency input component and has a corresponding eigenvalue that is at least times larger than the mean eigenvalue of the input covariance matrix. Even in datasets of fairly large size, the extremely large variance along the coherent mode obscures many of the truly relevant dimensions for neural spiking (Fig. 1). These effects are also reproduced in our analysis of the responses of ganglion cells from the salamander retina probed with -type naturalistic Gaussian stimuli. We identify a close relationship between the covariance structure derived from natural scenes to that defined by the Spiked-Wishart matrix model [35], [36]. This allows us to explain the effects in the context of the STC method using results from random matrix theory, and suggest ways to bypass sampling variability along the outstanding modes.
Mathematically, the first step in the STC method is to compute the covariance matrix of stimuli that lead to a spike and the covariance matrix of all stimuli :(1)(2)Here, is the number of recorded spikes, is the number of stimulus frames, is the value of the stimulus along the th dimension at time , the hat denotes that this stimulus triggered a spike, the bar denotes the average across the input distribution and is the average across the distribution of inputs that triggered a spike (the so called “spike-triggered-average”).
As the second step, one computes the difference between these covariance matrices:(3)and finds the eigenvalues that are significantly different from zero. The corresponding eigenvectors span the neuron's relevant subspace.
To determine statistical significance of the eigenvalues, they need to be compared to the null distribution, which is the distribution of eigenvalues of the matrices . The matrices are formed assuming no association between the stimulus and the neural response, i.e. by using random spike times chosen at the same rate found for real neurons. If the spike train has particular temporal structure (e.g. bursting, a refractory period), the is obtained by random shifts of the spike train with periodic boundary conditions [20]. Significant eigenvalues of can be positive or negative. The procedures for determining statistical significance are detailed in Materials and Methods.
The final step of the STC method is to remove stimulus correlations from the estimate of dimensions found to be significant. This can be done by multiplying them with the (pseudo)inverse of (see Materials and Methods). The method which we use to find the optimal rank of the pseudoinverse matrix is detailed in [22], [37] and for completeness described in Materials and Methods.
We note that this approach, Eqs. (1)–(3), of finding the relevant stimulus dimensions by diagonalizing is equivalent to seeking eigenvectors of the following matrix [20]:(4)This matrix describes a change in the second moment between the distributions stimuli that elicit a spike and that of all stimuli, after subtracting the mean stimulus . Despite the fact that , their eigenvectors coincide.
In another formulation, instead of subtracting the matrix in Eq. (3), the stimulus is decorrelated (“whitened”) prior to its spike triggered characterization [7]. For completeness, the details of this method are brought in Materials and Methods. Throughout the manuscript, we will refer to this method as the “one centered” method, because the null distribution is centered around the identity matrix, rather than a matrix of zeros, as in Eq. (3). Correspondingly, we will refer to the version of the STC method obtained by diagonalizing Eq. (3) as the “zero-centered” method. In essence, both the one-centered and the zero-centered versions are similarly affected by inhomogeneous sampling variability.
The authors of [7] proposed a slightly different definition of the null distribution and a nested hypothesis technique for significance testing. For the model cell simulations we used both significance analysis methods, in both the “zero-centered” and “one-centered” STC formulations, and obtained similar results. For the rest of this paper we will refer to our significance testing method as the “global” one, and focus mainly on the “zero-centered” formulation of the STC method. Using this combination the important effects of the strong stimulus correlations on the analysis are more easily understood.
We begin with an illustration of the problems that arise when the STC method is used to analyze neural responses to strongly correlated Gaussian noise (Fig. 1). We simulated a model neuron where the neuronal responses were modulated by stimulus projections onto a single dimension (termed here the relevant feature). The stimuli were constructed to match the second-order statistics from the set of images in the van Hateren dataset [38] (see Materials and Methods). In this example obtained for dataset of a moderate size, no eigenvalues fell outside of the % confidence intervals (% significant bounds for the largest and smallest rank-ordered eigenvalues). Yet, the spike train contains enough signal about the cell's input-output function to identify the relevant feature for this level of significance. Specifically, the variance along the relevant dimension in the spike-triggered stimulus () is much smaller than can be explained by random spike times (Fig. 1E).
To understand the origin of such masking of the relevant feature(s), we consider the eigenstructure of covariance matrices for stimulus ensembles with strong pairwise correlations. For example, in the case of natural scenes that exhibit long range correlations over a very wide range of spatial scales [27], [39], principal component analysis (PCA) yields one outstanding eigenvalue (for example, see eigenvalue marked in Fig. 2A). The corresponding eigenvector has all positive components [28], [30] and is often referred to as the “coherent mode” [34]. To understand why such a coherent mode appears, one can consider the case where the correlations decrease only slightly over the range of image patches used to compute the covariance matrix. In this case, the correlation values in different image patches will be approximately the same. Such a matrix will have one outstanding eigenvalue with a corresponding eigenvector that has equal weights for all stimulus dimensions [40]. Small differences in the amount of covariation for pixel pairs with different spatial separation will lead to deviations in components of the coherent mode from each other, but the basic structure will remain the same as long as the mean of the correlation values exceeds the standard deviation of their fluctuations [40]. In fact, shuffling entries in the sample covariance matrices of natural stimuli yields matrices whose spectra follow the analytical predictions exactly [40], [41]. These analytical predictions generalize the Wigner semicircle law [42] for matrices whose elements have a non-zero mean:(5)where and are the mean and variance of matrix elements. The distribution follows the semicircle law with the addition of one outstanding mode that appears once the mean of matrix elements exceeds their standard deviation. The eigenvector corresponding to the outstanding eigenvalue is . The semicircle law appears because matrices are no longer positive-definite after shuffling. However, the outstanding eigenvalue is located at exactly the same value as the outstanding eigenvalue of the natural scenes covariance matrix (see Fig. 2C).
In our analysis of the van Hateren database, the largest eigenvalue tends to be at least times larger than the second largest eigenvalue. This shows how strong the coherent mode is compared to other modes. The principal components ranked below the coherent mode form a collection of orthogonal “edge detectors”, some of which correspond to an eigenvalue still much larger than the mean eigenvalue of , a signature of the stimulus' heavy-tailed covariance spectrum. Such large disparities in variance along the different dimensions in the stimulus space make it problematic to directly compare changes in variance induced by the observation of spikes along these different dimensions.
The detailed structure of sampling variability in the estimation of eigenvectors and eigenvalues can be understood in terms of the Spiked Wishart ensemble [35], [36]. In the Spiked Wishart matrix model, the true (population) covariance eigenvalues are all equal to one, except for a small number of outstanding modes with eigenvalues larger than one , where is the stimulus dimensionality. The distribution of sample covariance eigenvalues for a finite number of inputs has a positive bias, with the following analytical expressions [36]:(6)(7)(8)where and is the number of samples. The distribution representing the “bulk” of eigenvalues is the so called Marčenko-Pastur distribution given by:(9)(10)This distribution corresponds to the sample covariance eigenvalues obtained when the true covariance is the identity matrix. Using numerical simulations we verified that, although the Spiked Wishart ensemble is only an approximation to the covariance matrices derived from natural stimuli, Eqs. (7) and (8) accurately describe the scaling of the variance and the mean of sample eigenvalues as increases.
In addition to biases in eigenvalue estimates, there are also biases in the estimation of eigenvectors. The dot product between the true (population) th eigenvector and the th eigenvector of the sample covariance approaches(11)
In other words, the “mixing” of the outstanding sample eigenvectors seen in Eq. (11) (note the dependence of this mixing on through ) as well as the variance and bias in the sample eigenvalues seen in Eq. (7) means that whitening cannot be exact.
In the context of the spike triggered covariance the consequences of such properties of the distribution of sample eigenvalues are twofold. First, Eq. (8) indicates that the variance of the outstanding eigenvalues around their mean increases with the square of their value and is inversely proportional to the number of samples. Thus, for sample sizes that are not much larger than the stimulus dimensionality ( in the simulation results presented in Fig. 3A), the increased variance of the outstanding sample eigenvalue means that and will not cancel each other exactly along that vector. Second, the mean estimate contains a positive bias relative to the population values, cf. Eq. (7). The combination of these two effects widens the null-distribution used to test the significance of the resulting eigenvectors, effectively masking features that should otherwise be identified as being relevant.
One way to compensate for the symmetry breaking effects caused by strong correlations in the input space is to equalize variances before applying the STC method. This is the essence of the “one-centered” formulation of the STC method [7]. In principle, this “whitening” should work with Gaussian stimuli with any covariance structure. However, as discussed above, in the case of strongly correlated stimuli, the estimation of eigenvalues (i.e variances along different dimensions in the input space) possesses strong variability, cf. Eq. (7). As a consequence, normalization by a variance estimated from one part of the dataset does not fully remove correlations in a different subset of the data. With increasing dataset size, the estimate of the variance along the coherent mode improves. However, because the absolute value of variance is not relevant in the pre-whitening method, dimensions with smaller variance can cause just as much contamination as the coherent mode. In addition, the estimation of variance along dimensions corresponding to just larger than remains poor for large . If the sample eigenvalue estimation error diverges as , as follows from Eq. (8). In other words, as the number of samples and increase, the bulk of the distribution narrows, and new eigenvalues separate from the bulk. It is these eigenvalues with intermediate values that are poorly determined and make it problematic to equalize variance along different dimensions. Another signature of this phenomenon is that for these dimensions, as follows from Eq. (11). Thus, these dimensions are poorly estimated from the sample covariance and, as a consequence, the variance along one stimulus dimension in the training set will be inappropriately used to normalize variance along a different stimulus dimension in the test set. Altogether, we observed that pre-whitening stimuli did not improve the estimation of relevant stimulus features compared to the zero-centered method, compare panels A and B in Fig. 3. Intuitively, in the zero-centered method the dimensions with the largest variance provide the largest uncertainty in variance estimation, whereas in the one-centered version the problematic dimensions change depending on the dataset size, and are not easily identified a priori.
We have also explored the possibility of using a pseudoinverse of the covariance matrix instead of the full inverse to normalize variance along different dimensions (see Materials and Methods for details). When using the pseudoinverse (instead of the inverse), stimulus dimensions with small variance in the stimulus ensemble are removed to avoid noise amplification along these dimensions (see Materials and Methods for details). However, an immediate consequence of choosing a small pseudoinverse order is that the stimulus dimensionality is reduced to . This implies that the effective of the problem is now , i.e. times larger than . This could work well in some cases as illustrated in Fig. 3I. Here, in simulations based on a small number of spikes, the use of pseudoinverse can help recover one or two significant features while the standard zero-centered method fails to find any. However, the use of pseudoinverse only helps within a very narrow band of small pseudoinverse orders. This band may be difficult to determine when analyzing real neural data. In addition, this procedure limits the reconstruction to a linear combination of only a few leading stimulus dimensions. In many cases, the relevant features do include components along stimulus dimensions with smaller variance, and in those cases, the effective increase in will not improve the performance of the STC method. Indeed, one observes that in cases where two significant dimensions are obtained by using substantial reduction in dimensionality of the pseudoinverse, the resulting dimensions have the subspace projection onto the model features of whereas this value is when using the full inverse and a larger number of spikes to obtain for a comparable effective (Fig. 3I).
Finally, in the regime where (i.e. “almost full” inverse), the prewhitening approach works just as well as the “zero-centered” formulation, and a relatively high value of the signal-to-noise ratio parameter is required for recovery of the full relevant subspace.
As another way to compensate for the symmetry breaking effects caused by strong correlations in the input space, we propose to modify the “zero-centered” formulation of the STC method in the following way. Because the largest drop in variance is between the coherent mode and other dimensions, we propose here to test the significance of changes in variance separately along the coherent mode and in the subspace orthogonal to it. Explicitly, to do the analysis in the dimensional subspace, the coherent mode is projected out of all stimuli. If is a stimulus vector and normalized to length , one can perform the STC analysis using instead of where:(12)
In this approach the correct number of relevant dimensions is determined by evaluating significance in the subspace orthogonal to the coherent mode and then adding back their projections on the coherent mode from the corresponding eigenvectors evaluated in the full input space (see below).
We find that considering the coherent mode separately from the rest of stimulus dimensions reduces the value of for which the full relevant subspace is found to be significant by a factor of (Fig. 3C). This improvement can be approximated from Eqs. (7) and (8). Assuming the cell's relevant subspace is exactly orthogonal to the coherent mode, the extremal values of the null distribution are distributed as . The variance of is:(13)This implies that the number of stimuli sufficient for identifying the relevant features as significant increases with as:(14)Upon removal of the coherent mode, the minimum value of for which the signal to noise ratio will be high enough to identify the relevant dimensions scales as corresponding to the stimulus' second principal component. Therefore the improvement is proportional to . In our simulations (Fig. 3A,C) this corresponds to a predicted fold improvement. Given that our model features were not exactly orthogonal to the coherent mode, and that the spectrum obtained from the van Hateren dataset has a heavy tail and does not conform exactly to the Spiked Wishart ensemble, an approximate fold improvement represents a good agreement with the prediction.
It is noteworthy that the minimum requirement on the dataset size for obtaining the correct number of relevant dimensions is actually smaller with correlated stimuli than it is for white noise stimuli for the same neuron (compare Fig. 3 panels A–D) when the model parameters were matched such that the firing rate remains constant across different stimuli statistics. Another important point is that considering the coherent mode separately is different from simply discarding a “DC-like” component that could be found to be significant by the STC. This is because when is small, no dimensions are found to be significant with the coherent mode as part of the stimulus ensemble (Fig. 1).
An important consideration is that the final analysis can include the components of the relevant dimensions onto the coherent mode. This is possible for two reasons. First, the coherent mode does not represent an arbitrary dimension in the input space but is one of the eigenvectors of the sample covariance matrix. Second, the significant eigenvectors of have a form , where is the th eigenvalue corresponding to the th eigenvector of the sample covariance matrix, and describes one of the relevant features [20]. Because of these two properties, eigenvectors evaluated in the full input space and in the subspace orthogonal to the coherent mode differ only in their components along the coherent mode (see Materials and Methods for the details of the derivation). This makes it possible to analyze cells with features that have nonzero components along the coherent mode.
We have verified that this approach also works in a large number of cases where the relevant stimulus dimensions have a large projection on the coherent mode (Fig. 4). One concern is that when such neurons are probed with a relatively small number of stimuli, then projecting the coherent mode out may “push” the relevant feature into the null eigenvalue distribution. This does not appear to be a problem in our simulations for (Fig. 4B). If this does happen, the relevant subspace should be the one spanned by both the eigenvectors found to be significant in the full stimulus space and those found to be significant in the subspace orthogonal to the coherent mode.
We now demonstrate the importance of this correction scheme by analyzing recordings of 22 salamander retinal ganglion cells (RGCs). These neurons were probed with a correlated noise stimulus whose covariance matrix was matched to that of natural visual stimuli. Without correcting for the presence of the coherent mode, the STC analysis yielded no significant dimensions for a third of the cells, and very few for the rest (Fig. 5). This happens because the eigenvalue corresponding to the coherent mode injects large eigenvalues into the null eigenvalue distribution (as seen in Eq. (8)), thus masking the cell's true relevant features. Following the correction, the number of significant dimensions per cell increased from to (see Fig. 5A for the full population values). The dimensionality of the relevant subspace increased for 21 out of 22 cells. For one cell, we were unable to find a significant dimension either before or after the correction of the method. The distributions of null eigenvalues used to determine which of the eigenvectors of are significant (Fig. 5B,C) became much more narrow when evaluated in the subspace orthogonal to the coherent mode.
The goal of this work was to extend the range of applicability of a computationally simple method of spike-triggered covariance to strongly correlated stimuli. While the STC method in principle can be used with strongly correlated Gaussian stimuli, our results show that the inhomogeneous sampling variability can in practice make it difficult to recover the correct relevant subspace. We have characterized the effects generated by strong Gaussian correlations using simulations of two model neurons in a wide range of dataset sizes (which could also be viewed as an inverse measure of the neuron's level of internal noise). Results from random matrix theory, and specifically the Wigner and Spiked Wishart ensembles, suggest that the origin of these issues can be traced to the estimation bias and variance of covariance matrices with vastly different eigenvalues. We demonstrate that by considering the coherent mode, which corresponds to the largest eigenvalue, separately from the rest of stimulus dimensions, one can improve the method's sensitivity by .
One qualitative lesson offered by these analyses is that while the bulk of the eigenvalues of is a good proxy for the width of the null distribution in the case of white noise inputs, but not in the case of strongly correlated inputs.
Furthermore, our analysis suggests that sampling variability along the secondary outstanding modes corresponding to the next few principal components may have similar masking effects to the ones reported here for the coherent mode. Possible solutions to the full problem may include performing a sequence of analyses in subspaces of decreasing dimensionality, orthogonal to several leading principal components. However, the payoff from this procedure is (at most) of order which in our case is . At the same time, one runs the risk of losing the ability to resolve the remaining dimensions because of the reduced signal to noise ratio.
Another potential solution is to correct for the estimation bias and variance in eigenvalues and eigenvectors, described by Eqs. (7) and (8). However, this procedure is difficult computationally and in most cases can only be done for simple eigenvalue distributions [43].
The treatment of the artifacts caused by a large coherent mode present in the data has been previously discussed in analyses of stock-markets [44], [45], evolution of proteins [46], and Human Immunodeficiency Virus (HIV) mutations [34]. In these cases, the extra dimension was removed and the resulting covariance structure was compared against the Marčenko-Pastur eigenvalue distribution that assumes no correlation between the variables and uniform variable variances. The case of reverse correlation experiments discussed here is different from these analyses because the spike triggered ensemble is compared to the full stimulus distribution. In addition, our analyses provide two important novel contributions. First, we show there is a crucial difference between discarding the coherent mode and projecting it out. This is because of the way the coherent mode injects noise into the null distribution. Second, the approach described here also permits the inclusion of the components of the relevant dimensions along the coherent mode in the final results. We hope that the ideas for treating the coherent mode presented here will also be relevant in other areas of computational biology.
Experimental data were collected using procedures approved by the Institutional Animal Care and Use Committee of Princeton University, and in accordance with National Institutes of Health guidelines. Experimental and surgical procedures have been described previously [47].
Each stimulus frame was randomly drawn from a multivariate Gaussian distribution with zero mean and covariance matrix , In the correlated stimulus case, the population covariance was computed from the covariance of pixels patches from the van Hateren image database [38] (with no downsampling). In the uncorrelated (“white”) case, was the identity matrix.
We describe two approaches for determining significance of candidate features that were previously described in the literature: global and nested. When applied to our datasets, both of the approaches yielded similar results.
Within the STC method, stimulus correlations need to be removed from the estimates of eigenvectors obtained by diagonalizing matrix . This correction is needed, because the eigenvectors of have a form , where describe components of one of the relevant features [20]. As described above, one may wish to use a pseudoinverse, instead of the full inverse of the matrix to minimize noise amplification at higher spatial frequencies. Assuming that the eigenvalues are ordered to be monotonically decreasing, the pseudoinverse of order is given by(18)
In the analysis of data from retinal ganglion cells, the optimal order of the pseudoinverse was determined in the following way. The dataset was divided into the training and test sets. The features were computed by diagonalizing the matrix , cf. Eq. (3), in either the full input space or in the space orthogonal to the coherent mode using the training set. Following that, the optimal pseudoinverse order was selected as the one that yielded decorrelated features that convey the most information about, or give the largest predictive power for, the neural response. Explicitly,(19)(20)where is the probability distribution of the projections of stimuli onto the significant eigenvectors (), decorrelated by . are the decorrelated significant features, and:(21)(22)
As an alternative to removing stimulus correlations from the eigenvectors of , one can remove stimulus correlations from each of the stimulus vectors, prior to the diagonalization of , a procedure that is known as pre-whitening [7].
The sample stimulus covariance matrix from Eq. (2) can be written in terms of eigenvalues and eigenvectors as(23)We can now define a matrix . Then, the analogue of in the “one-centered” formulation is given by:(24)This procedure is equivalent to whitening each of the stimulus frames independently (by multiplying it with ) and then computing the spike-triggered covariance.
In the limit of infinite data, the null hypothesis corresponds to . In this case . For a dataset of finite size, the null distribution is computed from many realizations of the matrix(25)where is defined by Eq. (16). The eigenvalues of (most of which are close to ) can then be compared to the null eigenvalue distribution, using either the nested or global comparison tests described above.
In Fig. 3 we analyzed the simulated spike trains using every pseudoinverse order of . The prewhitening is then done using this matrix Eq. (18) instead of the full rank matrix .
Performing the pre-whitened STC analysis using all pseudoinverse orders is equivalent to testing models. Therefore, the confidence interval of the null distribution should be adjusted from the percentile range to , where is the Dunn-Šidák correction:(26)
We recall that according to Ref. [20], the significant eigenvectors of can be written as(27)
Thus, the eigenvectors of represent a sum of projection operators onto the principal components of the stimulus ensemble. When we perform the STC method in the subspace orthogonal to the first principal component of the stimulus, the eigenvectors of can be written as(28)(the coherent mode is exactly the vector ).
Comparing expressions for the eigenvectors of and , one observes that there is a one-to-one correspondence between them. This correspondence can be identified based on proportionality in components along second, third, and other principal components:(29)for any . In sum, once the eigenvector is found to be significant in the subspace orthogonal to , the eigenvector that should be identified as significant in the full stimulus space is that satisfies the condition of Eq. (29).
The nonlinearity was chosen to be a logistic function because such functions maximize the cell's noise entropy and thus minimize the assumptions imposed on the cell's response [48]. Using the models, we generated simulated spike trains in response to either a white or a correlated noise stimulus.
The model used in Fig. 3 (model “”) had a two dimensional relevant subspace with features orthogonal to the coherent mode. The probability of spiking was modeled to increase when the projection of the stimulus on either of the preferred features was large in absolute value (representing a logical OR function). If are the preferred model features and is the stimulus presented at time (here, the 's and are dimensional vectors) then the probability of a spike at time is:(30)where and are parameters that determine the width and (soft) thresholds of the sigmoid nonlinearities for the model.
We have also considered the case where the projection of the stimulus on the features was not taken in absolute value, corresponding to a monotonic nonlinearity. In that case (model “”, used in Fig. 1) the model was one dimensional, so the probability of a spike is(31)
The effects described above were observed for both symmetric (Fig. 3) and monotonic (Fig. 1) nonlinearities.
The second model had one relevant input feature with a large component along the coherent mode. In this case, the probability of a spike was modeled as:(32)where and are the width and the threshold of the sigmoid nonlinearity of this model.
In units of the standard deviation of the projection of the stimulus on the model features (, ) the model parameters were chosen to be:(33)(34)(35)
The overlap measure we use when the dimensionality of the relevant subspace is greater than one is given by [49]:(36)where and are matrices that hold the model and computed features, respectively, is the input dimensionality, and is the number of relevant features in the model.
|
10.1371/journal.pgen.1007753 | Evidence that regulation of intramembrane proteolysis is mediated by substrate gating during sporulation in Bacillus subtilis | During the morphological process of sporulation in Bacillus subtilis two adjacent daughter cells (called the mother cell and forespore) follow different programs of gene expression that are linked to each other by signal transduction pathways. At a late stage in development, a signaling pathway emanating from the forespore triggers the proteolytic activation of the mother cell transcription factor σK. Cleavage of pro-σK to its mature and active form is catalyzed by the intramembrane cleaving metalloprotease SpoIVFB (B), a Site-2 Protease (S2P) family member. B is held inactive by two mother-cell membrane proteins SpoIVFA (A) and BofA. Activation of pro-σK processing requires a site-1 signaling protease SpoIVB (IVB) that is secreted from the forespore into the space between the two cells. IVB cleaves the extracellular domain of A but how this cleavage activates intramembrane proteolysis has remained unclear. Structural studies of the Methanocaldococcus jannaschii S2P homolog identified closed (substrate-occluded) and open (substrate-accessible) conformations of the protease, but the biological relevance of these conformations has not been established. Here, using co-immunoprecipitation and fluorescence microscopy, we show that stable association between the membrane-embedded protease and its substrate requires IVB signaling. We further show that the cytoplasmic cystathionine-β-synthase (CBS) domain of the B protease is not critical for this interaction or for pro-σK processing, suggesting the IVB-dependent interaction site is in the membrane protease domain. Finally, we provide evidence that the B protease domain adopts both open and closed conformations in vivo. Collectively, our data support a substrate-gating model in which IVB-dependent cleavage of A on one side of the membrane triggers a conformational change in the membrane-embedded protease from a closed to an open state allowing pro-σK access to the caged interior of the protease.
| Regulated Intramembrane Proteolysis is a broadly conserved mechanism for transducing information across lipid bilayers. In these signaling pathways a protease on one side of the membrane triggers the activation of a membrane-embedded protease that cleaves its substrate within or adjacent to the cytoplasmic face of the membrane. Site-2 metalloproteases (S2P) are the most commonly used intramembrane cleaving proteases in these pathways but the mechanism by which cleavage on one side of the membrane triggers intramembrane proteolysis remains poorly understood. Here, we provide evidence for a substrate-gating model in which an extracellular signaling protease triggers a conformational change in a S2P family member from a closed to an open conformation allowing its substrate access to the catalytic center of the enzyme.
| Regulated Intramembrane Proteolysis (RIP) is a broadly used strategy to transduce information across lipid bilayers [1–3]. In many of these RIP pathways, proteolysis on one side of the membrane by a so-called site-1 protease leads to the activation of a membrane-embedded site-2 protease that cleaves and activates a substrate on the other side of the membrane, ultimately leading to the activation of gene expression. Although the signaling (site-1) and intramembrane cleaving (site-2) proteases have been identified in many of these pathways, how intramembrane proteolysis is regulated at the molecular level remains poorly understood.
The zinc metalloproteases of the Site-2 Protease (S2P) family are among the most commonly used intramembrane cleaving proteases in RIP signaling pathways [4–6]. This family is composed of four subfamilies that share a conserved catalytic core [7]. Well-characterized members include the mammalian Site-2 Protease (S2P) [8] and E. coli RseP [9, 10], both from subfamily 1, and the B. subtilis sporulation protease SpoIVFB (B) [11–13] from subfamily 3. There are currently no characterized members of subfamilies 2 and 4. Mammalian S2P is involved in signaling pathways that sense and respond to intracellular levels of sterols and misfolded proteins in the endoplasmic reticulum [5]. In both cases, the canonical two-step protease cascade results in intramembrane proteolysis of membrane-anchored transcription factors and their release into the cytoplasm, translocation into the nucleus, and activation of gene expression. E. coli RseP and orthologs in diverse bacterial phyla are involved in envelope stress response pathways [4, 6]. These pathways commonly involve extracytoplasmic function (ECF) sigma factors that are held inactive by membrane-anchored anti-sigma factors. Specific envelope stresses trigger site-1 cleavage of the anti-sigma factor on the extracytoplasmic side of the membrane followed by RseP-mediated intramembrane proteolysis and activation of the ECF sigma factor. B. subtilis SpoIVFB (referred to as B, for simplicity) is involved in cell-cell signaling during the morphological process of sporulation [14] and is the subject of this study.
In response to starvation, B. subtilis differentiates into a dormant spore [15–17]. The first morphological event in this process is the formation of a polar septum that divides the sporulating cell into a large mother cell and smaller forespore. Shortly after division, the mother cell membranes migrate around the forespore generating a cell within a cell (Fig 1A). The mother then packages the forespore in protective layers while the spore prepares for dormancy. Upon spore maturation, mother cell lysis releases the stress-resistant spore into the environment. Throughout this morphological process, the mother cell and forespore follow distinct programs of developmental gene expression that are linked to each by cell-cell signaling pathways [18, 19]. The final communication between the forespore and mother cell involves a RIP signaling pathway in which an inactive mother cell transcription factor (pro-σK) is proteolytically activated by the S2P protease family member B in response to a signal from the forespore [11, 12, 14]. The B protease is produced in the mother cell at an earlier stage in sporulation and localizes to the mother-cell membranes that surround the forespore [11, 12, 20]. B is held inactive at this site by two membrane proteins SpoIVFA (A) and BofA [14, 21] (Fig 1A). Upon the completion of engulfment, a site-1 signaling protease called SpoIVB (IVB) is produced in the forespore and secreted into the space between the two membranes that separate mother cell and forespore where it cleaves the extracytoplasmic domain of A [22–25] (Fig 1A). This cleavage relieves inhibition imposed on B, triggering pro-σK processing and the activation of late mother cell gene expression. Thus, in this RIP pathway, the site-1 signaling protease (IVB) cleaves a negative regulator of the S2P family protease (B). How A and BofA hold B inactive and how cleavage by IVB triggers relief of inhibition remain unclear.
The mechanisms by which site-1 cleavage activates intramembrane proteolysis by S2P family members are largely unknown. The pathway for which regulation is most well understood is the one that employs E. coli RseP. Site-1 cleavage of the periplasmic domain of the anti-sigma factor RseA triggers RseP-mediated intramembrane proteolysis of RseA and activation of the ECF sigma factor σE [4]. RseP and S2P subfamily 1 members contain extracytoplasmic PDZ domains [7, 26]. In vitro studies have found that the interaction between this domain and the exposed C-terminal residue of RseA generated by site-1 cleavage is critical for intramembrane proteolysis [27]. However, in vivo analysis indicates that this is likely to be only a part of the story [28]. In vivo, the PDZ domain on RseP appears to function as a size-exclusion filter to prevent substrates with large extracytoplasmic domains from gaining access to the active site of the membrane protease. In addition, a membrane-reentrant β-loop in RseP has been shown to participate in the recognition of the substrate transmembrane helix and its destabilization for presentation to the protease active site [29]. The absence of a high-resolution structure of a member of this subfamily has prevented a complete mechanistic picture of this regulatory pathway.
The only published S2P structures come from Methanocaldococcus jannaschii (mjS2P) [30], a member of subfamily 3. Unlike S2P and RseP, the proteases in this subfamily lack extracytoplasmic PDZ domains. Subfamily 3 members contain six N-terminal transmembrane helices that make up the membrane protease domain followed by a C-terminal cystathionine-β-synthase (CBS) domain [7, 31, 32]. The mjS2P structures comprised the N-terminal protease domain that shares 29% identity (54% similarity) with the membrane protease domain of the B. subtilis B protease. Two molecules of mjS2P were present in the asymmetric unit of the crystal forming an antiparallel pseudo-dimer. One of the monomers was in a "closed" conformation in which the active site was surrounded by transmembrane helices and thus inaccessible to substrate. The other protomer adopted a relatively open conformation, in which the first and sixth transmembrane segments were laterally displaced by 10–12 Å, generating a deep groove that runs the length of the molecule. Based on these findings, it was hypothesized that mjS2P could transition between closed and open states to allow substrates access to the active site of the protease, however the biological relevance of these conformations has never been established. Here, using the B. subtilis sporulation RIP pathway, we investigate this substrate-gating model. Our analysis provides evidence that A and BofA hold the membrane protease B in a closed conformation and IVB cleavage triggers a shift to the open conformation allowing pro-σK access to the interior of the protease.
In previous studies we used co-immunoprecipitation to investigate the interaction between B and its two negative regulators A and BofA [22, 33]. To prevent activation of the membrane-embedded protease, these experiments were performed in strains lacking the site-1 signaling protease IVB. Although we could efficiently recover the processing complex, we never detected pro-σK in the immunoprecipitates. Prompted by the two structural conformations of the mjS2P protease domain [30], we wondered whether A and BofA hold B in a closed conformation preventing an interaction with pro-σK and potentially IVB signaling could generate an open conformation allowing stable association between protease and substrate. To investigate this possibility, we repeated the co-immunoprecipitations comparing strains with and without IVB. To prevent pro-σK processing and release of the mature transcription factor, the B protease contained the catalytic mutant E44Q [11, 12]. In addition, B(E44Q) was fused to GFP, which we previously found stabilizes the protease from degradation upon IVB signaling [33]. The strains were induced to sporulate and harvested at hour 4 of sporulation. A crude membrane preparation was solubilized with the non-ionic detergent digitonin and B(E44Q)-GFP was immunoprecipitated with anti-GFP antibody resin. The immunoprecipitated material was then analyzed by SDS-PAGE and silver staining.
As reported previously [22, 33], we efficiently immunoprecipitated B-GFP and its regulator A (Fig 1B). BofA is a small (9 kDa) protein and is likely present in the dye front (S1A Fig). In support of the idea that the processing complex is anchored in the mother cell membranes that surround the forespore by SpoIIIAH (AH) and SpoIIQ (Q) [34], these proteins were also present in the immunoprecipitates (Fig 1B and S1 Fig). Furthermore, consistent with published work showing that the Q protein is a substrate of the IVB signaling protease [22, 35], the level of co-precipitated Q was lower from the IVB+ strain. For unknown reasons, the levels of co-precipitated A were similar in the presence and absence of IVB, despite being a substrate of the signaling protease. Importantly, a protein band similar in size to pro-σK (27 kDa) was present in the immunoprecipitate from the IVB+ strain but absent from the ΔIVB mutant (Fig 1B). This protein was also absent in an immunoprecipitate from a IVB+ strain that lacked sigK (Fig 1B). To determine whether this protein was indeed pro-σK, we excised the band from the silver-stained gel and the corresponding region from the ΔIVB and ΔsigK lanes, digested the proteins with trypsin, and analyzed the products by Mass Spectrometry. Three unique peptides (see Methods) corresponding to pro-σK were identified in the immunoprecipitate from the IVB+ strain while none was detected from the ΔIVB and ΔsigK immunoprecipitates. Finally, immunoblot analysis using anti-σK antibodies identified pro-σK specifically in the immunoprecipitate from the IVB+ strain (Fig 1B). Collectively, these data suggest that stable association between pro-σK and the B-A-BofA processing complex requires the IVB signaling protease.
To investigate whether a similar phenomenon could be observed in vivo, we used fluorescence microscopy, taking advantage of a functional fusion between pro-σK and the cyan fluorescent protein (CFP) (S2E Fig). To simultaneously visualize the B membrane protease while also protecting it from degradation, we fused B(E44Q) to YFP. Importantly, a fusion of the wild-type B protease to YFP supported efficient sporulation (S2E Fig) and processing of both pro-σK and pro-σK-CFP (S2A and S2B Fig). Strains harboring these fusions were induced to sporulate and then monitored by fluorescence microscopy over a sporulation time course (Fig 2 and S3 Fig). Previous immunofluorescence microscopy and cell fractionation studies indicate that pro-σK non-specifically associates with membranes in vivo [36]. However, in sporulating cells lacking the IVB signaling protease, pro-σK-CFP appeared to localize in the mother cell cytoplasm with weak enrichment on the nucleoid, presumably due to non-specific interactions with the chromosome. We suspect pro-σK-CFP transiently associates with the lipid bilayer but this interaction is too short-lived to be detected in live cells. By contrast and as expected, in IVB+ cells harboring a wild-type B-YFP fusion, mature σK-CFP co-localized with the mother-cell nucleoid (Fig 2B) consistent with its proteolytic processing (S2B Fig) and association with core RNA polymerase [36]. Importantly, in >80% of the sporulating cells harboring wild-type IVB and the B(E44Q) catalytic mutant, pro-σK-CFP accumulated in the mother cell membranes surrounding the forespore (Fig 2, S3 and S4 Figs). This localization pattern provides further evidence that stable association between pro-σK and its processing complex depends upon IVB. Finally, pro-σK-CFP failed to localize in the outer forespore membrane in a strain harboring a catalytic mutant (S378A) of IVB [25] (Fig 2 and S4B Fig) indicating that IVB protease activity is required for the stable interaction between pro-σK and the signaling complex.
B, A, and BofA reside in a multimeric membrane complex in the outer forespore membrane [33] (Fig 1A). To investigate which of these factors tethers pro-σK to the complex, we analyzed pro-σK-CFP localization in sporulating cells lacking each of these components. In all cases, these strains possessed an intact IVB signaling protease and, other than the ΔB mutant, the B(E44Q)-YFP fusion. As previously reported [33, 37], in sporulating cells lacking A, BofA, or both, the B protease was no longer exclusively anchored in the outer forespore membrane and instead was distributed in all mother cell-derived membranes (Fig 3 and S5 Fig). In these mutants, pro-σK-CFP was detectable in the membranes surrounding the forespore (Fig 3 and S5 Fig). The pro-σK-CFP signal was reduced compared to wild-type presumably due to the lower levels of B(E44Q) present in the forespore membrane. However, the percentage of sporulating cells with forespore-associated pro-σK-CFP was similar to wild-type (S4C Fig). Surprisingly, pro-σK-CFP did not appear to localize in the peripheral membranes of the mother cell. The results of experiments described in the next section suggest that some of the pro-σK-CFP fusion protein is present in these membranes but the signal is too weak to generate a membrane signal. Finally, in cells lacking the B protease, pro-σK-CFP localization phenocopied the IVB null and was largely present in the cytoplasm with some enrichment on the nucleoid (Fig 3, S4C and S5 Figs). Taken together with the data presented in Figs 1 and 2, these experiments suggest that pro-σK specifically associates with the membrane-embedded protease and does so in a IVB-dependent manner.
To determine whether additional sporulation-specific proteins were required for the association between B and pro-σK, we engineered strains to express pro-σK-CFP and B(E44Q)-YFP under IPTG control during vegetative growth. Exponentially growing cells were analyzed by fluorescence microscopy in a time-course after the addition of IPTG (Fig 4 and S6 Fig). In cells lacking the B protease, pro-σK-CFP appeared cytoplasmic (Fig 4A and S6 Fig). Co-expression of pro-σK-CFP and wild-type B-YFP in the absence of its negative regulators A and BofA resulted in the production of processed σK-CFP (Fig 4B) that co-localized with the nucleoid (Fig 4A and S6 Fig). In cells expressing the B(E44Q) catalytic mutant, pro-σK-CFP appeared cytoplasmic with some localization at division septa and along the cytoplasmic membranes. Although the membrane signal was weak, we note that the pro-σK-CFP signal in cells that were in contact with each other along their length had no gap in fluorescence between them as compared to the gaps observed for cytoplasmic pro-σK-CFP in cells lacking the B protease (Fig 4A). We further note that the pro-σK-CFP signal that appeared to be cytoplasmic in the B(E44Q) mutant was weaker and more pixelated than the signal observed in the cells lacking B, even though the levels of pro-σK-CFP were similar in the two strains (Fig 4B). Both phenomena are consistent with the association of pro-σK-CFP with B(E44Q) at the cell membrane. Importantly, pro-σK-CFP remained largely full-length with relatively little free CFP liberated from the fusion (S2C Fig). The cytoplasmic pro-σK-CFP signal was similarly weaker and more pixelated in sporulating cells lacking A, BofA, or both compared to the ΔB mutant in Fig 3. We suspect that this reduction in signal similarly reflects an association between pro-σK-CFP and B(E44Q) in the peripheral mother-cell membranes despite our inability to directly detect the membrane signal.
Finally, we investigated whether expression of A and BofA in vegetatively-growing cells producing wild-type B-YFP is sufficient to inhibit pro-σK-CFP processing and membrane association of the fluorescent fusion. As anticipated, pro-σK-CFP remained full-length (Fig 4B and S2C Fig) and failed to associate with the membrane (Fig 4A). Collectively, these experiments support the model that A and BofA maintain B in a conformation that cannot interact with pro-σK and IVB-dependent signaling promotes a stable association between B and its substrate.
The B protease, like other members the S2P group 3 subfamily, contains a C-terminal cytoplasmic cystathionine-β-synthase (CBS) domain [7]. These domains commonly bind ligands with adenosyl groups like ATP, AMP, S-adenosylmethionine, and c-di-AMP and regulate enzymatic or transport functions [31, 32]. In vitro, the CBS domain from the B protease has been shown to bind pro-σK [13]. We therefore wondered whether the stable association between B and pro-σK in response to IVB signaling was mediated by the CBS domain. Consistent with this idea, a deletion of this domain was reported to abolish pro-σK processing in an E. coli expression system [13, 38].
To investigate the role of the CBS domain in IVB-dependent signaling in B. subtilis, we generated a deletion (BΔ85) that lacks the CBS domain and interdomain linker and a smaller 66 amino acid deletion (BΔ66) that lacks the CBS domain but retains the linker [13, 38]. In addition, we generated a 10 amino acid C-terminal deletion (BΔ10) that was previously shown to be produced in E. coli at levels similar to wild-type but was impaired in pro-σK processing in this heterologous system [13]. All three deletions contained the E44Q catalytic mutation and were fused to YFP to monitor localization and help stabilize the proteins. All three fusion proteins localized properly (Fig 5A) and, based on fluorescence intensities, were produced at levels similar to the full-length protein (S7B Fig). Importantly, in sporulating cells harboring either the BΔ66 or BΔ10 mutant, pro-σK-CFP accumulated in the outer forespore membranes (Fig 5A and S7A Fig), consistent with a recent study in which B truncations similar to those used here were found to interact with pro-σK when co-expressed in E. coli [38]. In support of the idea that the CBS domain helps stabilize the interaction between B and pro-σK [13, 38], the pro-σK-CFP fluorescent signal around the forespore was reduced compared to wild-type, however and importantly the percentage of cells with forespore-localized pro-σK-CFP was similar (S4D Fig). Moreover, we found that this localization pattern was dependent on the IVB signaling protein (S8 Fig). By contrast and similar to previous work [13], the BΔ85 mutant phenocopied the ΔB null. In line with our pro-σK-CFP localization data, the BΔ10 and BΔ66 truncations with an intact protease domain supported efficient sporulation while the BΔ85 mutant sporulated at levels similar to the null (Fig 5B). Analysis of σK activity during a sporulation time course (S9A Fig) revealed that the BΔ66 and BΔ10 strains activate σK with kinetics similar to wild-type B. Furthermore, σK activity was dependent on IVB, indicating that these truncations are subject to inhibition by A and BofA and are activated by site-1 signaling. Finally, immunoblot analysis revealed that the BΔ66 mutant processed pro-σK to mature σK with slightly reduced efficiency compared to the matched wild-type control (S9B Fig). Altogether, these data indicate that the CBS domain is not critical for the IVB-dependent forespore localization of pro-σK, pro-σK processing, or efficient spore formation. Thus, these results suggest that pro-σK associates with the membrane protease domain of B and raised the possibility that IVB triggers a transition from a closed to open conformation of the caged protease allowing stable association between pro-σK and B's catalytic center (Fig 1A).
Prompted by these data, we sought to explore the biological relevance of the two conformations observed in the mjS2P structures. We used homology modeling to predict the structure of the B protease domain in both open and closed states (Figs 6A and 6B). Because the B protease domain shares relatively low sequence identity with mjS2P, we used evolutionary co-variation analysis to guide target-template alignment and validate the resulting models. Evolutionary co-variation analysis takes advantage of the fact that residues that interact with one another in a folded protein tend to co-evolve to maintain their interactions [39, 40]. Analysis of mjS2P by this method using 5,290 S2P subfamily 3 orthologs identified extensive co-variation in amino acids that interact in the crystal structures (S10 Fig). Importantly, the interactions between the two molecules of mjS2P that form the antiparallel pseudo-dimer in the asymmetric unit of the crystal were not observed by co-variation analysis (S10 Fig), consistent with the anti-parallel interface being a crystallographic artifact. Co-variation analysis of B identified 81 evolutionary coupled residues (90% probability threshold) (S10 Fig). The analysis confirmed our assignment of sequence register in each of the TM segments and indicates that our homology models are reasonable proxies for the B structures.
Examination of the models of the B protease showed that a hydrophobic membrane-reentrant β-hairpin is buried in the hydrophobic core of the protease in the closed state (Fig 6A). Phenylalanine 66 sits at the tip of this β-hairpin, where it makes extensive interactions with a cluster of hydrophobic residues in the closed conformation while being mostly exposed in the open-conformation model (Fig 6A and 6B). Specifically, F66 contacts V189, F188, W185, L136, and V128 in the closed state and only an interaction with F188 is preserved in the open state model. Interactions between F66 and these residues are predicted to occlude substrate and help stabilize the closed state of the lateral gate (Fig 6A). The analogous residue (I77) in the mjS2P protein is similarly buried in the core of the protein and is predicted to help stabilize the closed conformation. In the open conformation, F66 (and I77 in mjS2P) is displaced from the catalytic center and the first and sixth transmembrane segments have moved apart. To investigate whether the B protease domain adopts the closed conformation in vivo, we generated a B(F66A) mutant. Removal of the phenyl ring, which makes most of the hydrophobic contacts, is predicted to destabilize the closed state relative to the open, generating a constitutively active protease. In support of this idea, a strain harboring B(F66A)-YFP was capable of activating σK during sporulation in the absence of both the site-1 signaling protease IVB and an auxiliary signaling protease CtpB [22, 41] (Fig 6C and 6D and S11A Fig). The B(F66A) point mutant similarly supported pro-σK processing in the absence of IVB albeit with reduced efficiency compared to a wild-type IVB+ strain (S11B Fig). In the presence of IVB, the point mutant had a modest sporulation defect (S11C Fig), consistent with the pre-mature activation of σK observed in this background (Fig 6D). Importantly, B(F66A)-YFP specifically localized in the membranes surrounding the forespore (Fig 6E) in a manner that depended on A and BofA (S12 Fig), supporting the idea that the F66A substitution did not disrupt interactions between the protease and its negative regulators. Collectively, these data are consistent with the model that the B protease adopts open and closed conformations in vivo.
Altogether, the results presented here are consistent with the substrate-gating model originally proposed by Shi and co-workers [30]. In the context of the pro-σK processing pathway, we envision that A and BofA inhibit the site-2 protease by stabilizing a closed, substrate-inaccessible conformation that prevents pro-σK access to the caged interior of the protease (Fig 1A). IVB-dependent (site-1) cleavage of A destabilizes this conformation leading to lateral displacement of the first and sixth transmembrane segments of the membrane-embedded protease allowing the pro-domain of σK access to the catalytic center of the enzyme. How cleavage of A triggers the open conformation is not known, but our data suggest that A need not be released from the complex as B(F66A)-YFP can process pro-σK while likely still bound to A and BofA and previous studies suggest that A remains tethered to the complex after IVB cleavage [22]. However, we note that pro-σK processing and σK activation are delayed in the B(F66A) mutant in the absence of IVB (Fig 6D, S11A and S11B Fig) and the localization of pro-σK-CFP to the forespore membranes in a B(E44Q, F66A)-YFP strain still required the IVB signal, suggesting that the B(F66A) mutant samples both open and closed states with A and BofA promoting the closed conformation. Future experiments will be directed toward understanding how IVB-dependent cleavage of A destabilizes the closed conformation triggering lateral gate opening.
Reconstitution studies using recombinant B protease revealed that pro-σK processing requires ATP [13]. In line with this finding, the CBS domain of the B protease was shown to bind ATP in vitro [13]. Furthermore, as described above, a deletion of the C-terminal CBS domain did not support cleavage of a modified pro-σK substrate in an E. coli expression system [38]. Since CBS domains have been proposed to function as sensors of cellular energy status [31, 32] it was hypothesized that the CBS domain on the B protease might couple σK activation to the energy status of the mother cell. Here, we report that the CBS domain contributes to the association between B and pro-σK and the efficiency pro-σK processing in B. subtilis but is not strictly required for either nor is it necessary for efficient spore formation. We cannot account for the discrepancies between our in vivo analysis and the in vitro reconstitution and E. coli expression data reported previously, however, other differences between B. subtilis sporulation and these heterologous systems have been reported [38, 42]. In vitro studies with the purified CBS domain of B [13] and with the CBS module from the S2P from Archaeoglobus fulgidus [43] suggest that Mg2+-bound ATP favors monomerization. While it is possible that fusing YFP to BΔ66 influenced its ability to dimerize, we obtained similar results using a BΔ66 fusion to monomeric YFP and a YFP variant that retains the ability to dimerize (S13 Fig). Nonetheless, it is possible that replacing the CBS domain with YFP could mimic the activated conformation of this domain. We note that a BΔ66 truncation that lacks the YFP fusion is non-functional, however we do not have antibodies to the membrane protease domain to assess the stability of this mutant. That being said, we favor a deterministic model in which the activation of σK is principally governed by the developing spore via the production and secretion of the IVB signal and that mother cell energy status could dictate the efficiency of pro-σK processing and consequently the rate at which σK-directed genes accumulate in the mother cell.
The pro-σK processing pathway is distinct from the other characterized RIP signaling pathways in that the substrate of the intramembrane cleaving protease, pro-σK, is not an integral membrane protein and is not subject to sequential site-1 and site-2 cleavages. Yet, this pathway takes advantage of a 2-step proteolytic cascade to transduce information across the lipid bilayer. We wonder whether other RIP signaling pathways that employ members of the group 3 S2P subfamily function similarly. M. jannaschii does not appear to have a homolog of the A protein. However, it does encode a protein with homology to BofA and a IVB protease family member despite not being an endospore-forming organism. It will be interesting to identify native substrates of mjS2P and other members of this broadly conserved subfamily. In the case of pro-σK pathway, this variation on RIP signaling is perfectly matched with the morphological constraints that exist during sporulation. B, A, and BofA are produced in the mother cell during the morphological process of engulfment [20, 21], while pro-σK expression is subject to more stringent control [44, 45] and only accumulates around the time when engulfment is complete [44, 46]. Due to a membrane fission event at this late stage [47], the mother-cell membranes that surround the forespore become topologically distinct from the peripheral membranes of the mother cell (Fig 1A) [48]. Integral membrane proteins produced in the mother cell at this stage are exclusively inserted in the peripheral membranes and are therefore unable to access the membranes surrounding the forespore [49]. Accordingly, to have access to the B protease, pro-σK must be a peripherally-associated, rather than an integral, membrane protein. Thus, regulation of intramembrane proteolysis is achieved by a site-1 signaling protease that instead cleaves a negative regulator of the S2P family member.
As in the case with RseP and other S2P family members, a complete picture of the sporulation RIP signaling pathway awaits structural determination of B(E44Q) bound to pro-σK and the inhibited B-A-BofA signaling complex. With the recent advances in membrane protein crystallography using lipidic cubic phase and cryo-electron microscopy, these structures are now within reach.
All B. subtilis strains were derived from the prototrophic strain PY79 [50]. Sporulation was induced by resuspension at 37°C according to the method of Sterlini-Mandelstam [51] or by exhaustion in supplemented DS medium [52]. Sporulation efficiency was determined in 24–30 hour cultures as the total number of heat-resistant (80°C for 20 min) colony forming units (CFUs) compared with wild-type heat-resistant CFUs. All sporulation assays reported were based on three or more biological replicates. Expression of B, A, BofA and pro-σK during vegetative growth was performed in LB. IPTG was added to a final concentration of 0.5 mM at an OD600 of 0.1 and samples were analyzed every 30 minutes post induction. Deletion mutants were generated by isothermal assembly [53] and direct transformation into B. subtilis. Tables of strains (S1 Table), plasmids (S2 Table), oligonucleotide primers (S3 Table) and descriptions of plasmid construction and isothermal assembly deletion mutants can be found online as supplementary material.
Strains bearing the PgerE-lacZ fusion were sporulated by resuspension. 1 mL samples were collected by centrifugation every hour during sporulation and stored at -20 ˚C. Samples were processed by the method of Miller [51, 54], using ortho-nitrophenyl-β-D-galactopyranoside (ONPG) as substrate. β-Galactosidase specific activity was defined as the change in A420 per minute per milliliter of culture per OD600 × 1000 and is reported in Miller units. All β-Galactosidase assays reported were based on three or more biological replicates. σK activity during sporulation was also assessed on DSM agar plates containing 100 μg/mL 5-bromo-4-chloro-3-indolyl β-D-galactopyranoside (X-gal).
Whole-cell lysates from sporulating cells (induced by resuspension) or vegetatively-growing cells were prepared as described previously [34, 55]. Equivalent loading was based on OD600 at the time of harvest. Proteins were separated by SDS-PAGE on 12.5% polyacrylamide gels, electroblotted onto Immobilon-P membranes (Millipore) and blocked in 5% nonfat milk in phosphate-buffered saline (PBS)-0.5% Tween-20. The blocked membranes were probed with anti-σK (1:10,000) [37], anti-σA (1:10,000) [56], anti-GFP (anti-10,000) [33], or anti-SpoIIP (1:10,000) [57] diluted into 3% BSA in 1X PBS-0.05% Tween-20. Primary antibodies were detected using horseradish peroxidase-conjugated goat, anti-rabbit IgG (1:20,000; BioRad) and the Western Lightning reagent kit as described by the manufacturer (PerkinElmer). At least two biological replicates were performed for each immunoblot. To resolve pro-σK-CFP (55 kDa) and mature σK-CFP (53 kDa), lysates were separated on a 35 cm 10% polyacrylamide gel.
Fluorescence microscopy was performed with an Olympus BX61 microscope as previously described [58]. Cells were mounted on a 2% agarose pad containing resuspension medium using a gene frame (BioRad). Fluorescent signals were visualized with a phase contrast objective UplanF1 100x and captured with a monochrome CoolSnapHQ digital camera (Photometrics) using MetaMorph software version 7.7 (Molecular devices). The membrane dye 1-(4-trimethylammoniumphenyl)-6-phenyl-1,3,5-hexatriene p-toluenesulfonate (TMA-DPH, Molecular Probes) was used at a final concentration of 50 μM and exposure times were typically 500 ms. The DNA dye 4’, 6-diamidino-2-phenylindole dihydrochloride (DAPI, Molecular Probes) was used at 2 μg/mL and exposure times were typically 200 ms. At least two biological replicates were performed for all microscopy experiment. Images were analyzed, adjusted and cropped using MetaMorph software.
Immunoprecipitations were performed as described previously [59]. Briefly, 50 mL cultures were harvested at hour 4 after the initiation of sporulation by resuspension. Cell pellets were washed twice with 1X SMM (0.5 M sucrose, 20 mM MgCl2, 20 mM maleic acid pH 6.5) at room temperature and then resuspended in 5 mL 1X SMM containing lysozyme (0.5 mg/mL). The suspension was gently shaken for 30 minutes at room temperature to generate protoplasts. Protoplasts were collected by centrifugation and flash-frozen in N2(l). Thawed protoplasts were disrupted by osmotic lysis with 3 mL hypotonic buffer (Buffer H) (20 mM Hepes pH 8, 200 mM NaCl, 1 mM dithiothreitol, with protease inhibitors: 1 mM phenylmethylsulfonyl fluoride, 0.5 μg/mL leupeptin, 0.7 μg/mL pepstatin). MgCl2 and CaCl2 were added to 1 mM and lysates were treated with DNAse I (10 μg/mL final) and RnaseA (20 μg/mL) for 1 hour on ice. The membrane fraction was separated by ultracentrifugation at 35 krpm for 1 h at 4°C. The supernatant was removed, and the membrane pellet dispersed in 200 μL of Buffer G (Buffer H with 10% glycerol). Crude membrane preparations were aliquoted and flash-frozen in N2(l). 100 μL of crude membranes were diluted 5-fold with Buffer S (Buffer H with 20% glycerol and 100 μg/mL lysozyme), and membrane proteins were solubilized by the addition of the nonionic detergent digitonin (Sigma) to a final concentration of 0.5%. The mixture was rotated for 1 h at 4°C. Soluble and insoluble fractions were separated by centrifugation at 35 krpm for 1 h at 4°C. The soluble fraction (the load) was mixed with 20 μL of affinity-purified anti-GFP antibodies [22] covalently coupled to Protein A sepharose and rotated for 4 h at 4°C. The resin was pelleted at 5 krpm and washed five times with 1 mL of Buffer S + 0.5% digitonin. Immunoprecipitated proteins were eluted with 50 μL of sodium dodecyl sulfate (SDS) sample buffer (0.25 M Tris, pH 6.8, 6% SDS, 10 mM EDTA, 20% glycerol) and heated for 15 min at 50°C. The resin was pelleted and the supernatant (the IP) was transferred to a fresh tube and 2-mercaptoethanol was added to a final concentration of 10%. The immunoprecipitates were analyzed by immunoblot and SDS-PAGE followed by silver staining [59]. Individual bands were excised from silver-stained gels and trypsinized. Extracted peptides were then separated on a nanoscale C18 reverse-phase HPLC capillary column, and were subjected to electrospray ionization followed by MS using an LCQ DECA ion-trap mass spectrometer. Among the 5 peptides identified by MS, 3 (YLEILMAK, FGLDLKK, EIAKELGISR) were from σK. No σK peptides were identified from the same region of the gel in the controls.
To generate homology models of the B protease, the sequence was first aligned to that of mjS2P using the HHPred server to generate an initial alignment for homology modeling [60]. This was further adjusted manually to correct a register error in the last transmembrane segment revealed by evolutionary co-variation analysis. The resulting modified alignment was used to construct a homology model in MODELLER [61] using the structures of mjS2P as a template (PDB ID: 3B4R). Both conformations were modeled, using 3B4R chain A as the template for the open conformation, and chain B as the template for the closed conformation.
Multiple sequence alignments (MSA) were generated for both the B protease (SP4FB_BACSU, residues 1–210) and mjS2P (Y392_METJA solved in PDB 3B4R, residues 1–224). The MSAs were built using jackhmmer [62], an iterative hidden Markov model-based sequence search tool, with 5 iterations querying against the April 2017 Uniref100 database [63]. For SP4FB_BACSU, the alignment contained 5,390 sequences (2,005 effective sequences after downweighting sequences with more than 80% identity), with 92.4% of the input residues covered with less than 30% gaps. Y392_METJA was aligned with 5,290 sequences (1927 effective sequences) with 92% coverage at 30% gap allowance. Evolutionary couplings were then determined as previously described [39, 64, 65]. The full EVFold package and documentation can be found at https://github.com/debbiemarkslab/EVcouplings
|
10.1371/journal.pcbi.1000421 | Maximum-Likelihood Model Averaging To Profile Clustering of Site Types across Discrete Linear Sequences | A major analytical challenge in computational biology is the detection and description of clusters of specified site types, such as polymorphic or substituted sites within DNA or protein sequences. Progress has been stymied by a lack of suitable methods to detect clusters and to estimate the extent of clustering in discrete linear sequences, particularly when there is no a priori specification of cluster size or cluster count. Here we derive and demonstrate a maximum likelihood method of hierarchical clustering. Our method incorporates a tripartite divide-and-conquer strategy that models sequence heterogeneity, delineates clusters, and yields a profile of the level of clustering associated with each site. The clustering model may be evaluated via model selection using the Akaike Information Criterion, the corrected Akaike Information Criterion, and the Bayesian Information Criterion. Furthermore, model averaging using weighted model likelihoods may be applied to incorporate model uncertainty into the profile of heterogeneity across sites. We evaluated our method by examining its performance on a number of simulated datasets as well as on empirical polymorphism data from diverse natural alleles of the Drosophila alcohol dehydrogenase gene. Our method yielded greater power for the detection of clustered sites across a breadth of parameter ranges, and achieved better accuracy and precision of estimation of clusters, than did the existing empirical cumulative distribution function statistics.
| The invention and application of high-throughput technologies for DNA sequencing have resulted in an increasing abundance of biological sequence data. DNA or protein sequence data are naturally arranged as discrete linear sequences, and one of the fundamental challenges of analysis of sequence data is the description of how those sequences are arranged. Individual sites may be very sequentially heterogeneous or highly clustered into more homogeneous regions. However, progress in addressing this challenge has been hampered by a lack of suitable methods to accurately identify clustering of similar sites when there is no a priori specification of anticipated cluster size or count. Here, we present an algorithm that addresses this challenge, demonstrate its effectiveness with simulated data, and apply it to an example of genetic polymorphism data. Our algorithm requires no a priori knowledge and exhibits greater power than any other unsupervised algorithms. Furthermore, we apply model averaging methodology to overcome the natural and extensive uncertainty in cluster borders, facilitating estimation of a realistic profile of sequence heterogeneity and clustering. These profiles are of broad utility for computational analyses or visualizations of heterogeneity in discrete linear sequences, an enterprise of rapidly increasing importance given the diminishing costs of nucleic acid sequencing.
| Analysis of discrete linear sequences has played an increasingly important role in biology. In particular, the detection of heterogeneous regions among sequences can aid in understanding the heterogeneous processes that act upon those regions [1],[2]. Therefore, determining whether specified types or categories of sites, such as polymorphic [3] or substituted sites [4] within DNA or protein sequences, are concentrated in specific regions within DNA or protein sequences has become a key component of these analyses [5]–[8]. For instance, detecting regions that feature heterogeneity in substitutions may provide valuable information on the structure and function of DNAs or proteins [9]–[13].
Several parametric and nonparametric methods have been proposed and historically applied to sequence data. Parametric methods include applications of a Fisher's exact test to tallies of site types between regions, or of a likelihood ratio test to identify heterogeneous regions [14],[15]. Alternatively, several heuristic methods may be applied for this clustering [16]. For example, UPGMA (Unweighted Pair Grouping Method with Arithmetic-mean) or NN (Nearest Neighbor), are hierarchical methods that at each step combine the nearest 2 clusters into one new cluster. Iteration of this step is continued until the number of clusters is one. One of NN's variants, K-NN (K-Nearest Neighbor), differs in its termination condition, stopping the iteration until the K clusters are identified, where K needs to be defined in advance. Another heuristic approach, K-means, uses a partitioning algorithm to break data into K clusters, and also requires the number of clusters K as a prior knowledge. When regions of a sequence that are expected to have heterogeneous frequencies of a site type may be specified in advance or the number of clusters to be identified is known a priori, these methods have high power to detect clustering [17]. However, they require a priori assignment of partitions. When no a priori expectation of cluster size or cluster number may be specified, extant studies have usually relied on “sliding window” methods [18]–[23]. For example, Pesole et al. (1992) labeled invariable site as ‘1’ and variable site as ‘0’, and applied a sliding window to identify whether ‘1’s are significantly clustered [24]. Pesole et al. calculated a heuristic score based on the presence or absence of site types within a window that processes serially across the sequence of interest.
Advantages of sliding window methods include their intuitive conceptual basis and their striking output: an autocorrelated plot of the score that may be superimposed upon the sequence, providing a visual appraisal of the level of clustering at every site. However, sliding window methods have two related major disadvantages [25]. First, they generally offer only crude non-parametric means for statistical significance testing. The autocorrelation of serial scores severely complicates attempts to develop more insightful parametric approaches to sliding window significance testing, making parameter estimation with confidence intervals either challenging or impossible. Second, the need to specify a window size presents a user with a procedural ambiguity. Without a unified statistical framework, there is no strong justification for selection of one window size over another. In such a situation, it may even be tempting to invert the procedure of statistical inference and select a window size that produces an autocorrelated score plot consistent with a particular scientific hypothesis, as opposed to the valid procedure of selecting a window size by an objective statistical optimality criterion.
Because of these disadvantages of the sliding window methods, several nonparametric statistical methods that do not assume prior knowledge have been suggested or implemented to detect clustering in discrete linear sequences. These methods include runs tests [26]–[28] and empirical cumulative distribution function (ECDF) statistics [29],[30]. Runs tests use the “longest unbroken run” between sites of interest as a test statistic for clustering, where a run is defined as consecutive length between events [26]. This test statistic provides very weak power, because it uses very little of the relevant information about the phenomenon of interest, ignoring all runs other than the longest. Statistics based on the longest two runs, longest three runs, or even on a summary of the full distribution of run lengths have been discussed, but remain weak tests. For instance, the variance in distance between site types of interest may be calculated and used as a test statistic for the detection of clusters of sites, where a high variance is indicative of clustering [29]. This test statistic incorporates information about the length of all the runs, but does not capture all of the relevant information: it discards all information about the relative position of runs of different lengths. A sequence with all of its shorter runs in one region would be more clustered than one with short runs distributed evenly.
Currently, the most powerful nonparametric method is the ECDF. It features the cumulative difference between the observed and expected proportion of variant sites to identify regions that differ from other regions in number of substitutions. Under a null model that assumes no heterogeneous region(s) within sequences, this difference remains close to zero. Its significant departure from zero is an indicator for rejecting the null model [29],[30]. Although ECDF has been used to detect heterogeneity in several studies [31]–[35], its power can be affected by the location of the heterogeneous region [30]. Moreover, a parametric method may perform even better across a wide range of datasets.
Most extant methods that have been proposed to detect heterogeneous clusters among sequences suffer from poor power to detect clustering when it is present. The problem is made especially challenging by a tradeoff wherein increasing power to detect clustering also increases overparameterization or false positive rates. Methods that have high power are prone to identify clustering even in random sequences, because even in short sequences, there are so many potential patterns of clustering to evaluate. In this paper, we propose a hierarchical clustering method, model averaged clustering by maximum likelihood (MACML), requiring no priori knowledge of cluster size or cluster count, that provides greater statistical power in detecting heterogeneous regions. MACML adopts a divide-and-conquer approach to hierarchically detect heterogeneous regions and repeat similar analysis for each identified region, unlike most hierarchical methods that do not revisit clusters once they are constructed [17],[36],[37]. To address issues of overparameterization, MACML employs model selection and model averaging techniques that lead to intuitively appealing profiles of sequence heterogeneity and that facilitate description of clustered sites in discrete linear sequences. We describe MACML in detail and provide comparative results in the form of an in-depth evaluation of simulated datasets and an empirical sequence data set on polymorphisms in the Drosophila alcohol dehydrogenase gene.
To apply MACML to locate regional clusters with different specified site types requires a general input sequence X with N sites, denoted as(1)For example, to examine heterogeneity of substitution, an aligned set of homologous sequences is converted into X, in which each site is scored entries xi of 0 representing identity, and 1 representing a variant or variable site [30]. Similarly, a sequence to be analyzed for detection of GC heterogeneity can be converted by setting G/C = 1 and A/T = 0. Notations used to describe our algorithm are summarized in Table 1.
To test the performance of MACML and compare it to the most powerful extant method, ECDF, we simulated sequences for analysis for which the rates of variant sites were known a priori. For each simulated sequence, we randomly generated the start and end positions of the cluster, positions of variant sites within the cluster region, and positions of variant sites within the non-cluster region (see Figure 1). To avoid stochastic errors, we repeated simulations M = 10000 times for each parameter combination. Thus, each performance measure was determined from M replicates.
We retrieved the Drosophila alcohol dehydrogenase (Adh) gene within five species of Drosophila melanogaster species subgroup (D. melanogaster, D. sechellia, D. simulans, D. yakuba and D. erecta) from FlyBase [47]. The aligned sequences of Drosophila Adh gene can be available at http://www.yale.edu/townsend/datasets.html.
The powers of MACML and ECDF were plotted against the percentage of variant sites within the cluster (q) under different numbers of variant sites (n) in Figure 3 and the corresponding accuracy and precision were plotted in Figure 4. Evaluating the methods based on their power to detect clusters within sequences with different q and n, MACML outperformed ECDF for nearly all the parameter combinations tested (Figure 3). When n was very small, both methods exhibited extremely low power for detecting hot spots (n = 10 in Figure 3A). At intermediate values of n, MACML and ECDF exhibited increasing power with q (Figure 3B and 2C). While ECDF approached the power of MACML when q was large, MACML remained more powerful across the full range of q (Figure 3B to 2D).
The power of MACML and ECDF to detect cold spots was also low when n was small (n = 10 in Figure 3E). When n increased to 50, the power of MACML and ECDF peaked at intermediate values of q (Figure 3F). At higher levels of n = 100 (Figure 3G) and n = 200 (Figure 3H), ECDF continued to peak at intermediate values of q, whereas the power of MACML continued to rise with q. Across the parameter ranges examined, MACML consistently exhibited greater power than ECDF.
The accuracy and precision of MACML and ECDF were estimated by the Kullback-Leibler (KL) divergence, which is a measure of the difference between the expected and estimated distributions of variant rates. In assessing the accuracy based on the KL divergence, therefore, there are three potential scenarios: a good match between the estimated and expected variant rates when a KL divergence is near zero, an underestimation of variant rates when KL divergence is positive, and an overestimation of variant rates when KL divergence is negative. The precision based on the KL divergence is also better when it is closer to zero. Unlike the accuracy, precision based on the KL divergence cannot be negative (Equation 12).
Evaluating the accuracy and precision based on the KL divergence, MACML performed better than ECDF for most of the cases examined (Figure 4). The accuracy and precision of MACML and ECDF for detecting hot spots were very good (near zero) when n was small (Figure 4A). When n became large, MACML exhibited good accuracy and precision, whereas the accuracy and precision of ECDF diverged positively from zero with increasing q (Figure 4B to 3D). This divergence was augmented when n was extremely large (Figure 4D).
When n is small (n = 10 in Figure 4E), both MACML and ECDF also exhibited good accuracy and precision for the detection of cold spots. At large values of n (Figure 4F to 3H), ECDF exhibited good accuracy and precision only when q was smaller (10%) or larger (90%). At intermediate values of q, the accuracy of ECDF diverged from the ideal negatively. The precision of ECDF diverged from the ideal as well. This divergence was augmented when n was extremely large (n = 200 in Figure 4H). In summary, MACML exhibited good accuracy and precision for nearly all tested cases.
The powers of MACML and ECDF were plotted against the ratio of variant rates within cluster to outside of cluster in Figure 5, and the corresponding accuracy and precision were plotted in Figure 6. The difference in power between MACML and ECDF was least remarkable for the detection of cold spots (Figure 5A). At values of the ratio of variant rates within cluster to outside of cluster ranging from 0.3 to 0.9, differences in power between both methods were relatively small, whereas at values of the ratio <0.3, MACML showed much greater power to detect cold spots than did ECDF (Figure 5A). The power of MACML to detect hot spots consistently increased with increasing ratio (Figure 5B). Although the power of ECDF increased with the ratio as well, its power was much lower than the power of MACML across the examined ranges of values of the ratio (Figure 5B).
MACML provided good accuracy and precision (near zero) for detecting cold spots, whereas the accuracy of ECDF diverged negatively and the precision of ECDF diverged from the ideal as well (Figure 6A). This divergence was more notable at values of the ratio <0.7 (Figure 6A). With regard to hot spots, the accuracy and precision of ECDF diverged positively across values of the ratio from 2 to 10 (Figure 6B). As the ratio was increased, this divergence became more remarkable. In contrast, MACML exhibited better accuracy and precision for most of the examined cases (Figure 6B).
According to their definitions, the ratio of variant rates within cluster to outside of cluster = 1∶1, q = 0%, or q = 100% represent sequences with entirely randomly located substitutions under the null model. Therefore, we compared three criteria adopted by MACML and examined their errors of overparameterizing the clustering model when no clustering was imposed during the sequence generation. MACML and ECDF demonstrated high overparameterization and false positive rates, respectively (Table 2). The overparameterization rate of MACML markedly exceeded the false positive rate of ECDF for n = 10, n = 100 and n = 200. Implementing the AIC and AICc did little to moderate overparameterization, whereas implementing BIC significantly moderated overparameterization. Implementing the BIC did not bring overparameterization down to the false positive rate of ECDF for n = 10, 100, and 200, but did limit the overparameterization rate to approximately the false positive rate of ECDF for sequences with n = 50.
The powers of MACML and ECDF were plotted against sequence length in Figure 7 and the corresponding accuracy and precision were plotted in Figure 8. When sequence length increased from 100 to 1000 sites, MACML and ECDF provided decreasing power to detect both hot spots (Figure 7A) and cold spots (Figure 7B). This decrease was more prominent for MACML than for ECDF. Nonetheless, MACML outperformed ECDF for most of these cases.
The accuracy and precision of MACML and ECDF varied little across all values of sequence length. With increasing sequence length, the accuracy of ECDF diverged from zero positively for hot spots and diverged slightly negatively for cold spots. The precision of ECDF diverged from the ideal positively for both hot spots and cold spots (Figure 8A and 7B). Overall, MACML exhibited better accuracy and precision than ECDF as sequence length increased from 100 to 1000 (Figure 8).
The powers of MACML and ECDF were plotted against the number of clusters in Figure 9. Under the parameters examined for multiple clusters (see Materials and Methods), MACML and ECDF performed similarly when the sequence had only one cluster to be detected. However, when the number of clusters ranged from 2 to 10, ECDF was unable to detect more than one cluster, whereas MACML had significant power to detect multiple clusters, especially for large values of n. In general, the power of MACML was limited for small values of n = 10 (Figure 9A) and n = 50 (Figure 9B), but much greater for large values of n = 100 (Figure 9C) and n = 200 (Figure 9D).
We applied MACML to detect heterogeneous clusters of polymorphisms within the Drosophila Adh gene and to profile potential for polymorphism for each site based on model selection and model averaging, respectively. Identified clusters as well as profiles of the potential for polymorphism were plotted against sequence coordinate (Figure 10). As expected, profiles of potential for polymorphism based on model selection (Figure 10A and 9C) are highly discrete, whereas smoother, continuous profiles are produced based on model averaging (Figure 10B and 9D). When using BIC, MACML detected two clusters along the Adh gene and both are cold spots residing between sites 98 and 189 and between sites 26 and 70 (Figure 10A and 9B). In addition to these two cold spots, when using AIC or AICc, MACML also identified two hot spots between sites 80 and 84 and between sites 212 and 218 (Figure 10C and 9D). In contrast, ECDF detected only one cold spot between sites 98 and 211 (data not shown), consistent with previous applications of the method [29],[30].
Detailed clustering results for the Adh gene are summarized in Table 3. For the AIC or AICc, the four detected clusters all deviate significantly from the null model (ΔAIC<0 and ΔAICc<0 in Table 3). When sample size is large, like sequence from sites 0 to 253, the ΔAICc asymptotically approaches ΔAIC, and thus their values are nearly same. However, for a smaller sample size, for example, when detecting sub-sequence from sites 71 to 97, ΔAICc is much larger than ΔAIC. By contrast, BIC incorporates a heavier penalty than AIC or AICc and ΔBIC>0 indicated no significant cluster among sub-sequences from sites 71 to 97 or from 190 to 253, whereas AIC and AICc identified two clusters along these two sub-sequences.
The power to detect heterogeneous clustered sites within sequences depended in moderately complex ways on the parameters we examined in this report. Consistent with expectations, our results show that the power of MACML to detect hot spots and cold spots increased with increasing percentage of variant sites within the cluster (Figure 3). Across simulations comparing different percentages of variant sites within the cluster, MACML exhibited both high accuracy and high precision: the estimated variant rates within and outside clusters were close to the expected ones across all parameter combinations (Figure 4). In contrast to MACML, ECDF performed more variably across different percentages of variant sites within the cluster. This inconsistency of performance agrees well with our theoretical analysis on ECDF (Text S1) as well as with results from a previous study [30]. The hot spots and cold spots estimated by ECDF tend to be narrower than the simulated hot spots and cold spots [30]. The misattributed region between the boundary of the estimated hot or cold spot and the corresponding boundary of the simulated hot or cold spot generally gives rise to much greater KL divergence than any other region of the sequence. Thus, the KL divergence of the full sequence tends to be dominated in direction and magnitude by the KL divergence of the region between these boundaries, a region that is usually present as a consequence of the bias in estimation of the width of hot and cold spots. Accordingly, positive divergence from perfect accuracy and precision for hot spots (Figure 4A to 3D) follows from underestimation of the variant rate of this region. Likewise, negative divergence from perfect accuracy and positive divergence from perfect precision for cold spot (Figure 4E to 3H) follows from overestimation of the variant rate of this region.
Across a range of ratios of variant rates within the cluster to outside of the cluster, MACML and ECDF exhibit similar trends in power, but different trends in accuracy and precision. With both methods, a significant difference between variant rates within the cluster and outside of the cluster leads to greater power, and nearly equal rates for all sites results in lower power (Figure 5). The KL divergence measure of the accuracy of ECDF is negative for cold spots and positive for hot spots, respectively (Figure 6). When the variant rate inside of the cluster approaches the variant rate outside of the cluster, estimated and actual variant rates are very close for any cluster model. Therefore, the accuracy and precision of ECDF approach those of MACML, consistent with simulation results (Figure 6). In contrast, as variant rates within the cluster diverge from rates outside the cluster, MACML produces incrementally better accuracy and precision across all parameter combinations (Figure 6).
Both MACML and ECDF exhibit decreasing power with increasing sequence length (Figure 7), presumably as a consequence of the decreasing proportion of variant sites relative to sequence length. Increasing sequence length with a fixed number of variant sites is equivalent to decreasing the number of variant sites with a fixed sequence length. Therefore, it is consistent that the power decreases with decreasing variant sites in Figure 3. This relationship between variant sites and power also agrees well with the results observed when varying the number of clusters (Figure 9), with the additional note that ECDF fails to detect more than one cluster. It is notable that simulations performed by Tang and Lewontin [30] were less general in scope than ours. That is, in Tang and Lewontin [30], the heterogeneous cluster was always centered and the two regions flanking the cluster were always equal in length. As noted by Tang and Lewontin, the power of ECDF is affected when the cluster moves off center [30]. In our simulations, the starting position and ending position of cluster are randomly generated, leading to a random location of the cluster and thus to an unequal length of the two flanking regions (see details in Materials and Methods). For these reasons, our simulations that incorporated random positions of clusters yielded different results in terms of success detecting multiple clusters than were yielded by the simulations of Tang and Lewontin [30].
False positive rates and overparameterization for clustering models were high, as expected as a consequence of the large number of potential cluster boundary sets that are possible. Powerful methods for this class of problem are expected to display high false positive rates, a tradeoff that is natural in statistical inference. Although ECDF presents lower false positive rates, MACML achieves more power than ECDF to reject the null hypothesis when it is not true (Figures 3, 4 and 6). Moreover, MACML achieves markedly greater accuracy and precision of variant rates as determined by the KL divergence (Figures 3, 5 and 7), demonstrating the marked superiority of MACML in selecting the best model of variant rates across a discrete linear sequence. Furthermore, MACML is more capable of detecting multiple clusters among sequences, as demonstrated by simulation (Figure 9) and by application to the empirical data (Figure 10).
Unlike ECDF, which is not integrated into a model selection framework, MACML adopts AIC, AICc and BIC for model selection. To clarify the differences observed implementing these diverse criteria, the different penalties for additional parameterization that they entail may be compared. Based on the clustering model, two parameters (cs and ce) are evaluated (from which p0 and pc can be calculated). Therefore, the number of parameters under the clustering model is two, whereas the number under the null model is zero. From equations 4–6, then,(11)(12)(13)where l is sample size, that is, (sub-)sequence length.
The values of lnLc–lnL0 may be plotted against sample size (Equations 11–13, Figure 11). AIC yields constant penalties for all values of sample size. For smaller sample size, AICc yields larger penalties than AIC or BIC. When sample size increases to large numbers, the penalty of AICc approaches AIC, and BIC produces much larger penalties than AICc.
For a given value of lnLc–lnL0, the three criteria are most likely to give different results with regard to rejection of the null model. The three lines plotted corresponding to the three different criteria in Figure 11 may be helpfully related to the results of our application of MACML to the Adh gene. MACML started by detecting a cluster from site 0 to 253. The sample size was 254, and the corresponding value of lnLc–lnL0 was 6.53 (Table 3). This cluster is represented by a point (254, 6.53), located above all three lines. This location signifies that the three criteria all reject the null model. After locating the first cluster, MACML proceeded to detect clusters along sub-sequences from 0 to 97, from 98 to 189, and from 190 to 253, until all possible sub-sequences had been tested. As a consequence, it identified several clusters. Two of them are located above the three lines, signifying that all three criteria reject the null model. The remaining two points are located below the BIC line and above the other lines, signifying that BIC does not reject the null model, but that the rest do (Figure 11). This graphical analysis clarifies results in which BIC identified only two cold spots, whereas the other criteria identified an additional two hot spots (Figure 11 and Table 3).
The Drosophila Adh is the most studied enzyme that catalyzes the oxidation of alcohols to aldehydes/ketones [48]. It has been extensive reported that several functionally important residues reside in the Adh gene: tyrosine-152, lysine-156 and serine-139 are conserved in homologous dehydrogenases and have important roles in catalysis [49]–[53]; glycine-130, glycine-133 and glycine-184 contribute substantially to the structure of the active form [50]; and aspartic acid-64 lies within a coenzyme-binding domain [51]. As shown in Figure 10 and Table 3, these residues were all clustered into the cold spots by MACML, indicating not only their functional conservation and relevance, but also the extent of the region of near-neighbor amino acids that are also conserved. Near-neighbors may be conserved due to their structural and biochemical effects on the known function of these residues. In addition, according to its gene structure, two introns in the Adh gene reside between the nucleotide sequences coding for residues 32 and 33 and between the nucleotide sequences coding for residues 167 and 168 [54],[55]. Therefore, the two cold spots identified by MACML extending from residues 26 to 70 and from residues 98 to 189 indicate conservation around the introns.
Heterogeneity of variant rates among specified site types is thought to commonly occur [56]–[59] and may derive from many sources, including functional constraint, gene structure, 3D protein structure, composition bias, mutation bias or recombination [1], [18], [34], [60]–[62]. As indicated by our results based on the simulated data and real data, MACML, equipped with model selection and model averaging, features smooth and continuous profiles of variant rates for each site, and is more accurate and more informative for the detection of multiple clusters among sequences. Therefore, MACML furnishes broad utility for any computational analyses of heterogeneous discrete linear sequences and provides valuable information to aid for a better understanding of the structure and function of DNAs or proteins.
In addition, MACML can be applied to a broad range of applications. For example, MACML would be appropriate for determining whether components of any multicomponent polymer have a clustered structure [33],[63]. It can also be used to detect compositional heterogeneity within sequences [64]–[66] (e.g., heterogeneous GC content by setting G/C = 1 and A/T = 0). Moreover, MACML may provide a framework upon which future modeling of the substitution process may be overlain, assessing heterogeneity in selective pressure acting on different coding sequence regions [60], [67]–[70] and detecting fast-evolving regions in noncoding sequences [71],[72].
Here we have presented a method, MACML, to detect clustering of a site type in discrete linear sequences. MACML features maximum likelihood estimation, model selection criteria (AIC, AICc, and BIC) and model averaging to profile sequence heterogeneity. It employs a divide-and-conquer approach to hierarchically detect multiple clusters within sequences, without requiring a priori knowledge for cluster size or number. We compared MACML with the most powerful competing method, the ECDF, by exploring a full range of parameter space using computer simulations, and by performing an analysis of empirical data. Our comparative results show that across a wide range of parameter combinations, MACML outperforms ECDF not only by exhibiting greater power to detecting hot spots and cold spots. Thus, it represents a powerful exploratory tool for profiling clustering in discrete linear sequences. Although discoveries using MACML should be considered tentative, it yields greater resolution than any other method, providing a significant advance for the analysis of clustering of sites within discrete linear sequences.
|
10.1371/journal.pcbi.1004952 | Filopodial-Tension Model of Convergent-Extension of Tissues | In convergent-extension (CE), a planar-polarized epithelial tissue elongates (extends) in-plane in one direction while shortening (converging) in the perpendicular in-plane direction, with the cells both elongating and intercalating along the converging axis. CE occurs during the development of most multicellular organisms. Current CE models assume cell or tissue asymmetry, but neglect the preferential filopodial activity along the convergent axis observed in many tissues. We propose a cell-based CE model based on asymmetric filopodial tension forces between cells and investigate how cell-level filopodial interactions drive tissue-level CE. The final tissue geometry depends on the balance between external rounding forces and cell-intercalation traction. Filopodial-tension CE is robust to relatively high levels of planar cell polarity misalignment and to the presence of non-active cells. Addition of a simple mechanical feedback between cells fully rescues and even improves CE of tissues with high levels of polarity misalignments. Our model extends easily to three dimensions, with either one converging and two extending axes, or two converging and one extending axes, producing distinct tissue morphologies, as observed in vivo.
| The development of an embryo from a fertilized egg to an adult organism requires not only cell proliferation and differentiation, but also numerous types of tissue restructuring. The development of a relatively round initial embryo into one elongated along its rostral-caudal axis involves coordinated tissue elongation and cell reorganization in one or more groups of cells or tissues. Counterintuitively, in many organisms, cells in elongating tissues elongate and increase their protrusive activity in the direction perpendicular to the axis of elongation (convergent extension). Experimental and theoretical studies have not determined how this cell-level oriented protrusive activity leads to observed tissue-level changes in morphology. We propose a filopodial-tension model that shows how tension from oriented cell protrusions leads to observed patterns of tissue CE.
| Embryonic development requires numerous changes in tissue morphology. Convergent-extension (CE) is a basic tissue shape change [1–9], during which cells in an epithelial sheet rearrange to narrow (converge) the tissue along one planar axis while lengthening (extending) it along the perpendicular planar axis (Fig 1). Although CE has been observed in the development of many organisms [1–8], the specific cellular mechanisms that drive such movements are still subject of investigation [10].
Both asymmetric external forces on a tissue (passive CE) and asymmetric forces generated by the cells within a tissue (active CE) can lead to CE (Fig 1) [10]. Hypothesized mechanisms for CE include anisotropic cell edge/actin contraction [11,12], anisotropic cell adhesion and elongation [13,14], cell shape extension/retraction [11,15], combinations of a constraining boundary with undirected cell elongation [16] or with directed leading edge protrusion [17], and increased cell adhesion within tissue segments [18] (see Supplemental Material for a more detailed discussion of previous models). Existing models of CE, however, neglect the experimentally observed prevalence of filopodial extension parallel to the direction of tissue convergence [3,9,19–24], which could produce anisotropic traction forces between cells or between cells and the extracellular matrix [25–28].
The observed asymmetry of filopodial protrusion led us to propose a filopodial-tension mechanism for CE based on anisotropic filopodial pulling forces between cells. We explicitly model the number of cell-cell connections, their range, angular distribution, strength, and frequency of formation and breakage. We define an appropriate set of metrics to quantify both the effects of model parameters and planar-polarization defects (such as misalignments and the passive cells) on the dynamics of tissue-level CE. Since our filopodial-tension model extends naturally to three dimensional tissues, we discuss the two types of 3D CE and their corresponding tissue morphologies.
Experiments show that long filopodia continuously form and retract during CE in epithelial sheets and that these filopodia preferentially form in-plane along angles near the axis of tissue contraction. Each model cells therefore extends and retracts filopodia (which we represent using the model concept of a link) distributed within a range of angles around the directions perpendicular to the cell’s planar-polarity axis. To simulate the observed binding of filopodial tips to other cells and the roughly length-independent pulling forces which retracting filopodia generate, in our model, an extending link binds to the cell it contacts, then generates a constant (length independent) tension force between the cells it connects [20,25,29]. We then test whether this tension force is sufficient to explain observed local cell intercalation and global tissue CE.
In the filopodial-tension model (Fig 2, S1 Movie) cells form and eliminate links representing filopodia with a defined set of neighboring cells (terms in boldface identify model objects). Each cell carries a polarization vector (perpendicular to its planar-polarity axis) (Fig 2, red arrow) that defines its preferred direction of filopodial protrusion (Fig 2, blue horizontal line). We simplify the model by having the links connect the centers-of-mass of cells rather than connecting the actin cortex of one cell to the actin cortex of the contacted cell, as do real filopodia. Because filopodia typically form in a pair of growth cones roughly along the convergence axis and with a typical maximal length, we allow a cell to form links within a range of angles ±ϑmax around this axis on either side of the cell with a maximum length of approximately rmax. Specifically, a cell can form a link only with those cells whose centers-of-mass lie within a distance rmax from its center of mass and within an angle ±ϑmax of its polarization axis (Fig 2, blue horizontal line). A cell can have at most nmax links to other cells at any time (including links formed and received) and only one link is allowed between any pair of cells. The actual number of links a cell forms may be less than nmax. Each link between a pair of cells exerts a tension force of magnitude λforce along the line connecting the cells’ centers-of-mass. To model the finite lifetimes of filopodia, we define a relaxation time, tinterval, after which we remove the links of all cells and create new ones. In a simulation in which the links form and then persist indefinitely, the cells only move a few microns (lattice sites) from their original locations and the tissue does not converge or extend.
We implement the filopodial tension model using the Cellular Potts model (CPM, also known as the Glazier-Graner Hogeweg model, GGH), where each cell is represented as a collection of lattice sites with the same cell index. An effective-energy cost function, H, specifies the cell’s properties (see supplemental material). The tension force along a link between a pair of cells is independent of its length and acts along the vector between their centers-of-mass. In the GGH/CPM formalism, the tension has the form:
H=H0+∑σ,σ′λforce(σ,σ′)lσ,σ′
(1)
where the sum is over all pairs of linked cells, λforce is the strength of the pulling force between cells σ and σ’, lσ,σ’ is the current distance between the cells, and the term H0 aggregates all the other GGH/CPM cost function terms. The GGH/CPM simulations evolve stochastically from random lattice-site updates subjected to the effective-energy cost function, H. The time unit is the Monte Carlo Step (MCS), defined as the rate of lattice-site updates (see supplemental material for more details on the GGH/CPM formalism).
The filopodial-tension model has five intensive parameters (λforce, tinterval, rmax, nmax, ϑmax) and one extensive parameter (N, the number of cells), making a complete sensitivity analysis computationally costly. We therefore fixed all parameters to reference values that are within the ranges observed in vivo and produced biological plausible convergent-extension (Table 1), then studied the effects of varying each intensive parameter one-at-a-time. The biological parameters proposed by the model can be directly measured experimentally, but since the concept of a filopodial-based CE is new and applies more readily to CE of deep tissues, which are not as easily visualized as epithelial sheets, appropriate experimentally-derived values are harder to find. The most studied cases are chicken limb-bud mesenchymal intercalation [30] (tinterval = 2.2 hours; rmax = 3 cell diameters; nmax = 11; ϑmax = 45°), Xenopus gastrulation and notochord formation [31–33] (tinterval = 2.0–2.7 min; rmax = 1.5 cell diameters; nmax = 8–9; ϑmax = 60°), and Xenopus Keller explants [23,34] (tinterval = 0.5–1.0 hour; rmax = 1.5 cell diameters; nmax = 8–9; ϑmax = 30°).
All simulations start with a mass of identical cells uniformly distributed inside a rough circle. Each cell has the same planar-polarization vector (V). To quantify the degree of tissue deformation we calculate the inverse aspect ratio between the length of the minor (L-) and major (L+) axes of the tissue (Fig 3B, green line). Initially the aspect ratio is close to 1 and decreases in time to a final value κ (Fig 3B, dashed red line) that depends on the filopodial tension parameters (λforce, tinterval, rmax, nmax, ϑmax), the number of cells in the tissue (N) and the surface tension of the tissue γ (defined below).
The final inverse aspect ratio quantifies the maximum elongation of the tissue, but does not convey how fast the tissue elongates. To quantify the elongation rate, we define the elongation time (τ) the time an initially isotropic tissue takes for its major axis (L+) to double the length of its minor axis (L-), which is equivalent to the time when the inverse aspect ratio (L-/L+) first decreases to 0.5 (Fig 3B, dashed blue lines). We consider CE to fail if L-/L+ never reaches 0.5.
Since both the filopodial-tension model and the GGH/CPM are stochastic, we average the value of the elongation time (τ) over 10 simulation replicas. Because the tissue inverse elongation ratio converges to the same value independent of the simulation seed or initial conditions (S1 Fig), unless specified otherwise, we calculate the final inverse aspect ratio κ for a single simulation replica, with the standard deviation indicating the fluctuations in κ around its final value for that replica.
Successful CE depends on the ability of intercalating cells to generate forces stronger than the internal and external forces that oppose tissue deformation. Here, the opposing forces come from the superficial tension (γ) between the cells and the external medium, defined as [35]:
γ=Jc,M−Jc,c2,
(2)
where Jc,c is the contact energy between cells and Jc,M is the contact energy between cells and medium (see supplemental material).
When the filopodial tension is weak compared to the surface tension (λforce < 2γ), cells do not intercalate and CE fails. For larger filopodial tensions, the elongation time (τ) decreases as a power of λforce (τ ∝ λforce-1.25±0.03) (Fig 4A, red line). The final inverse aspect ratio (κ) decreases monotonically with increasing λforce (Fig 4B). Increasing γ shifts the κ vs. λforce curve to the right and decreasing γ shifts the κ vs. λforce curve to the left (Fig 4B, inset). Normalizing the filopodial tension by the surface tension (λforce/γ) collapses the κ vs. λforce curves (Fig 4B), showing the linear relationship between λforce and γ. The surface tension (γ), however, has little effect on the elongation time (τ), which depends on λforce, but is relatively insensitive to surface tension (Fig 4A). The κ vs. λforce/γ curve is sigmoidal on a log-log scale (Fig 4B), because the shape of the tissue changes little for weak filopodial tensions and because the total number of cells limits κ for strong filopodial tensions (see Fig S3A). At the inflection point of κ vs. λforce/γ, the tensions of the links (λforce) balances the external surface-tension forces that oppose tissue elongation (λforce/γ ~ 6). Near this inflection point κ varies as an approximate power law of λforce (κ ∝ λforce-1.51±0.08).
Next we studied how the remaining filopodial tension parameters affect CE, specifically, the mean lifetime of the filopodia, modeled as the time interval between link formation and breakage (tinterval); the maximum length of the filopodia, modeled as the maximum distance of interaction between the cells’ centers-of-mass (rmax); the maximum number of filopodial interactions per cell (nmax); and the maximum angle between the filopodial direction and the cells’ convergence axis (ϑmax).
Fig 5A shows that, for the reference parameter values (Table 1) the lifetime of filopodia, tinterval, has no effect on τ or κ for tinterval ≲ 200 MCS. For the reference parameter values, 200 MCS corresponds to the typical time the cells require to rearrange their positions in response to a given set of neighbors interactions. Increasing filopodial lifetimes above 200 MCS slows cell intercalation (increasing the elongation time) and increases the tissue’s final inverse aspect ratio (corresponding to less deformation).
The maximum range (rmax) of filopodia interaction has different effects on the final inverse aspect ratio (κ) and elongation time (τ). For rmax < 2 cell diameters, κ decreases as a power law in rmax (κ ∝ rmax-3.5±0.2), then saturates for rmax ≥ 2 cell diameters, while the elongation time (τ) decreases monotonically with increasing rmax (Fig 5B). The same effect is seen with respect to the maximum number of links (nmax): κ decreases as a power law in nmax for nmax < 4 (κ ∝ nmax-1.5±0.03) and saturates for nmax > 4 (this saturation makes sense since the cell typically has 4 neighbors within the range of its filopodia for rmax = 2 and ϑmax = 45°) while τ decreases monotonically with increasing nmax (Fig 5C). Thus rmax and nmax have affect the rate of cell intercalation more than the final inverse aspect ratio, while the tissue’s surface tension affects only the final inverse aspect ratio and not the rate of cell intercalation (Fig 4).
Both κ and τ are concave with respect to the maximum angle of filopodial protrusion (ϑmax), since for small ϑmax the number of cell center-of-mass within the cones defined by ϑmax and rmax is very small, while for ϑmax = 90° the forces on the cell are symmetric since it extends filopodia uniformly in all directions. In both limits CE fails (Fig 5D). Since the net intercalation force is the difference between the tension forces parallel and perpendicular to the convergence axis (roughly ∫0θmax(cos(θ)-sin(θ))dθ ), we might expect the force to be greatest (and thus κ and τ to be smallest) when ϑmax = 45° and for their values to increase symmetrically away from ϑmax = 45°. The curves, however, have different minima and are not symmetric: the smallest final inverse aspect ratio (κ) is around ϑmax = 40° (Fig 5D, red dots) and the smallest elongation time (τ) is around ϑmax = 30° (Fig 5D, blue squares).
This asymmetry is caused by the limited number of neighbors with which a cell can form a link. Both the maximum number of links per cell (nmax) and the number of cells within the link interaction range (rmax) can limit the actual number of links a cell forms. If the maximum number of links per cell is lower than the number of cell neighbors within a cone of range rmax (e.g. nmax = 3) and angle ϑ < ϑmax, increasing ϑmax leads to more links with cells at larger ϑ and thus reduces the net tension force applied along the direction of the convergence axis. In effect, large ϑmax causes the cell to waste its limited number of filopodia. For large nmax, links form to all cells within the cone of range rmax and small θ regardless of the value of ϑmax.Thus, for large nmax (e.g. nmax = 7), the κ and τ curves are roughly symmetrical around their minima at ϑmax ~ 45° (blue and red lines in Fig 5D).
The filopodial tension model assumes that cells can extend filopodia, contact and pull other cells that lie within a given distance, even if they do not touch each other before filopodial extension. An example would be the formation of adhesion junctions between cells which coupled to a contractile stress fiber in both cells. To model these cases, we defined a contact-mediated cell tension model, which is identical to the filopodial tension model except that the maximum link length rmax in the filopodial tension model is replaced with the condition that cells must be in touch before pulling on each other (Fig 6A).
The qualitative results for the contact-mediated cell tension model do not differ much from the filopodial tension model. The κ x λforce curve is sigmoidal on a log scale, τ decreases with a power law (κ ∝ λforce -1.18±0.06) and CE fails for λforce < 20 (Fig 6B). The dependence of κ on the number of filopodial interactions (nmax) is still a power law (κ ∝ nmax-1.5±0.03) and saturates when nmax = 4. The elongation time (τ), however, does not keep decreasing as it does for the filopodial tension model, but also saturates around nmax = 4 links (Fig 6C), as few cells have more than 4 neighbors with centers near the convergence plane. The (κ, τ) x ϑmax curves have minima at ϑmax = 40° and ϑmax = 35°, respectively, but are less skewed than in the filopodial tension model (compare Figs 6D and 5D). CE fails for ϑmax < 10° and ϑmax > 70°.
Convergent-extension requires cells to have consistent planar polarity throughout an extensive region of tissue. This correlated orientation might result from a long-range bias from a morphogen gradient, cellular or intercellular differences in protein expression [36], or from a boundary-relay mechanism [37,38]. In our previous simulations we assumed that all cells had perfectly aligned polarization vectors (Fig 2, red arrows), i.e., they all pointed in the same direction with the same magnitude, and they maintained their internal orientation throughout the simulation. To study the effect of polarization misalignment on CE we added a zero-mean Gaussian distributed displacement angle to the cells’ polarization vectors and varied the standard deviation of the distribution (σ) while keeping the mean direction (here, the vertical axis) constant. Since the final elongation ratio is sensitive to the distribution of polarization vectors, the values of κ were averaged over 5 simulations.
The filopodial tension model tolerates small polarization misalignments, with a tissue with a displacement angle of σ = 10° reaching the same final inverse aspect ratio as in the perfectly aligned case with little decrease in elongation rate (an 11% increase in τ). The tissue remained aligned with the mean direction of cell polarization (the vertical axis) for small misalignments (σ < 40°, Fig 7B), but bent at around σ = 50° (Fig 7C). For polarization misalignments with σ > 60°, CE fails and the tissue breaks its symmetry, acquiring more complex shapes such as the caltrop (see Fig 7D). Both metrics are exponential functions of the variance σ2 (Fig 7A).
So far we have assumed that the polarization vector of the cells remains constant throughout the entire process. In reality, however, cells are constantly communicating with their neighbors either through signaling or through mechanical interactions. During CE cells establish and maintain their polarity through the planar cell polarity (PCP) pathway, however there is growing evidence that mechanical feedback may also play a role in the maintenance of global tissue polarity during development [39–42].
The presence of a mechanical feedback mechanism may rescue CE in tissues with high polarization alignment defects. In order to investigate this we developed a simple model of mechanical feedback and applied it to the misalignment polarization cases that have been described in the last section (Fig 7). The feedback model assumes that the pulling forces on a cell due to filopodial interactions affect its polarization vector. We implement a simple phenomenological version of such an interaction by calculating the line of tension from the sum of all filopodial interactions of the cell with its neighbors (Fig 8A). From this line of tension we extract an orthogonal vector T which is averaged with the previous cell polarization vector V in the following way:
Vt+Δt=Vt*(1−w)+Tt*w,
(3)
where Vt+Δt is the polarization vector of a cell at time t+Δt, Vt* is the normalized polarization of the same cell at time t, Tt* is the normalized tension vector of the cell at time t, and w is a feedback weighting factor ranging from 0 (no feedback) to 1 (no memory) (Fig 8A). This iterative processes repeated at discrete time intervals set equal to filopodia lifetime (Δt = tinterval). For simplicity, we do not distinguish between the pulling forces generated by the cell from the pulling forces that their neighbors exert on it. The normalized tension vector of a cell is calculated from the vector that maximizes the sum of all projections of the normalized lines of force from all the cells’ neighbors:
Σi cos(∅i− ∅T)= 0,
(4)
where the sum is over all the cell’s neighbors that pull on it, ϕi is the angle of the line of force between the cell and the pulling neighbor i, and ϕT is the angle that defines the tension vector T* = (cos ϕT, sin ϕT).
For tissues with a high starting level of polarization misalignment (σ ≥ 40°), addition of this mechanical feedback mechanism usually leads to a lower final elongation ratio κ, as long as the feedback factor w is below 0.1. For these cases, tissue elongation times (τ) decrease with higher feedback levels (Fig 8C, green and red lines), while k usually decreases with lower feedback levels. For tissues with a low starting level of polarization misalignment (σ ≤ 20°), addition of this mechanism leads to lower final elongation ratios only for small levels of feedback (w ≤ 0.001) (Fig 8B, blue and black lines), while the time of tissue elongation (τ) remains relatively unchanged with respect to the case with no feedback (Fig 8C, blue and black lines). In all simulations where the addition of mechanical feedback rescues CE, the cells established a global polarization axis emergently. We chose to implement the feedback update iteratively, which leads to fast destabilization of the tissue for high levels of feedback, as is expected for a case with no memory. Although a continuous model would be relatively more robust, we expect the same destabilization effect when the weighting factor w approaches 1.
Next we varied the number of intercalating cells in the tissue to check if there is a minimum number of active cells needed to drive CE and how this change tissue dynamics. We defined two types of cells without filopodia: passive cells, which lack filopodia but can be pulled by the filopodia of other cells; and non-responsive, or refractory cells, which cannot be pulled by the filopodia of other cells. The former would correspond to cells whose surface adhesion molecules were compatible with those of the cells extending filopodia and the latter to cells with incompatible adhesion molecules. The parameters for cells which produced filopodia were the same as in Table 1. Since the final elongation ratio is sensitive to the distribution of active/non-active cells, the values of κ were averaged over 5 simulations.
For tissues with a mixture of active and passive cells, both κ and τ decrease monotonically with the percentage of active cells in the tissue (Fig 9, red dots). However, even a fraction of active cells can drive CE. For 40% or more active cells (S2 Movie), the tissue deforms almost as much as a tissue composed entirely of active cells (Fig 9B, red dots), though the elongation time increases with the percentage of passive cells up to twice that for a tissue of all active cells (Fig 9A, red dots). For higher fractions of passive cells the final inverse aspect ratio increases significantly with the fraction of passive cells (Fig 9B). E.g., for 90% passive and 10% active cells (Fig 9C and S3 Movie), the tissue’s final inverse aspect ratio never drops below 0.3 (Fig 9B) and the elongation time τ is more than ten times that for a tissue of all active cells (Fig 9A). In all simulations, the active cells migrate towards the midline of the elongating tissue, leaving the passive cells at the lateral margins (Fig 9D).
The presence of relatively high CE despite the presence of only 10% of active cells can be explained by the relative tissue area covered by the filopodia. Every active cell can pull on neighbors that lie up to a distance rmax from its center of mass and within an angle ϑmax on each side of its convergence plane (see Fig 2), thus covering an area of 2ϑmaxrmax2. For the reference parameters (ϑmax = π/2 and rmax = 2 cd, see Table 1) and a population of 10% of active cells (N/10, where N is the number of cells in the tissue) this amounts to ~1.2N(cd)2, which more than covers the whole area of the tissue (N(cd)2).
Refractory cells have a stronger effect on CE than passive cells. CE fails when the percentage of refractory cells is above 60% (Fig 9B, blue squares), while with passive cells it only fails for percentages higher than 90% (Fig 9B, red squares). For higher fractions of active cells, the two populations sort out, with the active cells extending normally and the refractory cells displaced to both sides of the elongating tissue (Fig 9F). Surface tension between the cells and the surrounding medium causes the refractory cells to form droplet-like clusters which bend the extending active-cell tissue into a wavy bar (Fig 9F and S4 Movie).
The 2D filopodial tension model is a reasonable description of cells within epithelial sheets, where cell movement is confined to a plane. However, in many situations cell intercalation occurs in 3D. That is the case in radial intercalation during epiboly of the developing Xenopus Laevis embryo, where cells in a multilayered epithelium intercalate and converge perpendicular to the plane of the sheet [43]. The filopodial tension model can be easily extended to three dimensions, but due to the extra degree of freedom, it breaks in two versions, depending on which axis is rotated:
In equatorial or extensional intercalation, obtained by rotating the 2D model around the polarization vector (the red arrow in Fig 2), the cells pull on all neighbors that lie in a convergence plane (Fig 10A). At the tissue level, equatorial intercalation results in the convergence of the tissue along the two directions perpendicular to the polarization vector and its extension along the polarization vector (Fig 10A’ and 10A”).
In bipolar or convergent intercalation, obtained by rotating the 2D model around the convergence line (the blue line in Fig 2), the cells pull on all neighbors that lie along a convergence axis (Fig 10B). At the tissue level, the bipolar intercalation results in the convergence of the tissue along the axis of convergence and its expansion in the other two directions (Fig 10B’ and 10B”).
Beginning with a spherical tissue with all the cells polarized in the same vertical direction, the 3D equatorial model produces a tissue resembling a prolate spheroid (cigar shaped, Fig 10A”, S5 Movie), while the bipolar model produces a tissue resembling an oblate spheroid (lentil shaped, Fig 10B”).
The bipolar model has more biological correspondence than the equatorial model: cells with unipolar or bipolar protrusive activity are much more common during development than cells with equatorial protrusive activity, and the resulting tissue shape from the 3D bipolar model corresponds to the thinning and expansion associated with radial intercalation.
For both versions of the 3D model, the dependence of the parameters κ and τ with λforce, rmax, nmax and tinterval are qualitatively the same as in the 2D model. The results only differ qualitatively with respect to ϑmax. For the same values of rmax and nmax, the 3D convergence model is slightly less skewed than the 2D version, with the best value for κ around ϑmax = 45° and the best value for τ around ϑmax = 35° (Fig 11B). The 3D extension model, however, presents a more drastic change in the (κ and τ) vs. ϑmax curve when compared to the 2D. While the 3D convergence model was slightly more symmetrical around ϑmax = 90°, the 3D extension model is very skewed towards small angles, with the best values for κ around ϑmax = 30° and the best value for τ around ϑmax = 15° (Fig 11A). In the extensional model CE fails for ϑmax < 3° and ϑmax > 60°, while in the bipolar model CE fails for ϑmax < 10° and ϑmax > 75°.
The reasons for the asymmetry is that the final shape of the tissue in the 3D extension model—a two cell diameter tube orthogonal to the convergence plane—is more sensitive to perturbations than the lentil shape tissue obtained by the 3D bipolar model. Two cells pulling each other along a convergence axis leads them to be aligned in a plane perpendicular to the direction of the pulling. This plane fully coincides with the 3D bipolar model extension plane (Fig 10B”), but only partially with the extension plane of the 3D extension model (Fig 10A”). In the simulations results shown in Fig 11A, the value of ϑmax = 30° represents the optimal maximum angle value where the pulling forces in the 3D extension model are still able to align the tissue without destabilizing it.
Here we developed a new 2D model of for active cell intercalation. The model drastically differs from previous existing models by its explicit use of pulling forces between cells rather than anisotropy on adhesion energies or surface tensions on the cell surface. A recent experimental work by Pfister et al. [28] supports our force-driven model. The use of forces makes the model easily adaptable to three dimensions, but the extra degree of freedom gives two ways in which this can be achieved: either by rotating the model around the polarization vector or around the convergence line (see Fig 10A and 10B).
The model was implemented in CompuCell3D using the Cellular Potts formalism. The core of the model, however, is independent of the mathematical formalism and can be easily implemented on other types of agent-based formalisms such as cell-center or vertex models, as long as they provide a volume exclusion mechanism. Although some quantitative results might differ, we expect the same qualitative results. The advantages of implementing and simulating our model using the Cellular Potts formalism include the ease to manipulate and study the effects of the surface tension on the tissue dynamics (Fig 4) and the addition of the model as an integrated part of the CompuCell3D simulation package [44] that allows for it to be immediately reusable by others.
Model validation can be done at two different levels quantitative and semi-quantitative: when both microscopic and macroscopic measurements are available for a specific tissue, using model parameters that agree with those measured for the cells in the tissue should result in tissue-level model output (here, the rate of convergence and final aspect ratio) agreeing quantitatively with that of the experimental tissue. This agreement should persist under different experimental conditions. This form of validation shows that the hypothesized mechanisms included in the model are sufficient to reproduce the experiment quantitatively. Note that a model can only show the sufficiency of modeled mechanisms, not their necessity, since a different set of mechanistic hypotheses might yield the same results. In our case, since we are not modelling a specific tissue, model validation can only be semi-quantitative. We show that changes in the properties of the modeled cells (such as average number of filopodia, length and angular distributions) and tissue properties (such ratio of active and non-active cells) predict relative changes in the rate of CE and final tissue aspect ratios in real tissues that agree with those in experiments when the corresponding parameter and conditions are similarly modified. In this case, agreement demonstrates the plausibility of the hypothesized mechanisms, but detailed quantitative validation requires additional experimental measurements. We hope that our semi-quantitative analysis of the filopodial tension model of CE will inspire the additional experimental measurements that a more detailed quantitative validation requires.
Our model predicts that external forces, such as surface tension and pressure, can only affect the final degree of tissue elongation (κ) (Fig 4), whereas the internal parameters that regulate cell-intercalation can affect both tissue dynamics (as measured by τ) and the final tissue shape (Figs 4 and 5). This can be easily tested experimentally by either changing the properties of the external environment (the surrounding cells/matrix) of the intercalating tissue or culturing it ex vivo.
Of the five cell intercalation parameters, the time interval between link formation/breakage (tinterval) had negligible effects as long as it is below the typical time that the cells take to rearrange positions and/or shapes in response to a given set of external forces. This might be different if a refractory time interval between pulls is added to the model. We expect that in the presence of such refractory time, an increased frequency in link formation/breakage would slow down the speed of intercalation and reduce the final elongation ratio.
The model also predicts that the maximum range of cell interaction (rmax) and the maximum number of links per cell (nmax) had no effect on the final tissue elongation after rmax = 2 and nmax = 3, but the time of elongation kept decreasing for higher values of rmax and nmax (Fig 5B and 5C). It was not possible to increase those parameters indefinitely to check if the speed of intercalation would also saturate because the simulated cells start to fragment past rmax ≥ 6 or nmax ≥ 7. The current implementation of the model, however, does not allow for more than one active link between cell pairs, which would likely decrease elongation time.
All cell intercalation models assume some type of increased cell activity along the convergence axis, which is often translated, as is the case here, into the assumption that the cells are bipolar (one exception being our 3D equatorial/extensional filopodial tension model). This however is not necessarily true and we expect the model to also work in cases where the simulated cells are either monopolar in opposite directions of the same convergence axis or randomly alternate being monopolar in each direction of the convergence axis.
Our model also suggests that CE can be successfully achieved even in the presence of relatively high degrees of polarization defects. We predict that tissues containing polarization misalignments of up to ±10° will be practically indistinguishable to the optimally aligned case. Even severe misalignments (about ±50°) would still lead to some CE, although to a much lesser degree of final elongation ratio and with longer tissue elongation times (Fig 8). Experimental disruptions of the PCP pathway that alter the global alignment of cells in a dose-dependent manner would provide a way to test some of these predictions. Addition of a simple mechanical feedback mechanism by which the cells readjust their polarization in response to the pulling forces from the neighbors does not have major effects on the speed of tissue elongation (Fig 8C), but can fully rescue and even improve on the final elongation ratio of tissues with low or even severe polarization misalignments (Fig 8B). We choose to implement a minimal phenomenological feedback mechanism to explore the general response and self-organization of the tissue, but the formulation of the model allows the replacement of this generic mechanism with more detailed feedback models that reflect a specific tissue.
In cases where some cells fail to polarize, the severity of the effects on CE will depend on the type of interaction between the polarized (intercalating) cells and the unpolarized (non-intercalating cells). If the polarized (or active) cells can still pull on the unpolarized cells, then CE still happens even in a situation where the vast majority of cells (95%) are not active (Fig 9), although at a great reduction in both speed and final elongation ratio. If, on the other hand, the unpolarized cells are non-responsive and cannot be pulled, then the reductions in speed and final elongation ratio are much more sensitive to the presence of unpolarized cells (Fig 9) and CE completely fails when the population of active cells falls below 25% (Fig 9C). Another prediction of the model is the separation between the intercalating cells and the non-responsive/refractory cells (Fig 9D). Such defects could be induced experimentally by randomly distributed knock-out of intercalating cells, e.g. using electroporation of tissues with a dominant negative or RNAi, and would provide a way to further test model predictions.
Finally, the model reduces to the more common implementations when the maximum range of interaction is replaced by the common contact area condition. In this case, instead of contracting (or increasing the tension of) the cell’s surfaces that are aligned with the polarization vector or the global direction of convergence, we pull the neighbors that are not aligned with it (Fig 6D). In both cases active CE is achieved by the same principle, promoted cell-cell activity along one axis and inhibition along the other.
|
10.1371/journal.pbio.1002089 | ATPase-Dependent Control of the Mms21 SUMO Ligase during DNA Repair | Modification of proteins by SUMO is essential for the maintenance of genome integrity. During DNA replication, the Mms21-branch of the SUMO pathway counteracts recombination intermediates at damaged replication forks, thus facilitating sister chromatid disjunction. The Mms21 SUMO ligase docks to the arm region of the Smc5 protein in the Smc5/6 complex; together, they cooperate during recombinational DNA repair. Yet how the activity of the SUMO ligase is controlled remains unknown. Here we show that the SUMO ligase and the chromosome disjunction functions of Mms21 depend on its docking to an intact and active Smc5/6 complex, indicating that the Smc5/6-Mms21 complex operates as a large SUMO ligase in vivo. In spite of the physical distance separating the E3 and the nucleotide-binding domains in Smc5/6, Mms21-dependent sumoylation requires binding of ATP to Smc5, a step that is part of the ligase mechanism that assists Ubc9 function. The communication is enabled by the presence of a conserved disruption in the coiled coil domain of Smc5, pointing to potential conformational changes for SUMO ligase activation. In accordance, scanning force microscopy of the Smc5-Mms21 heterodimer shows that the molecule is physically remodeled in an ATP-dependent manner. Our results demonstrate that the ATP-binding activity of the Smc5/6 complex is coordinated with its SUMO ligase, through the coiled coil domain of Smc5 and the physical remodeling of the molecule, to promote sumoylation and chromosome disjunction during DNA repair.
| The modification of target proteins by conjugation to SUMO—a small protein that acts as a regulatory tag—is essential for maintaining the integrity of genomes in most eukaryotic organisms. One critical step during the attachment of SUMO is the activation of the enzymes that catalyze this reaction—E1, E2, and the SUMO ligases. However, we currently do not fully understand how the different enzymes in the SUMO pathway are regulated. The SUMO ligase Mms21 is known to bind to Smc5/6, a large protein complex involved in the structural maintenance of chromosomes. Both Mms21 and Smc5/6 counteract the accumulation of recombination intermediates, which otherwise join replicated chromosomes, preventing their separation. Not surprisingly, the few known targets of the Mms21 ligase are mostly related to the repair of sister chromatids by recombination. Here, we show that the Mms21 SUMO ligase needs to bind to the Smc5/6 complex to promote chromosome separation. We used two Mms21-dependent SUMO conjugation targets—Smc5 and cohesin—to study the connection between the Mms21’s SUMO ligase activity and its binding partner, Smc5/6. Our results indicated that Mms21 activation is tightly coordinated with the intrinsic ATPase function of the Smc5/6 complex. However, the SUMO ligase and the ATPase lie in different domains of the Smc5/6-Mms21 complex that are normally distant from each other; we show that communication between these enzyme sites is enabled by the presence of conserved joints, which we suggest allow the necessary conformational changes required for SUMO ligase activation. This coordination of activities is extremely helpful for the cell, enabling it to integrate a structural role on chromatin during DNA repair with a signaling function, thereby promoting correct separation of the chromosomes.
| During mitotic division, cells dedicate a large part of their efforts to accurately maintain and transmit genetic material to their offspring. The Structural Maintenance of Chromosomes (SMC) complexes play key structural roles in chromosome organization and dynamics and are crucial to maintain the integrity of the genome [1]. SMC proteins are rod-shaped molecules with a long coiled coil that separates a hinge or dimerization domain at one end and a nucleotide binding domain (NBD) at the other. Eukaryotes encode three different SMC complexes, known as cohesin, condensin, and Smc5/6. Heterotypic interactions between hinge domains lead to the formation of V-shaped molecules, which then bind to a variable number of non-SMC proteins [2]. The coiled coil domain of SMC proteins displays a remarkable flexibility, most probably due to the presence of conserved disruptions, which allow SMC complexes to adopt a wide variety of conformations [3–6]. Dimerization through the hinge and persistent connection of the NBD heads by a kleisin subunit generate large ring-like structures able to bind chromatin [7,8].
Smc6 was originally isolated in Schizosaccharomyces pombe as rad18, a gene involved in DNA repair [9]. The complex has been subsequently shown to be required during DNA double-stranded break repair and in response to perturbed replication forks, by either preventing the accumulation of recombination intermediates or promoting their removal, thus allowing chromosome segregation [10–19]. All subunits of the Smc5/6 complex are essential for viability in budding yeast. The complex is composed of the Smc5-Smc6 heterodimer, plus 6 Non-Smc Elements (named NSE1 to NSE6), which collectively regulate its function [20]. Nse4 binds to the ATPase head region and has been proposed to be its kleisin subunit [21,22]. Nse4 also interacts with the Nse1 and Nse3 subunits, which together function as an heterodimeric ubiquitin ligase [23]. The Nse5 and Nse6 subunits are the least conserved proteins in the complex and, in budding yeast, bind to the hinge domains of the SMCs [21]. Finally, the Nse2 subunit, also known as Mms21, docks to the middle of the coiled coil region in the Smc5 molecule. The N-terminal part of the protein contains the Smc5-interacting domain and is essential for cell viability; in contrast, the C-terminus codes for a SUMO ligase SPL-RING domain and only becomes critical under conditions of genotoxic stress [24–27]. The SUMO protein can be covalently conjugated to lysine residues trough an enzymatic cascade [28], requiring activation by an E1 enzyme, transfer of SUMO to the E2 (Ubc9), and final E2-dependent direct conjugation to the target protein, or in collaboration with E3 SUMO ligase enzymes (Siz1, Siz2, and Mms21 in budding yeast).
The small number of SUMO ligases, relative to the large population of E2 and E3s in the ubiquitin system, raises the question of how sumoylation is regulated. Actually, sumoylation of Mms21-dependent targets is up-regulated by DNA damage through an unknown mechanism [25,29]. Sumoylation is essential for DNA damage repair, and mutations in E1, E2, or SUMO all lead to genotoxic sensitivity [30]. In contrast, Mms21 is the only E3 in yeast that renders cells sensitive to DNA damage when mutated [31], highlighting its central role in DNA repair. Different lines of evidence suggest that Mms21 promotes DNA repair from its location on the Smc5/6 complex. First, inactivation of the Mms21-branch of the SUMO pathway leads to a general decrease in genome integrity, and this phenotype is shared with mutants in other subunits of the Smc5/6 complex [20,32]. Second, differently to the related Siz/PIAS E3 ligases, Mms21 lacks a DNA-binding domain, which suggests that it must dock to other proteins to reach its chromatin-associated targets. And third, the mms21-M5 allele, which is partially affected in its binding to Smc5, is also sensitive to various DNA-damaging agents [24]. Although these observations suggest that Mms21 needs to bind Smc5 to promote DNA repair, it is currently unknown if the Smc5/6 complex controls the activity of its associated SUMO ligase.
To investigate the relation between Mms21-dependent sumoylation, the association of the ligase with the Smc5/6 complex, and its role in maintaining the integrity of the genome, we have analyzed mutants in the Smc5/6 complex that block Mms21-dependent sumoylation. Here we report that Mms21 needs to bind an active Smc5/6 complex to reach its sumoylation targets and to promote sister chromatid disjunction. We also provide evidence demonstrating that Mms21-dependent sumoylation is controlled distally by the ATPase activity in the Smc5/6 complex, which is part of the E3 ligase mechanism that promotes sumoylation. Furthermore, we show that specific articulations in the Smc5 coiled coil structure allow communication between the ATPase heads and Mms21 in order to enable the activation of the SUMO ligase. Our findings suggest that the ATP-dependent chromosome structural role of the Smc5/6 complex and its SUMO-ligase activity are coordinated to ensure proper chromosome segregation.
The Mms21 SUMO ligase may promote sumoylation and chromosome disjunction from its location on the Smc5/6 complex, or independently from the complex (Fig. 1A). To test the relation between the docking state of Mms21 and its DNA repair and sumoylation functions, we generated mutants that disrupt the Smc5-Mms21 interaction. Since mutation of the SUMO ligase itself could potentially prevent its recruitment to other targets, we decided to disrupt the Smc5-Mms21 interaction by mutating the coiled coil domain of SMC5, without affecting the MMS21 gene.
We designed three different sets of smc5 mutants, namely S1, S2, and S3. Each of them carries mutations in residues lying on the surface of Smc5 that contacts Mms21 (Fig. 1B and S1 Fig.) [24]: smc5-S1 has four mutations, I780R, I784R, F787A, and N791A; smc5-S2 has two, M769A and K773A; and smc5-S3 has three, Q752A, L755A, and L762A. The three smc5 mutant alleles were fused to the 9myc epitope and expressed from centromeric vectors. Co-immunoprecipitation analysis confirmed that the smc5-S1 protein cannot interact with a 6HA-tagged wild-type Mms21 protein, while the smc5-S3 mutation reduces the Smc5-Mms21 interaction to 50%, and the smc5-S2 mutation does not seem to have any effect (Fig. 1C). In contrast, none of the SMC5 mutants shows altered protein interactions with the Nse4 subunit of the Smc5/6 complex (Fig. 1C). To study their functionality, all mutants were ectopically expressed in GALp-SMC5 cells, which allow conditional depletion of the endogenous SMC5 gene. As shown in Fig. 1D, the smc5-S2 and smc5-S3 alleles can sustain growth of GALp-SMC5 cells in glucose, while the smc5-S1 mutant is lethal, further supporting the notion that the Smc5-Mms21 interaction is essential for cell viability [24].
SUMO-ligase impaired mms21 mutant cells display chromosome segregation and disjunction defects after exposure to DNA damage [10], which are particularly severe in the ribosomal DNA (rDNA) array locus (S2 Fig.). To test if this is due to diminished sumoylation from the Smc5/6 complex, we arrested GALp-SMC5 cells in G1 after switching off SMC5 expression, and then treated them with 0.01% of MMS for 30 min before release into a synchronous cell cycle (Fig. 1E,F). All cultures entered the first and second cell cycles after the G1 arrest with similar kinetics, as evidenced by the appearance of budded and re-budded cells respectively. While cells ectopically expressing wild-type SMC5 do not display any obvious mitotic defect, cells bearing an empty plasmid or expressing the smc5-S1 allele show gross failures in chromosome segregation (Fig. 1E); this is evident as a slight increase in anaphase cells, the virtual absence of cells that have completed chromosome segregation (two nuclei), and the accumulation of cells with unequal separation of DNA masses between mother and daughter cells (nuclear missegregation). Additionally, FACS analysis shows the appearance of cells with less than 1N DNA content in cultures expressing no SMC5 or the smc5-S1 allele, indicative of chromosome segregation failures when Mms21 is not recruited to the Smc5/6 complex (Fig. 1F). The more severe effect of the smc5-S1 mutation in chromosome segregation, relative to the SUMO-ligase defective mms21Δc mutant (S2B Fig.), is most probably due to the fact that the Smc5-Mms21 interaction is essential, while the SUMO ligase domain in Mms21 is not [24].
The experiments described above strongly support the idea that critical Mms21-dependent DNA repair targets are sumoylated from its location on the Smc5/6 complex, in contrast to a model where Mms21 sumoylates the repair targets independently from the rest of the complex (Fig. 1A). To directly test sumoylation of Mms21 targets, SUMO-conjugates were purified from cells that carry an N-terminal 6xHis-Flag (HF) epitope on the SUMO protein (Smt3). The endogenous wild-type Smc5 protein was depleted by shifting GALp-SMC5 cells to glucose-containing media. The smc5-S1 mutant protein displayed almost undetectable levels of sumoylation (Fig. 1G); on the other hand, sumoylation of Smc5 was slightly diminished, but not overly affected, by the smc5-S2 or smc5-S3 mutations, which do not substantially decrease its binding to Mms21. The sumoylation of the cohesin complex is also partially dependent on Mms21 [33,34], and this modification is required for recombination-dependent repair of chromosomes [15,34–36]. As shown in Fig. 1H, expression of the binding deficient smc5-S1 allele severely impairs Smc1 sumoylation. In summary, these results confirm that Mms21 binding to the Smc5/6 complex is required for modification of known Mms21 substrates.
The previous observations suggest that Mms21 binds the Smc5/6 complex to reach its substrates and promote DNA repair. Yet, and more appealingly, the structural and the SUMO-mediated signaling functions present in Smc5/6 might be coordinated to enhance DNA repair. The Smc5/6 complex can be dissected into different sub-entities (Fig. 2A) [24], including the Smc5-Smc6, Nse1-Nse3, and Nse5-Nse6 heterodimers, plus the Mms21 SUMO ligase and the Nse4 kleisin subunit. We therefore decided to test the participation of these sub-complexes on the E3 SUMO-ligase activity. Since Smc5 is a target of Mms21 [25] and the binding site for Mms21 in the complex [24], we used its sumoylation levels as an in vivo reporter for the activity of Mms21.
First, we tested cells that express their endogenous 3HA-tagged SMC6 gene from the GAL promoter. We observed that turning off the GAL promoter leads to a drastic reduction in Smc5 sumoylation (Fig. 2B). Interestingly, Mms21 SUMO-ligase activity cannot be restored to wild-type levels by the hypomorphic smc6-1 allele (Fig. 2B). Since the Mms21-Smc5 interaction is maintained, and Mms21 recruitment to chromatin is not overly affected in smc6-1 cells (Fig. 2C and D), these results indicate that the Mms21 SUMO ligase must be inactive when SMC6 function is impaired. Next, we studied the contribution of the Nse1-Nse3 and Nse5-Nse6 heterodimers using the thermosensitive and MMS-sensitive nse3-2 and nse5-2 hypomorphs (Fig. 2E). Co-immunoprecipitation analysis shows that the Smc5-Nse3 interaction is weaker in nse3-2 cells at 25°C and becomes severely impaired upon shift to the restrictive temperature (Fig. 2F). The nse5-2 mutant protein also interacts weakly with Smc5, even at the permissive temperature (Fig. 2F). However, neither the nse3-2 nor the nse5-2 mutations affect the Smc5-Mms21 interaction (Fig. 2G), and the nse5-2 mutation does not impair the binding of Mms21 to chromatin (Fig. 2H). Notably, both nse3-2 and nse5-2 mutant cells show reduced levels of Smc5 sumoylation, even at the permissive temperature for growth (Fig. 2I). These results indicate that proper recruitment of the Nse3 and Nse5 protein to Smc5/6 is required for the Mms21-dependent sumoylation of Smc5, in accordance with a previous report [37]. Finally, we confirmed that auxin-induced destruction of specific Smc5/6 subunits, including the Nse4 kleisin, also leads to a rapid loss of Smc5 sumoylation (S3 Fig.). Overall, our observations indicate that inactivation of different sub-entities in the Smc5/6 complex reduces Mms21-dependent sumoylation. Therefore, an active and intact Smc5/6 complex is required for the activity of its associated SUMO ligase.
Since the essential function of SMC complexes in chromosome maintenance requires the ATPase activity of its SMC subunits, we introduced mutations in the Walker A or B ATPase motifs of Smc5 (K75I or D1014A, respectively) to compromise its binding to ATP. The ATPase-mutant alleles were expressed in GALp-SMC5 cells and, as previously described, we observed that they render yeast cells non-viable (Fig. 3A) [38]. Co-immunoprecipitation experiments show that Mms21 binds with similar efficiency to either wild-type or ATPase mutant Smc5 proteins, indicating that there is no ATP-dependent modulation of the Smc5-Mms21 interaction (Fig. 3B). However, sumoylation of the ATPase-defective Smc5 proteins was almost undetectable (Fig. 3C), proving that they are not modified by their accompanying E3. As shown in Fig. 3D, the sumoylation deficiency affects other proteins in the complex, such as the Nse4 kleisin subunit, when cells only express the smc5(K75I) allele. Moreover, cells that only express the ATPase-defective version are also impaired in sumoylation of the Smc1 subunit of the cohesin complex (Fig. 3E), to a similar extent as mms21 mutants lacking the C-terminal E3 ligase domain (mms21ΔC), indicating that the ATPase mutant complex is poised in an inactive state for sumoylation. Therefore, sumoylation of Smc5/6-Mms21 targets depends on the ATP-binding ability of Smc5. Although diminished sumoylation in the ATPase mutants could stem from defective recruitment of the complex to chromatin, this possibility seems unlikely, since in vitro experiments show that Smc5 binds efficiently to DNA in the absence of ATP [38], and we did not observe major alterations in the chromatin binding of the ATPase mutant proteins or the SUMO ligase using a chromatin fractionation assay (Fig. 3F), not even under conditions of competition with the endogenous wild-type Smc5 protein (S4 Fig.). These results further argue that the recruitment of Smc5/6 to chromatin does not depend on the sumoylation state of the complex.
Mms21 could facilitate sumoylation-dependent DNA repair by bringing Ubc9 and its substrates into close proximity. In accordance, we observed that substitution of the SPL-RING sequence in MMS21 for that of the E2 conjugating enzyme (mms21ΔC-UBC9) partially suppresses most of the mms21ΔC temperature and DNA damage sensitivities (Fig. 4A,B). As expected for constitutive Ubc9 recruitment to Smc5/6, substitution of the SPL-RING domain for Ubc9 not only restores, but actually up-regulates Smc5 sumoylation levels (Fig. 4C).
We speculated that the ATPase-inactive Smc5 mutant protein may fail to recruit Ubc9, thereby precluding sumoylation. If this hypothesis was correct, artificial recruitment of Ubc9 to Smc5/6 would eliminate the sumoylation differences between the wild-type and ATPase mutant Smc5 proteins. To explore this possibility, we integrated a second copy of UBC9 fused to the C-terminus of the endogenous wild-type MMS21 gene, using a 3xHA epitope as a linker (Fig. 4A). This fusion is functional and displays higher levels of Smc5 sumoylation, as expected from constitutive recruitment of both the E2 and E3 enzymes (Fig. 4B,C). Besides, the E3-E2 fusion co-immunoprecipitates similar amounts of both the ATPase active and inactive Smc5 proteins (Fig. 4D), indicating that the Smc5-Mms21 interaction is not affected in the ATPase mutant. Strikingly, and in spite of the constitutive binding of Ubc9 to the Smc5 protein, its full sumoylation still required an ATP-dependent step (Fig. 4E). To test the sumoylation of the non-SMC protein Nse4 in E3-E2 cells, we integrated the Mms21-Ubc9 fusion in GALp-SMC5 cells, and transformed them with centromeric vectors that express 9myc-tagged versions of the Nse4 and/or Smc5 proteins. As shown in Fig. 4F, we could not detect Nse4 sumoylation when no SMC5 is expressed (second lane in Input and Pull Down); while ectopic expression of wild-type Smc5-9myc led to detectable levels of Nse4-9myc modification, the sumoylation of Nse4 could not be restored by expression of the ATPase-defective Smc5(K75I) allele. Altogether, these results suggest that the ATPase function of Smc5 is part of the ligase mechanism that enables sumoylation.
To further analyze the participation of the ATPase heads on sumoylation, we immunoprecipitated active and inactive Smc5/6 complexes from yeast cells and tested their ability to sumoylate Smc5 in vitro (Fig. 5A). Addition of human E1, E2, and SUMO to the immunoprecipitates led to the appearance of sumoylated species and sumoylation of Smc5-9myc in an ATP-dependent manner (Fig. 5B). Interestingly, the ATPase mutant Smc5 protein can also be sumoylated in vitro, arguing that the Smc5(K75I) allele is not intrinsically unsumoylatable. However, we noticed that the rate of Smc5 sumoylation was significantly lower in the mutant than in the wild-type protein (Fig. 5B and C). A more severe effect was observed when the yeast E1, E2, and SUMO proteins were used in the assay (S5 Fig.). To focus on an even smaller number of components, we expressed and purified from Escherichia coli the C-terminal domain of the Nse4 kleisin (which interacts with the ATPase head of Smc5), and the Smc5-Mms21 heterodimer. We chose to co-express the heterodimer because monomeric Smc5 is poorly expressed in bacteria, while the Smc5-Mms21 pair accumulates at higher yields (S6 Fig.). Addition of ATP to the sumoylation reaction promoted the mono-sumoylation of the Nse4 fragment (Fig. 5D). Yet again, we detected a significant reduction in the rate of Nse4 sumoylation when the reaction proceeded in the presence of the mutant Smc5(K75I) protein, relative to wild-type Smc5 (Fig. 5E). Therefore the SUMO ligase is less active when the ATPase head cannot bind ATP. Overall, these results support the notion that, in vivo, the Smc5/6-Mms21 complex operates as a giant E3 SUMO ligase, which is sensitive to the ATPase activity of the SMC heads.
The SUMO ligase binds in the middle of the coiled coil domain of Smc5, at 16–24 nm from the ATPase heads [24], and yeast two-hybrid experiments indicate that Mms21 does not seem to contact the NBDs of the Smc5 protein [21]. Therefore, it is possible that the ATP-dependent communication between the NBDs and the SUMO ligase is triggered through conformational changes in the Smc5-Mms21 molecule. To test this hypothesis, the Smc5-Mms21 heterodimer was expressed in E. coli, purified, and imaged by scanning force microscopy (SFM) (Fig. 6 and S7 Fig.). Visual inspection of the purified particles shows the presence of approximately 50 nm rod-shaped structures, as expected for an SMC protein. In addition, we also observed many globular particles, suggesting that the coiled coil domain may fold about or wrap around the ATPase heads of Smc5, as occurs in other SMC proteins [39,40]. Particles were automatically recognized and classified according to volume. The volumetric distribution of wild type and ATPase K75I mutants shows that most particles have the expected dimensions for individual Smc5-Mms21 heterodimers. Prior incubation with ATP produces a shift towards smaller particles, which could reflect partial loss of the Smc5-Mms21 interaction, an effect that is observed in both the wild type and K75I mutant heterodimers (Fig. 6A and B). Interestingly, ATP increased the frequency particles with greater height, indicative of a more condensed shape and a conformational change (Fig. 6C). Monomeric Smc5 molecules are expected to have a low rate of ATP hydrolysis, suggesting that the observed conformational change takes place in response to ATP binding. As expected, mutation of the nucleotide binding domain in Smc5(K75I)-Mms21 heterodimers substantially reduced the degree of ATP-dependent compaction (Fig. 6D). The smaller conformational change observed in the K75I molecule might be due to partial binding of ATP on the Smc5(K75I) head. Therefore, binding to ATP induces a conformational change in the Smc5-Mms21 molecule.
Because it was not possible to directly visualize the coiled coil structure of Smc5 in most particles, we could not determine whether it participates in the ATP-dependent conformational change. On the other hand, if the coiled coil of Smc5 collaborates in the ATPase-dependent activation of its SUMO ligase, it should be possible to identify specific features in this domain that are critical for Mms21 activity.
Proline residues are rarely observed in coiled coils because they do not favor α-helical structures. Secondary structure prediction shows that, in most species, the probability of coiled coil drops at three different positions in the first coiled coil (CC1) of Smc5 (Fig. 7A). The first two disruptions in CC1 frequently involve a proline residue (P271 and P305). Mutation of these residues to glutamic acid, an abundant amino acid in the coiled coils of Smc5, had little effect on Smc5 sumoylation (Fig. 7B). The third disruption in CC1 contains a well-conserved proline residue (P393 in budding yeast; Fig. 7A). Different amino acids, besides P393, contribute to this disruption. We noticed that the combination of H391D, P393E, and E394L mutations, which locally change the coiled coil sequence from HLPE to DLEL (smc5-DLEL), restored the heptad periodicity and substantially increased the coiled coil probability (S8 Fig.). The smc5-DLEL mutation does not affect the interaction between Smc5 and Mms21 (Fig. 7C); in accordance, smc5-DLEL cells are viable, indicating that the essential function of Smc5 is not affected (Fig. 7D). However, smc5-DLEL cells are sensitive to MMS and exhibit nuclear segregation defects after a pulse of MMS in G1, suggesting improper disjunction of sister chromatids (Fig. 7D and E). Moreover, smc5-DLEL cells are compromised in Smc5 and cohesin sumoylation (Fig. 7B and F), indicating that mutation of the coiled coil disruption down-regulates the activity of the SUMO ligase.
To prove that the MMS-sensitivity of smc5-DLEL cells is directly related to impairment in SUMO-ligase activation and not to defective Smc5 function, we forced sumoylation by introducing the E3-E2 allele at the MMS21 locus. We observed that constitutive recruitment of the E2 restores Smc5-DLEL sumoylation to levels comparable to wild-type cells (Fig. 8A). The fact that the double smc5-DLEL E3-E2 mutant displays lower sumoylation than the single E3-E2 is consistent with the idea that the third coiled coil disruption in Smc5 is also, as shown previously for the ATPase activity, part of the SUMO ligase activation mechanism. In accordance with restoration of Mms21-dependent sumoylation, constitutive E2 recruitment rescued the MMS sensitivity of smc5-DLEL cells (Fig. 8B), proving that their DNA damage sensitivity is not due to the structural alteration of the Smc5 protein, but to impaired activation of the SUMO ligase.
There is currently very little information about the regulation of SUMO enzymes. The Mms21 SUMO ligase is essential for the maintenance of genome stability [25–27,41], and it has been hypothesized that Mms21 docks to Smc5 to reach its few known substrates, which are mostly chromatin-associated [42]. Here we show that the SUMO ligase is physically and mechanistically coupled to the activity of Smc5/6. Remarkably, mutations that block Smc5 sumoylation also impinge on cohesin modification, which indicates that this mechanism is shared by other sumoylation targets outside the Smc5/6 complex.
In principle, cells could regulate Mms21-dependent sumoylation by triggering its recruitment on and off chromatin; for example, by binding to the Smc5/6 complex. However, our results indicate that the mere proximity of Mms21 is not sufficient for sumoylation of some of its targets. The clearest example is the Smc5 protein, which binds strongly to and is sumoylated by Mms21 [24,25]: ATPase-defective Smc5 proteins are not sumoylated, despite normal recruitment of both Smc5 and Mms21 to chromatin and proper binding of Smc5 to the SUMO ligase. Therefore, Smc5/6-Mms21 dependent sumoylation is only possible from an active Smc5 protein. Our findings do not exclude the possibility that Mms21 might target other proteins in an Smc5/6- (and hence ATPase-) independent manner. However, such targets do not seem to participate in chromosome disjunction, as a wild-type Mms21 protein is incapable of promoting chromosome segregation when not recruited to Smc5 (Fig. 1E and F). The smc5-S1 mutant developed in this study should thus become an indispensable tool to test putative Smc5/6-independent roles of the Mms21 SUMO ligase.
Our results also shed new light on how the Smc5/6-Mms21 branch of the SUMO pathway is controlled at the molecular level. Differently to the ubiquitin pathway, Ubc9 can directly transfer SUMO to its targets [28]. However, sumoylation of most proteins in budding yeast requires the presence of a SUMO ligase [43]. In the case of substrates that directly interact with the E2, the presence of an E3 might promote the correct orientation of the Ubc9-SUMO thioester for catalysis [44] or establish additional contacts with the substrate [45]. Sumoylation of Smc5/6 subunits also requires binding of the E3 to its substrate (Fig. 1), suggesting that Mms21 promotes sumoylation and chromosome repair by stimulating the formation of an E2-SUMO-E3-target complex. In accordance, artificial recruitment of the E2 to Smc5/6 in cells lacking the E3 SUMO ligase domain of Mms21 suppressed their DNA-damage sensitivity (Fig. 4). Importantly, our results also point to an unforeseen regulation of Mms21 in the Smc5/6 complex, since the ATPase activity of Smc5 is required for activation of the Mms21 ligase, even after recruitment of the E2 (Figs. 3 and 4). The ATP-dependency is not restricted to the Smc5 protein, as Nse4 and cohesin (Fig. 3D and E) are also hypo-sumoylated when Smc5 cannot bind ATP. Therefore, the Smc5/6-Mms21 complex must have lower activity when Smc5 is not bound to ATP. In accordance, our in vitro assays show a 2- to 3-fold lower rate of sumoylation in ATPase-defective Smc5 mutant proteins. Such defect might seriously compromise the ability of Smc5 ATPase mutants to sustain proper sumoylation levels in vivo, as SUMO enzymes are probably less accessible than in vitro, and SUMO peptidases are actively removing SUMO from targets. Different situations could account for the lower sumoylation in ATPase mutants, ranging from the incapacity to properly orient the Ubc9-SUMO thioester, the inhibition of SUMO discharge from Ubc9, or the presence of a molecular obstruction in those Smc5/6 molecules that are not charged with ATP. A detailed view of the Mms21-Ubc9-SUMO interaction structure, as well as the participation of Smc5/6 structural elements in this process, will be required to solve this issue.
Apart from the ATPase heads, the coiled coil domain to which Mms21 binds is also required for SUMO ligase activation and chromosome segregation. To our knowledge, this is the first report for a specific function of a coiled coil domain in an SMC protein. Our results are consistent with the smc5-DLEL mutation severing the communication between the ATPase heads and Mms21. Although we cannot formally discard that the DLEL mutation indirectly decreases the ATPase activity, this situation seems unlikely because (i) the mutation is located far away from the NBDs of the Smc5 protein, and close to the Mms21 docking site; (ii) differently than ATPase mutants, smc5-DLEL cells are viable; and (iii) the MMS-sensitivity of the DLEL mutant can be bypassed by constitutive tethering of Ubc9 to the complex, an observation that directly links this coiled coil disruption to activation of the SUMO-ligase. Moreover, the inability of the smc5-DLEL E3-E2 double mutant to reach the hyper-sumoylation state of the E3-E2 single mutant shows that this knuckle is part of the E3. It is worth noting that the proline residue present in this disruption is conserved in evolution, indicative of a vital function and a possible similar regulation of the Smc5/6-Mms21 ligase in humans.
The coiled coil domains in SMC proteins display a wide variety of conformations, most probably due to the presence of kinks at specific disruptions in this domain [39,40]. The coiled coil flexibility in SMC proteins might be important to accommodate chromatin fibers inside the ring structure, and it might also help to bring different domains of the molecule in close contact [46,47]. We have observed an analogous conformational heterogeneity for the Smc5-Mms21 heterodimer and further conformational changes upon ATP binding. ATPases are known to couple ATP binding and hydrolysis to mechanical work, and coiled coil domains can transmit this information to other regions of the molecule [48,49]. In the case of dynein, the communication is enabled by a change in the registry of the two short alpha-helical chains in the coiled coils [48], although this seems dubious for Mms21, given the distance separating the NBDs and the Mms21 binding site. Other possibilities could be the rotation of the ATPase heads along the coiled coils axis, as is the case for Rad50 [50] or folding of the molecule at specific articulated disruptions, as has been hypothesized for the cohesin complex [51]. The participation of the P393 disruption in Mms21-dependent sumoylation invokes a model where the ATP-dependent reshaping of the molecule allows activation of the SUMO ligase (Fig. 9).
If bending of the Smc5/6 molecule leads to Mms21 activation, this might only happen in the context of a competent Smc5-Smc6-Nse4 ring. Indeed, our results suggest that the Smc5/6 functions as a giant SUMO E3 enzyme, and the different sub-entities present in the Smc5/6 molecule are required for Mms21-dependent sumoylation. We hypothesize that the non-SMC elements, as well as Smc6, could directly participate in Mms21 activation, as we have shown here for Smc5 (Fig. 3 and 7). A second possibility is that the NSE mutants analyzed might diminish the ATPase activity of the complex. For example, the kleisin subunit in the cohesin complex is known to regulate the ATPase activity of the SMC heads [52], and loss of Nse4 could inactivate Smc5/6 in an analogous manner. A third option is that NSE malfunction might indirectly diminish the ATPase activity by precluding Smc5/6 recruitment to damaged DNA [37]. It has been proposed that the Nse5-Nse6 sub-complex might regulate chromatin association of Smc5/6 through opening of the Smc5-Smc6 hinge interface [21], analogously to what occurs during chromatin loading of cohesin [51]. Still, it is worth noting that Mms21 remains bound to chromatin in nse5-2 mutants, and docked onto Smc5 (despite down-regulation of Smc5 sumoylation) in all the thermosensitive smc5/6 mutants tested; the most plausible explanation being that Mms21 is not active. Interestingly, Nse5 is known to directly interact with proteins of the SUMO pathway, including Ubc9 and SUMO [37,53]; therefore, it is tempting to speculate that this sub-complex might play specific roles in SUMO conjugation through recruitment of Ubc9.
Our study emphasizes the importance of the intimate relation between the Mms21 SUMO ligase and its binding site, the Smc5/6 complex. The Mms21 branch of the SUMO pathway and the Smc5/6 complex are required to prevent the accumulation and/or promote the removal of pathological recombinogenic structures [10–13,16]. These structures are lethal, since they prevent segregation of sister chromatids. The meiotic program, which requires the induction of double-stranded breaks, also requires the Smc5/6-Mms21 complex to properly channel recombination intermediates [17–19]. Here we have shown that mis-regulation of the SUMO ligase activity in the complex renders cells unable to disjoin and segregate chromosomes after DNA damage. The integration of the ATPase and the SUMO ligase in the Smc5/6-Mms21 complex should help to coordinate a structural activity on chromosomes with a signaling role via sumoylation, both of which would be directed to the efficient resolution and proper segregation of sister chromatids. The relatively small number of known Mms21 targets, most of which participate in processing of double-stranded breaks, points in this direction. The case of cohesin is paradigmatic, as its Mms21-dependent sumoylation is known to be required for establishment of sister-chromatid cohesion and sister chromatid recombination [34–36]. The growing list of damage-induced targets of the Mms21 ligase should definitely contribute to our understanding of this branch of the SUMO pathway.
Yeast cells were grown in YP (Yeast extract Peptone), or minimum complete medium (SC) to select for plasmid auxotrophies, plus the indicated carbon source at 2% final concentration. For auxin-induced degrons, IAA (SIGMA) was added to 1 mM from a 0.5 M stock in water.
Exponentially growing cells were arrested in G1 by addition of 10–8 M alpha factor (Genscript) at 30°C for 2 h or until >95% of cells were arrested in G1. Cells were then treated with 0.01% MMS (SIGMA) for 30 min to induce a pulse of alkylation damage, and cultures were released by washing cells three times and re-suspension in media containing 0.1 mg/ml pronase E (SIGMA). Synchronic cultures were routinely checked by FACS analysis. DNA was stained using 4,6,-Diamidino-2-phenylindole (DAPI) at 1 μg/ml final concentration in the presence of mounting solution and 0.4% Triton X-100 to permeablize cells. For fluorescence microscopy, series of z-focal plane images were collected with a DP30 monochrome camera mounted on an upright BX51 Olympus fluorescence microscope.
Epitope tagging of genes and deletions were performed as described [54,55]. Fusion of genes to an auxin-induced degron was done as described [56]. SMC5-9myc was amplified by PCR from the yeast strain YTR914 (SMC5-9myc:hphNT1) and cloned into the SphI/KpnI sites in YCplac22. Then, the ADH1p promoter was cloned upstream of the SMC5 gene by recombination cloning in recA+ MC1061 cells to yield plasmid pTR1094 (YCplac22-ADH1p-SMC5-9myc). MMS21 was cloned at the KpnI site in pTR797 (pYES2-3HA) and then moved to the SalI site in pRS315 to yield pNC2275 (pRS315-GALp-MMS21-3HA). All other SMC5- or MMS21-expressing plasmids used in this study are derived by site-directed mutagenesis from pTR1094 or pNC2275, respectively, using QuikChange XL (Stratagene). The E3-E2 strain was created by fusion PCR of two partially overlapping sequences, one containing an MMS21-3HA-UBC9 sequence from plasmid pTR1138 (pYES2-MMS21-3HA-UBC9), the other an hphNT or natNT cassette for integration at the 3′ end of the MMS21 gene. Transformants were checked by PCR and western blot. Oligos used for PCR are provided upon request.
Pull-down analysis of sumoylated proteins was performed essentially as described [34]. In all pull downs (except those shown in Figs. 2B, 2I, 4C, 4E, and 6B), cells were denatured during harvesting, and prior to snap-freezing, by sequential resuspension of the yeast pellet in 12% TCA and in 1M Tris-HCl pH 8. Cells were mechanically broken in 8M urea, and incubated with Ni-NTA beads in the presence of 15mM imidazole overnight at room temperature. Bound proteins were eluted with SDS-PAGE loading buffer. In all cases, SUMO pull downs were loaded in SDS-PAGE gels next to protein extracts to confirm the slower mobility of SUMO conjugates with respect to the unmodified protein. All proteins were resolved in 10% SDS-PAGE gels, except SMC proteins (7.5%), histone H3 (15%), and Fig. 4F (4%–15% gradient gel; BioRad). For co-immunoprecipitation analysis, protein extracts were prepared in EBX as previously described [34]. myc-tagged proteins were immunoprecipitated using anti-myc antibodies (9E10, Roche) coupled to protein G Dynabeads (from Invitrogen). HA-tagged and Flag-tagged proteins were immunoprecipitated using anti-HA Affinity matrix (Roche) and Anti-FLAG M2 Affinity Gel (Sigma). Chromatin Binding Assay was performed as previously described [57]. Antibodies used in western blot analysis are anti-HA (3F10; Roche), anti-Flag (M2; Sigma), anti-myc (9E10; Roche), anti-Rpd3 (ab18085; abcam), anti-hexokinase (H2035-01; USBiological), anti SUMO2/3 (Enzo Life Sciences), and anti-histone H3 (ab1791; abcam).
The 6his-T7-Smc5 and 6his-HA-Nse2 proteins were co-expressed in Rosetta 2 (DE3) pLysS cells (Novagen) from pET28a-SMC5 and pET15b-HA-MMS21, respectively. Bacterial cultures were grown at 37°C to A600 = 0.6, before IPTG addition. Cultures were then incubated for 3–4 h at 30°C and harvested by centrifugation. Cell pellets were equilibrated in Lysis Buffer (20% sucrose, 20 mM Tris, 8.0, 1 mM β-mercaptoethanol, 350 mM NaCl, 20 mM Imidazole, 1 mM PMSF, 0.1% IGEPAL), and cells were disrupted by sonication. Cell debris was removed by centrifugation (40,000×g). Hexa-histidine tagged proteins were purified by metal affinity chromatography using Ni-NTA resin (Qiagen) and eluted with 20 mM Tris (pH 8.0), 250 mM NaCl, 1 mM β-mercaptoethanol, and 250 mM imidazole. Fractions containing the Smc5-Mms21 heterodimer were further purified by gel filtration (Superdex 200; GE Healthcare).
For in vitro sumoylation assays, 100 ODs of GALp-SMC5 cells that express SMC5-9myc from a centromeric vector were shifted to glucose for 5 h, collected and stored at −80°C. After anti-myc immunoprecipitation, reactions were directly performed on complexes immobilized on protein G dynabeads (Invitrogen). Sumoylation was conducted either at 37°C with the human E1, E2, SUMO1, SUMO2, and SUMO3 proteins (Enzo Life Sciences Sumoylation kit, according to the supplier instructions), or at 30°C with recombinant yeast 6 histidine-tagged E1, E2, and Smt3, as previously described [43]. Reactions were run in parallel for wild-type and K75I mutant Smc5/6 complexes, started by addition of ATP, stopped with SDS-PAGE loading buffer, analyzed by western blotting and quantified with Image Lab (Bio-Rad). Since basal sumoylation of wild-type Smc5 is often detectable in the immunoprecipitates, the rate of sumoylation was calculated as the increase in sumoylation divided by the time of incubation in ATP. Small-scale sumoylation reactions of the C-terminal region of Nse4 (residues 246 to 402) were performed in a reaction mixture containing 20 mM HEPES pH 7.5, 5 mM MgCl2, 0.1% Tween-20, 100 mM NaCl, 1 mM dithiothreitol, 1 mM ATP, 150 nM hE1, 150 nM hE2, 32 μM hSUMO2, 16 μM Nse4(ct), and 300nM Smc5/Mms21 (wild type or K75I mutant); all proteins were tagged with six histidines, expressed in E. coli and purified by chromatography on Ni-NTA and gel-filtration columns. Reactions were conducted at 30°C, and samples were taken at different times after ATP addition and stopped with SDS-PAGE loading buffer. SDS-PAGE gels were stained with SYPRO-Ruby (Life Technologies) and the accumulation of Nse4(ct)-SUMO2 was quantified with Image J. Only those time points where the reaction progressed linearly were taken into consideration.
Smc5-Mms21 or Smc5(K75I)-Mms21 heterodimers were diluted to 30 ng/μl in 50 mM Tris-HCl, pH 7.5, 100 mM NaCl, 2 mM MgCl2 with or without 1 mM ATP and deposited on freshly cleaved mica in the presence of 50 μM spermidine. After 1 min the mica was rinsed with milli Q water and dried with filtered air. Samples were imaged in air by tapping mode SFM using a Nanoscope III or IV (Digital Instruments; Santa Barbara, CA). Silicon tips (NHC-W) with resonance frequency 310–372 kHz were from Nanosensors supplied by Vecco Instruments, Europe. Images were collected at 2 μm × 2 μm, and processed only by flattening to remove background slope. SFM images of Smc5-Mms21 or Smc5(K75I)-Mms21 heterodimers in the absence or in the presence of nucleotide were used for automatic particle detection with custom-made software written in MATLAB. In brief, particles are detected after finding their edges by calculating the gradient of the image intensity at each pixel. The height and area of detected objects was used to calculate a volume in arbitrary pixel units, after subtracting the average background signal of an identical area. Volume units were then normalized using as standard EcRNA polymerase (450 kDa, 678.8 ± 124 measured volume units).
Coiled coil and heptad-repeat registry prediction were performed as previously described [58]. For each protein sequence, 14-, 21-, and 28-residue windows were used to plot the coiled coil probability in the upper, middle, and bottom rows, respectively.
Fig. 1C: YTR337, YMB2210, pTR1094, pCG2788, pCG2821, pPM2750; Fig. 1D, E, F: YMB1840, pTR1094, pCG2788, pCG2821, pPM2750; Fig. 1G: YSM2465, pRS315, pTR2395, pTR2400; Fig. 1H: YMB1840, YMB1902, pTR1094, pCG2788: Fig. 2B: Y557, YMB794, YTR1444, YMB1556; Fig. 2C: YMB794, YMB2315, YMB2309; Fig. 2D: YMB2315; YMB2309; Fig. 2E: YTR337, YTR788, YTR786; Fig. 2F: YTR82, YMB1424, YMB1330, YMB1410, YMB1432; Fig. 2G: YMB1424, YMB1448, YMB1430, YMB1446, YMB1432, YMB2210; Fig. 2H: YMB1446; YMB1432; Fig. 2I: YTR854, YMB1117, YMB1345, YMB1120; Fig. 3A: YTR31, pTR1094, pTR1621, pNC1828; Fig. 3B: YMB1925, YMB1949, YMB1950, YMB1951; Fig. 3C: YTR907, pTR1094, pTR1621, pNC1828; Fig. 3D: YMB1905, pTR1094, pTR1621; Fig. 3E: YTR2373, YMB2214, Y557, YMB1902, pTR1094, pTR1621; Fig. 3F: YMB2136, pTR1094, pTR1621, pTR1828; Fig. 4B: YMB794, YMB793, YTR1766, YTR1768; Fig. 4C: Y557, YTR27, YMB794, YMB793, YTR1766, YTR1768; Figs. 4D and E: YPM1812, pTR1094, pTR1621; Fig. 4F: YTR3119, YCplac22, pTR1094, pTR1621, pTR3154; Fig. 5B: YTR31, YMB1840, pTR1094, pTR1621; Fig. 5D: 28S1, pNC2089, pNC2279, p6his-NSE4(ct); Fig. 6: 28S1, pNC2089, pNC2279; Fig. 7B: YTR907, pTR1094, pTR2158, pTR1967, pTR1969; Fig. 7C: YTR31, YMB2136, YCplac22, pTR1094, pTR2158; Fig. 7D: YTR29, pTR1094, pTR2158, pTR1967, pTR1969; Fig. 7E: Y557, YPM2506; Fig. 7F: YMB1840, YMB1902, pTR1094, pTR2158; Fig. 8A: Y557, YMB794, YPM2506, YPM2759, YPM2724; Fig. 8B: YMB794, YPM2506, YPM2759, YPM2724; S2A Fig.: YTR622, YMB628; S2B Fig.: Y557, Y570, YMB1840, pCG2788, YTR622, YTR628, YTR506, YTR3135; S2C Fig.: YTR628; S3 Fig.: Y557, YTR1435, YMB1452, YMB1454, YMB1456; S4 Fig.: YTR907, pTR1094, pTR1621; S5 Fig.: YMB1840, pTR1094, pTR1621; S6 and S7 Figs.: 28S1, pNC2089, pNC2279, pNC2094.
|
10.1371/journal.pbio.3000094 | Contingency in the convergent evolution of a regulatory network: Dosage compensation in Drosophila | The repeatability or predictability of evolution is a central question in evolutionary biology and most often addressed in experimental evolution studies. Here, we infer how genetically heterogeneous natural systems acquire the same molecular changes to address how genomic background affects adaptation in natural populations. In particular, we take advantage of independently formed neo-sex chromosomes in Drosophila species that have evolved dosage compensation by co-opting the dosage-compensation male-specific lethal (MSL) complex to study the mutational paths that have led to the acquisition of hundreds of novel binding sites for the MSL complex in different species. This complex recognizes a conserved 21-bp GA-rich sequence motif that is enriched on the X chromosome, and newly formed X chromosomes recruit the MSL complex by de novo acquisition of this binding motif. We identify recently formed sex chromosomes in the D. melanica and D. robusta species groups by genome sequencing and generate genomic occupancy maps of the MSL complex to infer the location of novel binding sites. We find that diverse mutational paths were utilized in each species to evolve hundreds of de novo binding motifs along the neo-X, including expansions of microsatellites and transposable element (TE) insertions. However, the propensity to utilize a particular mutational path differs between independently formed X chromosomes and appears to be contingent on genomic properties of that species, such as simple repeat or TE density. This establishes the “genomic environment” as an important determinant in predicting the outcome of evolutionary adaptations.
| We address how predictable evolution is at the DNA sequence level in nature by studying the parallel evolution of a phenotype that is well understood at the molecular level: the acquisition of sex chromosome dosage compensation in fruit flies. While female flies have two X chromosomes, the males have to compensate by up-regulating genes on their single X chromosome. They do this by using specific sequence motifs on the X chromosome to recruit a cluster of proteins and RNAs called the male-specific lethal (MSL) complex. However, different fruit fly lineages have independently co-opted the MSL complex to dosage compensate newly formed X chromosomes through the acquisition of hundreds of new binding sites, thereby providing independent replicas of the evolutionary process both within and between species. Moreover, dosage compensation has often evolved very recently, allowing us to infer the causative mutations by which the novel binding motifs arose. Genome sequencing and genomic occupancy maps of the MSL complex allowed us to infer the location of novel binding sites on the recently formed sex chromosomes of flies from the Drosophila melanica and D. robusta species groups. We show that species use diverse mechanisms to generate novel MSL-binding sites, including the use of presites, expansion of runs of simple sequences, or insertion of transposable elements. We also show that the propensity for using different types of mutations differs between lineages and depends on genomic properties of a species.
| What would happen if we “replay the tape of life” [1]? The question of whether adaptation follows a deterministic route largely prescribed by the environment or whether evolution is fundamentally unpredictable and can proceed along a large number of alternative trajectories has until recently been a fascinating problem that could not be addressed directly.
In the past decade, however, advances in DNA sequencing technology have allowed researchers to tackle this question using two complementary approaches. Experimental evolution of viruses, bacteria, and yeast, in combination with genome sequencing, has allowed direct identification of adaptive mutations in order to address the relative contributions of determinism and stochasticity in the evolutionary process [2]. Genomic analysis of populations experimentally evolved under controlled laboratory conditions has consistently revealed parallelism in which mutations in certain genes are repeatedly selected [3,4]. These studies are typically limited to systems that can be rapidly propagated in the lab and many relevant evolutionary parameters (such as environment, population size, etc.) are controlled by the experiment, and their applicability to natural systems is sometimes unclear [5].
Studies of parallel adaptations in the wild are a complementary approach to understanding the repeatability of evolution [2]. Organisms evolving under similar ecological conditions often evolve similar traits, and striking examples of genetic convergence at the DNA level have been recently discovered. For example, plant-feeding insects have independently and repeatedly colonized many different plant taxa, and highly diverged insect orders have evolved cardenolide resistance through the exact same amino-acid substitution in Na+/K+-ATPases [6]. Similarly, three distantly related lineages of snakes have convergently evolved resistance to the tetrodotoxin found in their prey via the same amino-acid mutation in a voltage-gated sodium channel [7]. Phenotypic convergence can also result from noncoding changes. Parallel evolution of trichome patterning in Drosophila [8] or wing pattern mimicry in Heliconius butterflies [9] both involved regulatory mutations that altered the expression pattern of a transcription factor. The parameters of convergent evolution in protein-coding genes are fairly well understood and often involve a small number of amino-acid mutations of large effect size that are constrained to specific regions of the protein because of pleiotropy. By contrast, how changes in cis-regulatory regions contribute to convergence is less well understood and hampered by our limited understanding of the global cis-regulatory structure of a phenotype [10]. Convergent regulatory evolution involves a much larger set of mutational targets and mechanisms: A single regulatory mutation affecting a transcription factor could act in trans to change the expression pattern of a suite of target genes (as observed in Drosophila and Heliconius) or multiple independent cis-acting mutations could act in concert to produce the selected phenotype. Furthermore, these regulatory mutations can arise via a large variety of mechanisms, from transposable element (TE) insertions to microsatellite expansions, and identifying causative adaptive mutations in nature is challenging, especially at noncoding DNA [10]. Moreover, adaptation to novel habitats often involves multiple selective agents whose relative importance is often unclear, making interpretations of convergent evolution (or a lack thereof) challenging [2].
Here, we address how predictable evolution is at the DNA sequence level in nature by studying the parallel evolution of a regulatory phenotype that is well understood at the molecular level: the acquisition of dosage compensation in fruit flies. Many species with heteromorphic sex chromosomes have evolved mechanisms to equalize the amount of gene product from the X chromosome in males and females [11]. In Drosophila, males compensate for reduced dosage of X-linked genes by hypertranscribing their hemizygous X chromosome through epigenetic modifications [12]. At the molecular level, this is achieved by recruiting the male-specific lethal (MSL) complex to numerous chromatin entry sites (CESs) on the X in a sequence-specific manner [13] (Fig 1A). The MSL complex targets a 21-bp long, GA-rich sequence motif that is enriched on the X chromosome (the MSL recognition element, or MRE) [13]. The complex then spreads from CESs to the rest of the X and induces chromosome-wide hyperacetylation of H4K16, which results in up-regulated transcription on the X chromosome [14].
MSL-mediated dosage compensation evolved over 60 million years (MY) ago and is conserved across Drosophila species [15,16]. However, different species in this genus co-opted the MSL machinery to evolve dosage compensation on newly evolved neo-sex chromosomes [15,16]. In particular, fusions between the ancestral sex chromosomes (that is, the original X and Y chromosome shared by all members of the genus Drosophila) and autosomes have repeatedly created so-called neo-sex chromosomes [17]. Once fused, the neo-sex chromosomes follow a distinct evolutionary trajectory over tens of millions of years until they obtain the classical properties of ancestral sex chromosomes: The neo-Y chromosome degenerates as its protein-coding genes are inactivated, and the neo-X chromosome is up-regulated to compensate for this gene dosage imbalance [12,18]. During this transition, the age of the neo-sex chromosomes broadly correlates with their level of differentiation.
Dosage compensation evolves on newly formed X chromosomes by co-opting the MSL complex through the acquisition of new MSL-binding sites (Fig 1B). We recently studied the evolution of MSL-binding sites in D. miranda, a model species for sex chromosome evolution that possesses two neo-X chromosomes that were formed about 13–15 MY and 1.5 MY ago, respectively [19,20]. We found that diverse mutational paths contributed to MSL-binding–site evolution [21], but the majority of novel MSL sites on the younger neo-X were created by insertions of a domesticated helitron TE containing the GA-rich sequence motif recognized by the MSL complex (the MRE) [22,23]. We also detected highly eroded remnants of a related TE at the much older neo-X chromosome of this species, in which dosage compensation evolved around 13–15 MY ago [22].
Independently formed neo-X chromosomes are faced with the same evolutionary challenge: to co-opt the existing MSL machinery and up-regulate hundreds of genes simultaneously in response to neo-Y degeneration. This creates a set of fascinating questions: How are new binding sites acquired on different positions along the neo-X chromosome of a lineage or on independently evolved neo-X chromosomes between species? Does evolution predominantly follow the same molecular path to evolve new MSL-binding sites, do species-specific solutions evolve to the same problem, or do independent binding sites evolve by diverse molecular mechanisms even within a lineage?
To address these questions, we use comparative genomic and functional analysis to infer the mutational path evolution has taken to acquire novel MSL-binding sites. We focus on Drosophila species in the D. melanica and D. robusta groups, which are promising new systems to study the independent rewiring of the MSL complex. Cytogenetic studies have shown that species in this clade have independently evolved neo-sex chromosomes [24], thus ensuring that dosage compensation evolved in parallel for these species. Phylogenetic dating suggests that neo-sex chromosomes in this group are young [24], which should allow us to identify the causative mutations that created novel MSL-binding sites on neo-X chromosomes.
In this study, we generate genomic data for five species from the D. melanica/D. robusta groups to identify the specific X-autosome fusions in these species and date their formation. We create maps of MSL occupancy for three species in which dosage compensation on the neo-X has evolved independently and recently, using Chromatin Isolation by RNA Purification (ChIRP; see Fig 1C). Comparative analysis allows us to reconstruct the mutations generating novel MSL-binding sites, and we infer both heterogeneity and convergence of binding site evolution within and between species. Our results demonstrate that evolution is highly opportunistic yet contingent on the genomic background. We show that species use a diverse spectrum of mutational events to generate novel MSL-binding sites, but the propensity for different types depends on genomic contingencies of a species.
Young neo-sex chromosomes of Drosophila have formed independently by fusions between autosomes and the ancestral sex chromosomes [12]. Over time, they acquire the stereotypical properties of the ancestral X and Y, and their repeated formation in different lineages at different time points allows us to contrast neo-sex chromosomes at various stages of differentiation [25,26]. Cytogenetic comparisons suggest that neo-sex chromosomes have evolved independently multiple times in species from the virilis–repleta radiation [24], but their neo-sex chromosomes were not examined at the genomic level. To identify neo-sex chromosomes and infer their evolutionary history, we performed Illumina whole-genome sequencing of males and females from five species in the D. robusta/D. melanica sister groups (Fig 2; S1 Table). We generated de novo assemblies from the female sequencing data and created a whole-genome alignment to identify orthologous regions between all five species, which we used to infer their phylogeny. Our genome-wide phylogenetic analysis confirms previous inferred relationships among members of these groups based on a handful of genes [24,27], with D. melanica and D. nigromelanica being sister species and D. micromelanica as their outgroup (forming the melanica group), whereas D. robusta and D. lacertosa are more distantly related (Fig 2A). We used sequence divergence between species pairs to roughly date their split times. Assuming a neutral mutation rate of 3.46 × 10‐9 per year [28], we estimate that species from the D. melanica subgroup split very recently; D. melanica and D. nigromelanica diverged roughly 4.3 MY ago, D. robusta split from the D. melanica species about 9.4 MY ago, and D. lacertosa diverged 16 MY ago (Fig 2B).
Neo-sex chromosomes were previously reported for four of the five species investigated here, with D. micromelanica lacking an X-autosome fusion. The neo-sex chromosomes of D. nigromelanica and D. melanica were thought to be homologous [24], while X-autosome fusions occurred independently in D. robusta and D. lacertosa [24]. We used male and female genomic coverage data to infer which autosomal chromosome arm became the neo-sex chromosome in each species and to estimate the date at which the fusion occurred (see Materials and methods). As expected, we identified the ancestral X chromosome (Muller element A, which is shared across Drosophila) by reduced male coverage in each species, and we confirmed the presence of a neo-X chromosome in D. nigromelanica, D. melanica, D. robusta, and D. lacertosa, as well as the absence of a neo-X in D. micromelanica (Fig 2C). Intriguingly, however, we find that a different chromosome arm formed the neo-sex chromosome in the melanica group: Muller element D (corresponding to chromosome 3L in D. melanogaster) became the neo-sex chromosome of D. melanica, while Muller element C (chromosome 2R in D. melanogaster) formed the neo-sex chromosome of D. nigromelanica. Thus, contrary to parsimonious interpretations based on cytological data, our genomic comparison shows that neo-sex chromosomes formed independently at least two times in the melanica group and imply that they are younger than previously assumed. We also confirmed that Muller element D (chromosome 3L) became the neo-sex chromosome of both D. lacertosa and D. robusta (Fig 2C), but the lack of X-autosome fusions in multiple species of the lacertosa and robusta subgroups indicates that these fusions originated independently [24]. Muller element D has become sex–linked multiple times in the Drosophila genus and in Diptera [29], suggesting that this chromosome may have an intrinsic propensity to become a sex chromosome.
The age of each species group with unique or shared neo-sex chromosomes sets a limit to the age of their chromosomal fusions (Fig 2B). Phylogenetic analysis suggests that D. melanica and D. nigromelanica diverged about 4.3 MY ago, implying that both species groups’ neo-sex chromosomes are younger than that age. Sequence divergence of homologous neo-sex–linked genes provides an independent estimate of their age: older neo-Y chromosomes harbor fewer genes (and fewer neo-X/neo-Y gene pairs), and levels of sequence divergence between orthologous gene pairs increases with the age of the neo-sex chromosome. We identified 118 pairs of homologous neo-sex–linked genes in D. nigromelanica, 20 in D. melanica, 16 in D. robusta, and 11 in D. lacertosa, and sequence divergence between homologous gene pairs increases with decreasing gene number (S2 Table, Fig 2B). This analysis suggests that D. nigromelanica has a relatively young neo-sex chromosome (mean synonymous site divergence [dS] = 0.16, 4.6 MY), followed by D. melanica (mean dS = 0.26, 7.5 MY), while the sex chromosomes of D. robusta and D. lacertosa formed about 11–15 MY ago (mean dS = 0.39, 0.52; 11.3 MY, 15.0 MY). The inferred ages of neo-sex chromosomes are in between those for the neo-sex chromosomes of D. miranda, whose older neo-sex chromosome (chromosome XR, Muller element D) fused to the ancestral X about 13–15 MY ago, and its neo-X/neo-Y (Muller element C) formed about 1.5 MY ago [20]. Note that the estimated age of neo-sex chromosomes may exceed the inferred speciation time even if they formed after speciation because of faster sequence evolution on the neo-Y chromosome. Selection is less efficient on the non-recombining neo-Y chromosome [20], and slightly deleterious synonymous mutations may thus accumulate faster; a higher mutation rate in males relative to females (male-driven evolution) could further increase the rate of neutral substitutions along the neo-Y branch [30].
Degeneration of the Y chromosome creates selective pressure to dosage compensate the X chromosome [31,32]. Our genomic analysis demonstrates that the neo-Y chromosomes of members of the D. robusta and D. melanica species group have very few genes left (S2 Table), and their neo-X chromosomes may thus have already acquired full dosage compensation. We gathered male and female RNA sequencing (RNA-seq) data from D. melanica and D. robusta heads to test whether expression levels of the newly formed neo-X chromosomes are similar between males and females (that is, whether they have evolved dosage compensation). Dosage compensation is absent in testis of Drosophila [33,29], and genes expressed in gonads show a nonrandom distribution on sex chromosomes [34–37]; we thus chose heads (a somatic tissue) to test for dosage compensation on the neo-X of D. melanica and D. robusta. We assigned D. melanica and D. robusta genomic scaffolds to chromosomes based on homology to D. virilis and identified genes directly from RNA-seq alignments to those same genomic scaffolds. Fig 3A shows male/female expression ratios for genes on the different chromosomes. Genes located on the neo-X in both species show very similar male/female expression ratios to genes on the ancestral X, suggesting that they have evolved full dosage compensation. Interestingly, however, genes located on the ancestral X as well as the neo-X in both species show slightly higher expression in females than males compared to autosomes (Fig 3A). This is consistent with the female-biased expression pattern of X chromosomes previously observed in Drosophila [34,35].
Comparative studies in Drosophila have shown that dosage compensation by the MSL complex is conserved across species [15,16,38,39]. Moreover, newly formed neo-X chromosomes evolve dosage compensation by acquiring novel MSL-binding sites that are able to recruit the MSL complex [38,39]. In D. melanogaster, two noncoding RNAs (roX1 and roX2) are part of the MSL complex, and a recently developed technique known as ChIRP has been successfully used to map MSL binding in several Drosophila species by isolating and sequencing DNA bound by the roX noncoding RNAs [38] (see Fig 1C).
Noncoding RNAs evolve quickly at the DNA sequence level but can be identified based on microsynteny and their male-specific expression [38]. Previous work has shown that while roX1 strongly localizes to the X chromosome in D. melanogaster, it shows much weaker X localization in other species of Drosophila (including D. virilis) [38]. RoX2, on the other hand, shows strong localization to the X chromosome in all Drosophila species studied so far [38] and has male-specific expression in dozens of species across the Drosophila phylogeny [38]; we thus focused on identifying roX2 in our target species. In D. virilis, roX2 is located between the protein-coding genes ari-1 and e(y)2. To identify the roX2 locus, we first searched each genome for synteny blocks likely containing roX2 based on the conserved location of ari-1 and e(y)2 homologs and then mapped RNA-seq data from D. melanica and D. robusta males and females in order to identify genomic regions showing male-specific expression (Fig 3B). Indeed, we found male-specific RNA-seq reads from our candidate region in both species, and we assembled the RNA-seq reads mapping to this location to generate the full-length roX2 transcripts from each species. We identified roX2 in D. nigromelanica based on homology to the D. melanica and D. robusta transcripts (S1 Fig).
To map MSL-binding sites, we designed nonoverlapping oligos against roX2 in D. nigromelanica, D. melanica, and D. robusta using a split oligo design [38]. We performed two independent ChIRP experiments with different nonoverlapping oligo sets (S3 Table) and generated 100-bp paired-end (PE) sequencing libraries for each oligo set as well as an input control. We aligned the sequencing reads to their respective genomes and identified genomic regions showing roX2 enrichment peaks. We found that roX2 binding is highly correlated between independent probe sets for each of the species (S2 Fig), and we identified MSL-binding sites as overlapping peaks between the two independent ChIRP experiments. We identified between 980 and 1,570 peaks in each species, with the majority located on either the X or neo-X chromosome (approximately 75%–95%, Fig 4A). As done previously [38,39], we set an enrichment threshold (see Materials and methods) to identify the subset of the most strongly bound peaks as CESs. We found similar numbers of CESs on the ancestral X (212–295) and the neo-X (258–290), suggesting that the neo-X chromosomes in each species have evolved full dosage compensation (Table 1). Note that the slightly higher number of CESs on the neo-X of D. melanica and D. robusta compared to their ancestral X could be due to the larger assembled size of their neo-X (the neo-X assembly is roughly 10% larger in both species relative to the ancestral X). Indeed, the density of CESs is similar for the ancestral X and the neo-X in both species and similar to that of the ancestral X in D. nigromelanica (that is, a CES every 58–68 kb; S4 Table). This is consistent with gene expression patterns, and the number of CESs per chromosome are similar to what has been found in other species with fully dosage-compensated X chromosomes [13,38]. A slightly lower density of CESs on the neo-X of D. nigromelanica (a CES every 83 kb; S4 Table), on the other hand, may reflect the evolution of incomplete dosage compensation on this more recently formed neo-X, as has been found on the young neo-X chromosome of D. miranda [39].
We used the software package MEME [40] to identify motifs enriched within roX2-bound CES regions. Molecular studies in D. melanogaster identified a GA-rich sequence motif that is targeted by the MSL complex [39]. Consistent with previous studies, we find a GA-rich sequence motif to be highly enriched on the ancestral X in every species studied, and the same GA-rich sequence motif is found on the newly evolved X chromosomes in species of the D. robusta and D. melanica species group (Fig 4B, S5 Table). A subset of CESs, called pioneering sites on the X (pion-X sites), are thought to be responsible for the initial recruitment of the MSL complex to the X chromosome [41]. Pion-X sites share the low-complexity GA-rich motif with canonical CESs (that is, they are generally a subset of MRE motifs) but contain a more complex CAC 5´ extension [41]. We searched each CES to identify matches to both the canonical MRE motif and the pion-X site motif (S3 Fig, S5 Table). Between 85%–91% of the CES sites identified contain either an MRE and/or pion-X site motif on both the ancestral X and on the neo-X (174/175/139 MRE and 84/65/57 pion-X sites on the ancestral X and 137/169/157 MRE and 73/82/86 pion-X sites on the neo-X in D. nigromelanica/D. melanica/D. robusta; Table 1). Thus, this confirms that the molecular machinery for dosage compensation is conserved in Drosophila and that the MSL complex has been independently recruited to transcriptionally up-regulate newly formed X chromosomes by acquisition of the MSL binding motif.
Contrasting CES evolution on homologous chromosomes that either ancestrally or convergently evolved dosage compensation allows us to study evolutionary patterns and constraints of CES conservation, acquisition, and turnover [21,22,38,39]. In particular, the ancestral X was fully dosage compensated by the MSL complex in the last common ancestor of D. melanica, D. nigromelanica, and D. robusta, and contrasting CES locations on the ancestral X chromosome can inform us of the evolutionary stability and turnover of shared CESs. In contrast, chromosome 3L independently evolved dosage compensation in D. melanica and D. robusta, and convergent acquisition of CESs on homologous positions may reflect either the presence of dosage-sensitive genes in a particular location and a strong need to evolve CESs or the presence of presites (that is, nucleotide sequences that resemble the MRE motif) and thus an easy mutational path to acquire CESs.
Overall, we find 72 CESs (about 28%) and 46 motifs (20%) to be syntenic between all three species on the ancestral X chromosome, and 50% of CESs (and 41% of motifs) are shared between the more closely related D. melanica and D. nigromelanica species pair (S6 Table). Each species from the D. melanica subgroup shares about 43% of its CESs with the more distantly related D. robusta. Inspection of species-specific CESs reveals that the majority of orthologous regions in other species (77% on average) tend to be bound by roX2, but at too low a level to pass our genome-wide threshold for CES identification; the majority of orthologous regions also contain the MRE/pion-X site motif (73% on average), suggesting that most CESs on the ancestral X are conserved between species.
On the other hand, about 25% of CESs evolved independently at syntenic positions on chromosome 3L in D. melanica and D. robusta (that is, 54 out of 251/243), suggesting that they arose from a presite present in their common ancestor. Indeed, for 90% of these sites, the orthologous region in D. nigromelanica, in which this chromosome is an autosome, shows homology to the MRE/pion-X site motif, indicating that the MSL-binding site evolved from a preexisting prebinding site. Five genomic regions evolved CESs independently in D. melanica and D. robusta by either GA expansions or insertions.
How are novel CESs acquired at the molecular level? Previous analysis of CES evolution in Drosophila revealed diverse mutational paths by which novel CESs originated in different fly species [21,22,38]. They include the use of prebinding sites (that is, sequence motifs that resemble the MRE motif and predate CES formation), simple GA expansions [42], or the spreading of CESs by TE mobilization [22]. Overall, we find that about 40%–60% of CESs on the newly formed neo-X chromosomes evolved from a presite. Thus, CESs often evolve from sequences that ancestrally resemble the MRE or pion-X site motif, and variability in sequence composition within similar GA-rich motifs [43] or changes to either the flanking sequences of CESs [44] or the repeat composition of the X [45], or possibly changes to the 3D organization [46], may allow these sequences not previously targeted by the MSL complex to function as CESs. In D. melanica and D. nigromelanica, most remaining CESs (approximately 25%) are created by simple GA expansions, whereas in D. robusta, 28% of novel CESs were created by TE insertions (Fig 5A and 5B). Thus, all three species utilized diverse mutational paths to evolve hundreds of novel CESs on their independently formed neo-X chromosome.
We also compared the mechanisms that gave rise to MRE motifs versus those that produced pion-X site motifs separately (S4 Fig). The relative frequencies of each mutational mechanism were similar for both types of motifs, with the exception that MRE motifs were much more likely to arise from GA expansions compared to pion-X site motifs (Fisher’s exact test P = 4.9 × 10‐8), consistent with the more complex sequence content of the pion-X site motif. A recent study suggested that CESs might originate by the co-option of GA-rich polypyrimidine tracts that are located at the 3´ 100 bp of introns and are used for splicing [38]. We found that 7%–12% of neo-X CES motifs are within the 3´ 100 bp of introns, the expected location for motifs arising from a co-opted polypyrimidine tract [38]. Both MRE and pion-X site motifs arose from these locations (55%–76% MRE, 24%–45% pion-X sites), and in roughly half of these cases, the polypyrimidine tract served as a ready-made motif (that is, presite), whereas in the remaining half, there were additional GA-expansion mutations, suggesting that the co-option of these features can involve multiple mutational paths (S5 Fig). The majority of MSL-binding motifs within CESs on neo-X chromosomes lie within gene bodies (introns and UTRs; S6 Fig), consistent with dosage compensation up-regulating gene expression [38].
We investigated the D. robusta TE-derived motifs in more detail and found that both MRE and pion-X site motifs were derived from a total of 18 different families (S7 Table). While most of these TE families were associated with only a single motif, we also identified a single family of elements that gave rise to 24 MSL binding motifs (17 MRE motifs and 7 pion-X site motifs). This TE family contains terminal inverted repeats with weak homology to those of two related galileo elements that have been identified in D. buzzatii and D. melanogaster (Fig 5B). The TE consensus sequence contains a match to the pion-X site motif near its 5´ end as well as an MRE motif near its 3´ end. We identified several hundred fragmented copies of this TE in our D. robusta genome assembly. These copies are highly enriched on the D. robusta neo-X chromosome (binomial test P = 3.2 × 10‐14, Fig 5C) and overlap CESs more often than expected by chance (Fisher’s exact test P < 2.2 × 10‐16). Of the 66 copies that are located on scaffolds that were assigned to Muller elements, 32 are from the neo-X chromosome, and 24 of these 32 copies lie within CESs. The average pairwise genetic distance between these copies is 0.47, which suggests they were active around the time when the neo-sex chromosomes of D. robusta were formed (mean dS: 0.39). Thus, enrichment of this TE on the neo-X of D. robusta, its occurrence within CESs and the presence of a binding motif for the MSL complex, and its inferred time of mobilization around the formation of neo-sex chromosomes strongly suggests that this TE was actively involved in dispersing MSL-binding sites along the neo-X.
What drives heterogeneity in CES acquisition across lineages? While D. melanica and D. nigromelanica evolved most novel MSL binding motifs by GA expansions, D. robusta utilized a TE for acquiring novel CESs. To investigate whether certain genomic factors prime a genome to preferentially evolve CESs by a particular mutational path, we analyzed the content of repetitive DNA—both microsatellite and TE density—in the different lineages. Interestingly, we found that D. melanica and D. nigromelanica differ in their overall repeat composition from D. robusta: they both show a higher density of simple repeats but a lower density of TE sequences (Wilcoxon test P < 2.2 × 10‐16 for both comparisons) (Fig 5D). Higher TE density in D. robusta is consistent with its larger assembled genome size (185 Mb in D. robusta versus 150 Mb in D. melanica and 164 Mb in D. nigromelanica). Thus, this observation is consistent with the notion that historical contingencies constrain evolutionary patterns of MSL-binding–site evolution. TEs may be more often utilized for rewiring regulatory networks in species with a higher number of TEs, but a larger TE burden may also contribute to increased genome sizes. On the other hand, a higher density of simple satellites in D. melanica and D. nigromelanica may have preadapted them to evolve novel MRE sites by GA-sequence expansion.
We took advantage of naturally occurring variation in sex chromosome karyotype in Drosophila species to study independent replicates of solving the same evolutionary challenge: to dosage compensate newly formed neo-X chromosomes by acquiring hundreds of MSL-binding sites in response to Y degeneration.
The independent acquisition of dosage compensation in Drosophila allows us to address several important questions in evolutionary biology and gene regulation: first, how repeatable is evolution? Evolutionary biologists have long debated the predictability of the evolutionary process. At one extreme, evolution could be highly idiosyncratic and unpredictable, since the survival of the fittest could occur along a great number of forking paths. Alternatively, constraints on evolution may force independent lineages to frequently converge on the same genetic solutions for the same evolutionary challenge. Second, how do regulatory networks evolve? And what is the contribution of TEs to regulatory evolution? Evolutionary innovations and adaptations often require rapid and concerted changes in regulation of gene expression at many loci [47]. TEs constitute the most dynamic part of eukaryotic genomes, and the dispersal of TEs that contain a regulatory element may allow for the same regulatory motif to be recruited at many genomic locations, thereby drawing multiple genes into the same regulatory network [48–50]. Third, what makes a binding motif functional? The genomes of complex organisms encompass megabases of DNA, and regulatory molecules must distinguish specific targets within this vast landscape. Regulatory factors typically identify their targets through sequence-specific interactions with the underlying DNA, but they typically bind only a fraction of the candidate genomic regions containing their specific target sequence motif. An unresolved mystery in regulatory evolution is what drives the specificity of binding to a subset of genomic regions that all appear to have a sequence that matches the consensus binding motif.
Several features make dosage compensation in Drosophila a promising system to tackle these questions. The genetic architecture for most adaptations—especially those involving regulatory changes—as well as the timing and exact selective forces driving them is generally little understood. In contrast, we have detailed knowledge of the molecular mechanism of dosage compensation in Drosophila. We know the cis- and trans-acting components of this regulatory network and the regulatory motif for targeting the MSL complex to the X. We have clear expectations of which genomic regions should acquire dosage compensation and about the timing and the evolutionary forces that drive wiring of hundreds of genes into the dosage-compensation network on newly evolved X chromosomes. Specifically, Y degeneration is a general facet of sex chromosome evolution, creating selective pressures to up-regulate X-linked genes in males. Dosage compensation should thus only evolve on neo-X chromosomes whose neo-Y homologs have started to degenerate and should evolve simultaneously or shortly after substantial gene loss has occurred on the neo-Y [31,32]. Indeed, comparative data in Drosophila support this model of dosage-compensation evolution. Drosophila species with partially eroded neo-Y chromosomes exist that have not yet evolved MSL-mediated dosage compensation, including D. busckii [25] and D. albomicans [51,52], lending empirical support to the notion that dosage compensation evolves in response to Y degeneration and not the other way round. Thus, our refined understanding of how, when, why, and where dosage compensation in Drosophila evolves makes this an ideal model system to study the repeatability of evolution and the evolution of regulatory networks.
Results from evolution experiments indicate that although evolution is not identical in replicate populations, there is an important degree of predictability [53]. Experimentally evolved populations under controlled, identical conditions consistently show parallelism in which mutations in certain genes are repeatedly selected [3,4]. However, organisms adapting to similar environments are not genetically identical, but their genome instead carries the legacy of their unique evolutionary trajectory, raising the question of how genomic differences affect genetic parallelism.
Sex chromosome–autosome fusions have independently created neo-sex chromosomes in different Drosophila lineages. This provides us with several independent replicates to study how, on the molecular level, evolution has solved the same adaptive challenge: acquiring hundreds of binding sites to recruit the MSL complex to newly formed X chromosomes. This allows us to quantify how much variation there is, both within and between species, in the underlying mutational paths to acquire hundreds of MSL-binding sites on neo-X chromosomes and identify genomic contingencies that will influence the repeatability of evolutionary trajectories. Importantly, neo-sex chromosomes of Drosophila are evolutionarily young (between 0.1–15 MY old), which allows us, in many cases, to infer the causative mutations that have resulted in the gain of a regulatory element and decipher the evolutionary processes at work to draw hundreds of genes into a new regulatory network.
Our results suggest that the evolution of MSL-binding sites is highly opportunistic but contingent on genomic background. In particular, we find that each independently evolved neo-X chromosome uses a diverse set of mutational pathways to acquire MSL-binding sites on a new neo-X chromosome, ranging from microsatellite expansions to the utilization of presites to TE insertions. However, different lineages differ with regards to the frequency of which mutational paths are most often followed to acquire novel binding sites, and this propensity may depend on the genomic background. In particular, we find that the two species with the higher density of simple repeats are more prone to utilize expansions in GA microsatellites to gain a novel MSL-binding site. In contrast, D. robusta has an elevated TE density compared to its sibling species, and we find that the dispersal of a TE has played an important role in the acquisition of MSL-binding sites on its neo-X chromosome. Thus, this suggests that the genomic background of a species predisposes it to evolve along a particular path, yet the evolutionary process is random and resourceful with regards to utilizing a variety of mutations to create novel MSL-binding sites. However, as discussed in the introduction, different phenotypes show drastic differences in their underlying genetic architecture, and the importance of genomic background likely differs among traits and species [2].
Evolutionary innovations and adaptations often require rapid and concerted changes in regulation of gene expression at many loci [47]. It has been suggested that TEs play a key role in rewiring regulatory networks, since the dispersal of TEs that contain a regulatory element may allow for the same regulatory motif to be recruited at many genomic locations [48–50]. A handful of recent studies have implicated TEs as drivers of key evolutionary innovations, including placentation in mammals [54] or rewiring the core regulatory network of human embryonic stem cells [55]. While these studies demonstrate that TEs can, in principle, contribute to the creation or rewiring of regulatory networks, they do not address the question of how often regulatory elements evolve by TE insertions versus by other mutations. That is, the importance of TEs in contributing to regulatory evolution is not known. Quantification of the role of TEs would require a priori knowledge of how and when regulatory networks evolve and a detailed molecular understanding of which genes are being drawn into a regulatory network and how. As discussed above, these parameters are well understood for dosage compensation in flies.
Our previous work in D. miranda has shown that a helitron TE was recruited into the dosage-compensation network at two independent time points. The younger 1.5-MY-old neo-X chromosome of D. miranda is in the process of evolving dosage compensation, and dozens of new CESs on this chromosome were created by insertions of the ISX element [22]. We showed that the domesticated ISX TE gained a novel MRE motif by a 10-bp deletion in the ISY element, which is a highly abundant TE in the D. miranda genome [22]. We also found the remnants of a related (but different) TE at CES on the older neo-X of this species (which formed roughly 13–15 MY ago), but the TE was too eroded to reconstruct its evolutionary history. Here, we identified another domesticated TE that was utilized to deliver MSL-binding sites to a newly formed neo-X chromosome, but no significant TE contribution was found for MSL-binding site evolution in two independent neo-X chromosomes.
Our data shed light on the question of when we expect TEs to be important in regulatory evolution. For TEs to contribute to regulatory rewiring, two conditions have to be met: a regulatory element (or a progenitor sequence that can easily mutate into the required binding motif) needs to be present in the TE, and that TE needs to be active in the genome (and not yet silenced by the host machinery). TEs undergo a characteristic life cycle in which they invade a new species (or escape the genome defense by mutation) and transpose until they are silenced by the host genome [56]. Once a TE is robustly repressed, it no longer can serve as a vehicle to disperse regulatory elements, so the time window when a particular TE family can be domesticated is probably short and needs to coincide with a necessity to disperse regulatory motifs. A high TE burden does increase that chance, but at a cost: maintaining active TEs in the genome allows a rapid response to evolutionary challenges but also creates a major source of genomic mutation, illegitimate recombination, genomic rearrangements, and genome size inflation [57].
Our findings support this view of a TE tradeoff. The ISY element in D. miranda is the most highly abundant transposon in the D. miranda genome and is massively contributing to the degeneration of the neo-Y in this species [26]. Indeed, our genomic analysis has revealed >20,000 novel insertions of the ISY element on the neo-Y, often within genes [26]. Yet, it contained a sequence that was only one mutational step away from a functional MSL-binding site (that is, a single 10-bp deletion), and domestication of this element allowed for the rapid dispersal of functional binding sites for the MSL complex along the neo-X. The domestication of the TE in D. robusta occurred too long ago for us to reconstruct its exact evolutionary history and the potential damage its mobilization may have caused while it was active. However, consistent with a tradeoff that the host genome faces, we find that D. robusta has a higher TE density than its sister species and also a considerably larger genome size, yet a TE contributed to wiring hundreds of genes into the dosage-compensation network on its neo-X.
Perhaps surprisingly, in many instances, we are unable to detect specific mutations that would generate a novel MSL binding motif. Instead, we find that functional MSL-binding sites are derived from presites containing the GA-rich motif that was already present in an ancestor in which the neo-X is autosomal and in which these sequences do not recruit the MSL complex. The MSL binding motif is only modestly enriched on the X chromosome compared to the autosomes (only approximately 2-fold), and only a small fraction of putative binding sites are actually bound by the MSL complex [13]. The dosage-compensation machinery shares this characteristic with many other sequence-specific binding factors whose predicted target motifs are often in vast excess to the sites actually utilized. It has been speculated that other genomic aspects, such as chromatin context or the 3D organization of the genome, could help to distinguish between utilized and nonutilized copies of a motif. Our finding that a large number of sites can acquire the ability to recruit the MSL complex, without any apparent associated changes at the DNA level, supports the view that epigenetic modifications or changes to the 3D architecture of the genome help to ultimately determine which putative binding sites in the genome are actually utilized [44,46]. In D. melanogaster, the X chromosome has a unique satellite DNA composition, and it was suggested that these repeats play a primary role in determining X identity during dosage compensation [45]. Furthermore, localization of the MSL complex to MREs is dependent on an additional cofactor, the CLAMP protein [58]. CLAMP binds directly to GA-rich MRE sequences and targets MSL to the X chromosome but also binds to GA-rich sequence elements throughout the genome [58]. Recent work has shown that variability in sequence composition within similar GA-rich motifs drive specificity for CLAMP binding [43], and variation within seemingly similar cis elements may also drive context-specific targeting of the MSL complex. Future investigations of changes in the chromatin level, the repeat content, and the genomic architecture of these newly formed sex chromosomes will help to resolve this outstanding question.
DNA was extracted from single flies using the Qiagen PureGene Kit (Qiagen, Hilden, Germany), and two PE Illumina sequencing libraries (male and female) were prepared for each species. The Illumina Nextera library prep kit (Illumina, San Diego, CA, USA) was used for D. melanica and D. robusta (150-bp PE reads), while the Illumina TruSeq kit (100-bp PE reads; Illumina) was used for the remaining species. Genome assemblies were generated for males and females separately by first error-correcting reads using BFC [59] and then assembling the corrected reads using IDBA [60]. A whole-genome alignment was constructed using the female assemblies for the five species studied here plus D. virilis using Mercator [61].
To create a whole-genome phylogeny, the D. virilis genome was split into 250-bp windows. Each window was extracted from the Mercator whole-genome alignment, and windows were retained if the aligned sequence from each species contained no more than 10% of positions as gaps. Retained windows were further filtered to ensure that each window was at least 1 kb from the closest neighboring window. These windows were concatenated to produce a multiple-sequence alignment containing 1.1 million positions. The RAxML rapid bootstrapping algorithm [62] was used to produce a maximum likelihood phylogeny from this alignment.
Chromosome assignments for D. virilis scaffolds were obtained from [63]. The scaffolds from each species studied here were assigned to Muller elements based on their alignment to D. virilis scaffolds from the Mercator whole-genome alignment.
To determine which Muller elements are X linked in each species, male and female Illumina reads were aligned separately to the female genome assemblies using bowtie2 [64], and male/female coverage ratios were calculated for each female scaffold.
Y-linked scaffolds were identified from the male assemblies using YGS [65].
For each sex, heads were removed from five flies, flash frozen in liquid nitrogen, and placed into Trizol for RNA extraction. The Illumina TruSeq RNA kit was used to prepare unstranded, single-end 50-bp sequencing libraries for each sex. RNA-seq data were aligned to the female reference genome assembly using Hisat2 [66], and gene models were generated from the merged male + female spliced alignments, along with normalized expression values, using StringTie [67]. Male and female RNA-seq read coverage was used to identify the location of roX2 in D. melanica and D. robusta (see Fig 3B), and roX2 transcripts were extracted from the genome assemblies based on the StringTie gene models. The D. nigromelanica roX2 transcript was identified based on homology to the D. melanica transcript using Exonerate [68].
For each species, D. melanogaster peptides were searched against the set of neo-X- and Y-linked scaffolds using a translated BLAST search [69]. The resulting neo-X- and Y-linked gene models were further refined using Exonerate, and their coding sequence was aligned using the codon model in PRANK [70]. Ks values were calculated for each neo-X/Y pair using KaKs_Calculator [71]. For species divergences, orthologous genes were identified using the D. robusta gene models from StringTie and the Mercator whole-genome alignment. Refinement of gene models, alignment, and Ks values were obtained as described above.
We used a neutral mutation rate estimate of 3.46 × 10‐9 per base per generation, which was experimentally determined from D. melanogaster [28]. The species studied here have a generation time that is roughly twice as long as D. melanogaster, and we therefore used the lower bound of the estimate of the number of generations per year for Drosophilids (5 generations) [72] to convert the mutation rate to time-based units (1.73 × 10‐8 mutations per base per year).
ChIRP sequencing libraries were prepared according to the published protocol [73] using the Drosophila-specific modifications described in [38]. For each species, ChIRP libraries were prepared from 2 different pools of 6 probes (S3 Table), which were tiled across the roX2 transcript. Input control libraries were also prepared for each species by extracting DNA from an aliquot of the cell lysate immediately prior to probe hybridization. Wandering third-instar larvae were used for D. melanica and D. robusta. Because of the difficulty in collecting sufficient larvae for D. nigromelanica, adult males were used instead. 100-bp PE Illumina reads were generated for each pool for each species and aligned to the female reference genome assembly using bowtie2. Peaks of roX2 binding were identified by running MACS [74] on each pool separately, along with the control library, and a final set of peaks was generated by retaining only the subset of peaks that were identified in both pools.
The ChIRP libraries varied in overall signal versus background, likely because of differences in the hybridization efficiency of the different probe sets. For each species, we calculated the average fold enrichment (treatment versus control) across all peaks as a measure of overall ChIRP signal. The D. robusta libraries showed the highest signal, and we used the same enrichment threshold (20) that was previously used to identify CESs from roX2 peaks [38]. For the remaining species, we identified CESs by scaling the enrichment threshold in proportion to our measure of ChIRP signal.
We defined the location of CESs as a 500-bp region centered on the summit of the roX2-bound peak. We extracted the sequence from these regions for the ancestral and neo-X chromosomes separately and used the ZOOPS model in MEME [40] to identify enriched sequence motifs. We used FIMO [40] to determine the location of MRE and pion-X site motifs within each CES and assigned each CES as containing either an MRE motif or a pion-X site motif (whichever match had a higher score) or no motif (if there was no match to either motif). We used the Mercator whole-genome alignment to assess orthology of CES as well as individual motifs. For each neo-X CES, we manually viewed the alignment of its sequence motif with the orthologous sequences from the other five species to determine the mutational mechanism that gave rise to the motif.
De novo TE identification was performed for each species using RepeatModeler (https://github.com/rmhubley/RepeatModeler). To identify the genomic locations of TE families, RepeatMasker (https://github.com/rmhubley/RepeatMasker) was used with the RepeatModeler consensus sequences as the repeat library. D. robusta TEs that overlapped CESs were further classified using CENSOR [75] and RepBase [76].
For each species, we used RepeatMasker to separately identify simple repeats and TEs. Because the percentage of the genome assembly that falls into these two categories will be affected by differences in total assembly size between species, we used an alternative approach for determining the density of these repeat classes. For each species, we permuted the location of 1,000 1-kb windows for 1,000 permutations. For each iteration, we determined the number of windows that overlapped a simple repeat and the number of windows that overlapped a TE, which we termed the “repeat density metric.”
|
10.1371/journal.ppat.1005974 | Dual microRNA Screens Reveal That the Immune-Responsive miR-181 Promotes Henipavirus Entry and Cell-Cell Fusion | Hendra and Nipah viruses (family Paramyxoviridae, genus Henipavirus) are bat-borne viruses that cause fatal disease in humans and a range of other mammalian species. Gaining a deeper understanding of host pathways exploited by henipaviruses for infection may identify targets for new anti-viral therapies. Here we have performed genome-wide high-throughput agonist and antagonist screens at biosafety level 4 to identify host-encoded microRNAs (miRNAs) impacting henipavirus infection in human cells. Members of the miR-181 and miR-17~93 families strongly promoted Hendra virus infection. miR-181 also promoted Nipah virus infection, but did not affect infection by paramyxoviruses from other genera, indicating specificity in the virus-host interaction. Infection promotion was primarily mediated via the ability of miR-181 to significantly enhance henipavirus-induced membrane fusion. Cell signalling receptors of ephrins, namely EphA5 and EphA7, were identified as novel negative regulators of henipavirus fusion. The expression of these receptors, as well as EphB4, were suppressed by miR-181 overexpression, suggesting that simultaneous inhibition of several Ephs by the miRNA contributes to enhanced infection and fusion. Immune-responsive miR-181 levels was also up-regulated in the biofluids of ferrets and horses infected with Hendra virus, suggesting that the host innate immune response may promote henipavirus spread and exacerbate disease severity. This study is the first genome-wide screen of miRNAs influencing infection by a clinically significant mononegavirus and nominates select miRNAs as targets for future anti-viral therapy development.
| The henipaviruses Hendra and Nipah are bat-borne paramyxoviruses that are highly pathogenic in humans. Until recently the constraints of working at biosafety level 4 had hindered the large scale study of host factors associated with henipavirus infection. MicroRNAs are a class of single-stranded non-coding RNAs that regulate biological processes in eukaryotes. An emerging body of evidence suggests that host microRNAs may favour infection of vertebrate RNA viruses. We have performed high-throughput agonist and antagonist screens at biosafety level 4 to identify host-encoded microRNAs impacting henipavirus infection in human cells. Members of the miR-181 and miR-17~93 families strongly promoted Hendra virus infection and appear to suppress multiple antiviral host molecules. Infection promotion is primarily mediated via the ability of miR-181 to repress Eph receptors that negatively regulate henipavirus glycoprotein-mediated cell-cell fusion. This study is the first large-scale screen of host-encoded microRNAs influencing infection by a clinically significant mononegavirus, and of a BSL-4 virus, and supports the emerging notion that host miRNAs can play a role in supporting infection of RNA viruses.
| Hendra virus (HeV) and Nipah virus (NiV) are highly pathogenic zoonotic paramyxoviruses belonging to the genus Henipavirus [1]. First isolated in Australia in 1994, HeV disease has caused seven clinically confirmed human cases with four fatalities. NiV initially appeared in Malaysia in 1998–1999, resulting in 105 human fatalities. Since 2001, recurring outbreaks of NiV have been reported in South Asia, resulting in more than 211 deaths and an average case-fatality rate of approximately 75% [2, 3]. Both bat-borne henipaviruses cause severe respiratory illness and encephalitis in humans, however there is a lack of therapies and vaccines. With high fatality rates emphasising the need for effective anti-viral strategies [4–6], a better understanding of henipavirus biology is required.
Viruses may co-opt or alter a range of host cell processes that optimise replicative efficiency. One such process is the RNA interference (RNAi) pathway [7]. Conventionally, in chordates RNAi involves the base-pairing of small non-coding microRNA (miRNA) molecules in a multi-protein complex to complementary mRNA sequences, often resulting in post-transcriptional silencing of host gene expression [8]. Some DNA viruses (i.e. herpesviruses) in particular, which also encode their own viral miRNAs, are known to subvert this fundamental host process to promote infection [7]. For RNA viruses however, the pro-viral roles of host miRNAs remain poorly characterized. Up until recently, the general thought was that the multifaceted dependence of hepatitis C virus infection on hepatocyte-specific miR-122 is the exception, not the rule, for RNA viruses [9, 10].
More recently, a few high profile studies have highlighted that the usurping of host miRNAs by RNA viruses might previously have been underappreciated. Trobaugh et al. showed that the alphavirus Eastern equine encephalitis virus (EEEV) utilizes host-derived miR-142-3p to define cell tropism and to suppress innate immunity, indirectly promoting neuropathogenesis [11]. A comprehensive survey of 15 RNA viruses from 7 families identified miR-17 and let-7 binding to pestivirus 3’ UTR as critical for enhanced viral translation, RNA stability and virus production [12]. The Argonaute protein, a key component of functional miRNA complexes, was also found to be associated with viral RNA of virtually all of the viruses assessed, including paramyxoviruses [12]. These Argonaute-viral RNA interactions also often exhibit preferential clustering on the viral subgenomes, implying specificities in the miRNA targeting. Enterovirus infection induces host miR-141 expression, which is then co-opted by the virus to silence cellular translation initiation factor eIF4E, resulting in host translational shutoff [13]. One emerging concept from such studies is the sequestration or “sponging” of anti-viral host miRNAs by genomes of some RNA viruses to derepress cellular transcripts that might enhance infection [9, 12]. These reports suggest that RNA viruses can adopt host miRNAs for their own utility via a diversity of mechanisms, and that this aspect of virus-host interactions is currently understudied.
With technological advancements in high-throughput techniques making the comprehensive study of both physical and genetic virus-host interactions a possibility [5], we have started executing functional genomics screens using fully infectious biosafety level 4 agents [14]. Despite the power of functional genomics as a research tool, thus far only two comprehensive RNAi screens investigating the contributions of miRNAs to pathogenesis of RNA viruses have been reported [15, 16]. No such study has been done for BSL-4 viruses and for any of the medically relevant mononegaviruses, such as paramyxoviruses or filoviruses. In light of recent studies underscoring the potential significance of miRNAs for RNA virus replication as well as the therapeutic promise of miRNA antagonists [6, 17], we sought to address this gap in our knowledge of virus-host interplay.
Here we present findings from two high-throughput genome-wide screens, conducted at BSL-4, of host-encoded miRNAs associated with HeV infection. The screens, in addition to subsequent validation work, demonstrate a key role for miR-181 family members in regulating henipavirus syncytia formation and infection, and suggest several host miRNAs, including miR-17~93, as potential candidates for novel therapeutic targets.
To identify host-encoded miRNAs that regulate HeV infection, we performed two complementary high-throughput screens at BSL-4 that targeted 834 human host-encoded miRNAs (Fig 1A). This first involved the reverse transfection of HeLa cells with a library of 1,239 synthetic miRNA agonists (i.e. mimics), which are double-stranded RNA molecules that functionally imitate native miRNAs, to over-express each of 834 miRNAs in a 384-well plate format. In a concurrent screen silencing each of the 834 miRNAs, HeLa cells were transfected with a library consisting of 1,225 antagonists (i.e. inhibitors), which are single-stranded RNA molecules that bind and sequester native mature miRNAs [18]. After library preparation and transfection were performed under BSL-2 conditions, one set of the daughter plates for each screen was moved to BSL-4 (red box, Fig 1A). 72 h after transfection, these cells were infected with a recombinant HeV that expresses the firefly luciferase reporter gene [19]. After 24 h, a luminometer was used to measure luciferase expression of the infected cells for both screens at BSL-4. The impact that each miRNA agonist or antagonist had on reporter expression was normalized to values from mock-transfected cells, and then expressed in terms of a robust Z-score, which is a commonly used measure of hit identification for RNAi screens [20, 21]. Z-scores of 2 or greater indicate an increase in infection relative to control, whereas Z-scores of -2 or lower indicate a decrease in infection.
The other set of daughter plates with transfected cells were processed for cell viability analysis by nuclei quantification using an automated fluorescence microscope (Fig 1A). Cells that were transfected with a siRNA targeting the PLK1 gene served as a positive control for cell death. 54 of the agonists and none of the antagonists caused significant cell death relative to mock-transfected cells (≥70% cut-off) and were thus subsequently eliminated from the robust Z-score analysis and final hit list generation (S1 Table and S2 Table).
The agonist and antagonist screens identified 35 and 61 miRNAs respectively that significantly promoted HeV infection, and 19 and 83 miRNAs respectively, that significantly inhibited virus infection (Fig 1B and 1C, S1 Table and S2 Table). Eight miRNAs exhibited pro-viral characteristics in both agonist and antagonist screens (Fig 1D), including all four members of the miR-181 family. Conversely, both screens identified miR-532 as a miRNA that inhibits infection. In regards to miRNAs that promote virus infection, screen results from two miRNA families were notable. All four members of the miR-181 family significantly promoted HeV infection (Fig 2A). These miRNAs all share the same seed sequence (ACAUUC), implying significant congruency in function(s). The scale of the pro-viral impacts of miR-181 members is especially remarkable if we compare their effects to that of miR-146a (Fig 2A), which we previously validated as pro-viral for HeV [22]. In addition to miR-181, most members of the miR-17~93 family were pro-viral (Fig 2B). The seed sequence of the miRNAs in this family (AAAGUG) is distinct from that of miR-181.
MiRNAs regulate gene expression by binding to complementary sequences typically located in the 3’ untranslated region (3’ UTR) of the mRNA target [23–26]. Depending on the degree of complementarity, this generally results in the suppression or degradation of target mRNA, thereby preventing encoded proteins from being translated [23, 25, 26]. As each miRNA can act as a suppressor of many target genes, we hypothesized that miR-181 and miR-17~93 families promoted henipavirus infection by suppressing multiple anti-viral host genes. To test this hypothesis, we firstly mined the miRTarbase database [27] to identify all experimentally-validated target genes for miR-181 and miR17~93 families. Next, the effects of these genes on HeV infection, as represented in robust Z-scores, were cross-referenced from results of our published genome-wide siRNA screen that identified pro- and anti-viral host genes for henipaviruses [14] (S3 Table). This analysis demonstrates that for both miRNA families these genes were more likely to be anti-viral (Fig 2C and 2D). For instance, the ratio of anti-viral to pro-viral hits for validated miR-181 targets was 2.2 to 1 (Fig 2C). In contrast, this ratio for all unbiased gene hits in the entire screen was 1 to 1 [14]. Collectively, these data suggest that the net outcome of miR-181 or miR17~93 expression is a cellular microenvironment that is more conducive for henipavirus infection. The results also indicate a reasonable level of congruency between our miRNA and siRNA gene screen datasets.
We also sought to determine whether miR-181 preferentially regulates the expression of host proteins localized in a particular subcellular compartment. To this end, the list of experimentally-validated miR-181 targets (n = 78 genes) was obtained from miRTarBase was subjected to annotation enrichment analysis using the DAVID web service. Functional annotation clustering analysis was performed using default settings (Fisher’s exact test to calculate p-values, followed by multiple testing correction using the Benjamini method). This analysis demonstrated an enrichment of miR-181 target genes associated with the nucleoplasm (p = 4.6e-5), while proteins associated with plasma membrane localization were not significantly enriched (p = 0.7).
All four members of the miR-181 family exhibited consistent pro-viral phenotypes in both the agonist and antagonist screens (Figs 1D and 2A). We thus decided to investigate the role of this miRNA family in henipavirus infection further. Firstly, since the screens were performed using a recombinant reporter HeV, we validated the observations using a wild-type virus. Screen results suggested that miR-181d is one of the most pro-viral members of the family (Fig 2A, S1 Table and S2 Table), hence miR-181d was chosen as a representative member of the miR-181 family in the majority of subsequent experiments.
HeLa cells were transfected with miR-181d-specific agonists to effectively over-express miR-181d. Non-targeting miRNA agonists (miNEG), or transfection reagent devoid of agonist (mock), were also included as negative controls. At 72 h post-transfection, cells were infected with wild-type HeV and incubated for 24 h. The cell supernatants were then collected and applied to a TCID50 assay to quantify infectious virus titre. As indicated in Fig 3A, the results showed a 350% increase in HeV infectious titres in cells transfected with miR-181d agonists compared to control miRNA agonists.
In addition to TCID50 measurements of HeV titres in cell supernatants, the impact of miR-181d agonists on the proportion of infected cells and HeV protein production was investigated using quantitative immunofluorescence imaging. Transfected cells were infected with HeV for 24 h, before being fixed and stained with fluorescently labelled antibodies. The HeV phosphoprotein (P) was selected for analysis due to its high protein expression levels during infection [28, 29]. Compared with control miRNA agonists, cells transfected with miR-181d agonists showed a significant increase in both the percentage of cells infected (>50% increase) (Fig 3B) and also the levels of virus antigen per infected cell (>25% increase) (Fig 3C). In support of the quantitative data, both an increase in HeV P protein concentration and also syncytia formation (characteristic of paramyxovirus infection) could be visually observed by fluorescence microscopy (Fig 3D).
To assess whether the pro-viral effects of miR-181 are specific to HeV, the in vitro activity of miR-181d agonists were tested on a range of viruses from different subfamilies of the Paramyxoviridae family. These included the closely related Nipah virus (subfamily Paramyxovirinae, genus Henipavirus), but also measles virus (subfamily Paramyxovirinae, genus Morbillivirus), mumps virus (subfamily Paramyxovirinae, genus Rubulavirus) and respiratory syncytial virus (RSV, subfamily Pneumovirinae, genus Pneumovirus). Influenza A/WSN/33(H1N1), an orthomyxovirus, was also included to compare with a virus from a different family. TCID50 measurements of the supernatants collected from transfected cells infected with NiV revealed a significant >300% increase in virus production in cells treated with miR-181d agonists compared to control miRNA agonists (Fig 4). On the other hand, infectivity assays showed no significant differences in virus titre between cells transfected with miR-181d or control agonists and infected with either measles virus, mumps virus, RSV or influenza A/WSN/33 (H1N1). These results indicate that the enhancement effects of miR-181 are specific to the henipavirus genus.
In order to narrow down the possible mechanisms by which miR-181 promotes henipavirus infection, we next sought to delineate the part of the virus life cycle at which miR-181 promotes infection. We first looked at whether viral RNA synthesis was induced by miR-181 during a single round of HeV infection. Cells were transfected with miR-181 agonists, and then infected with a high MOI (5) of HeV. At 0, 12 and 24 h post-infection (h.p.i.), cell lysates were harvested, and intracellular viral RNA levels were assessed by quantitative RT-PCR. Our previous studies have shown that in HeLa cells the first cycle of HeV infection is completed by 24 h but not at 12 h [14]. Nonetheless, here cell supernatants were also collected and virus titres from all time-points were determined by TCID50 assay (S1 Fig). Congruent to our previous observations, production of progeny virions was only detected at 24 h.p.i.
At 12 h.p.i., a four-log10 increase in viral RNA above inocula levels was observed in cells transfected with either miNEG or siNEG (Fig 5A). This increase could be suppressed by siRNA-mediated knockdown of the entry receptor for HeV, ephrin-B2. In contrast, pre-treatment of cells with miR-181d agonists increased the amount of HeV RNA by approximately 3-fold relative to cells treated with control agonist. Thus, miR-181 promotes henipavirus infection at, or prior to, the step of viral RNA synthesis. Similar trends in viral RNA levels amongst the treatment groups were also observed at 24 h.p.i.
Changes in viral RNA synthesis as measured by qRT-PCR during a single cycle infection could be due to effects on viral entry, genome replication, viral transcription or translation. To address whether miR-181 promotes entry of henipaviruses, a cell-cell fusion assay was performed using 293T effector cells expressing HeV F and G-glycoproteins [14]. Target cells (HeLa) were pre-transfected with either miR-181d agonist or the control agonist (miNEG), stained with a live cell membrane dye, and co-cultured for 24 h with effector cells expressing both glycoproteins or HeV G alone. Syncytia formation was imaged after fixation of the co-cultures and immunofluorescent staining for surface G glycoproteins. As controls, target cells were pre-treated with either siRNA duplexes targeting ephrin-B2 (positive control for impaired cell entry) or siNEG.
As expected, cells transfected with either siNEG or miNEG formed many multinucleated cells with the effector cells (Fig 5B). In contrast, cells with depleted ephrin-B2 either did not fuse with the effector cells, or fused into smaller syncytia with less nuclei. Additionally, effector cells expressing G singly did not develop syncytia with any of the target cells. Interestingly, miR-181-transfected cells induced substantially more, and larger, syncytia. Cell-cell fusion was so extensive, in a few instances polykaryons with at least 100 nuclei were observed. The extent of fusion for each treatment group was measured using automated image analysis. Results corroborate what was observed by visual inspection, indicating that miR-181 overexpression induced a drastic 9- to 10-fold increase in fusion events relative to control (Fig 5C). Conversely, ephrin-B2 knockdown caused about a 90% reduction in syncytia formation.
Considering the striking pro-fusogenic activity of miR-181, we wondered whether this effect is unique to the miR-181 family of miRNAs. The miR-17~93 family was another high-ranking pro-viral hit from the dual miRNA screens (Figs 1D and 2B). We decided to test the impact of miRNAs from this family on henipavirus-induced cell-cell fusion.
We first sought to validate the pro-viral effects of the miR-17~93 family using wild-type HeV. Agonists for two representative members of the miR-17~93 family, miR-17 and miR-93, were transfected into HeLa cells, and the permissiveness of these cells to HeV infection was evaluated using quantitative immunofluorescent analysis. The proportion of infected cells was 2-fold higher in the cells treated with miR-17 agonists compared to cells treated with control agonists (Fig 6A). This was commensurate with the uptick in infection ratio in the miR-93-treated cells (170% relative to control). Virus yields in the supernatants of miR-17 agonist-treated cells were also enhanced, as measured by TCID50 assays (Fig 6B). For comparison, the activity of these agonists on another paramyxovirus, RSV, was also analysed using quantitative immunofluorescence microscopy (Fig 6A). Akin to the impact of miR-17~93 on HeV infection, we found that the ratios of infected cells were also significantly higher in the miR-17 and miR-93 agonist transfected cells (189% and 180% of control, respectively). These results further validate the datasets from the complementary screens (Fig 1), and indicate that members of the miR-17~93 family are indeed promotive for wild-type henipavirus infection. However, and rather intriguingly, unlike miR-181 (Fig 4), members of the miR-17~93 family appear to also exhibit pro-viral effects on a paramyxovirus from a different subfamily than the henipaviruses.
Once the pro-viral ability of miR-17 was established, we then performed the cell-cell fusion assay to assess the fusogenic effects of miR-17 agonists. Consistent with our previous observations (Fig 5B), target cells depleted of ephrin-B2 exhibited muted fusion activity relative to control cells (Fig 6C). In contrast, target cells treated with miR-181d agonists formed larger multinucleated cells. Interestingly, even though miR-17 enhanced HeV infection, cells loaded with miR-17 agonists did not fuse at significantly higher efficiencies than negative control cells. Of note, smaller syncytia were formed in this experiment as less target cells were added to the co-cultures. The extent of syncytia formation in the captured images of all microscopy fields were additionally quantified using automated image analysis software (Fig 6D).
In sum, these results indicate that HeV-glycoprotein mediated cell-cell fusion is greatly stimulated by miR-181, but not by miR-17, suggesting that miR-181 specifically facilitates henipavirus infection by enhancing host entry and, quite possibly, by supporting cell-to-cell spread during late stages of infection via syncytia formation.
Since miR-181 specifically promotes infection of henipaviruses but not other paramyxoviruses, it is quite likely miR-181 increases membrane fusion by directly targeting viral and/or host molecules unique to the henipavirus fusion machinery. Although most miRNAs reduce expression of its target mRNA, there have been instances where miRNA binding improves stability and translation of the mRNA [30, 31]. Thus, we next investigated if miR-181 overexpression would enhance expression of the virus entry receptors ephrin-B2 and -B3, as well as the viral fusion glycoproteins F and G. miR-181a agonists were included in this analysis, subsequent to the validation of their pro-fusion nature in infection and fusion assays (S2 Fig). Perhaps surprisingly, S3 Fig shows that expression levels of the host and viral molecules which are known to be directly involved in fusion were not appreciably boosted by miR-181d. Similar results were observed for miR-181a as well. In fact, miR-181 downregulated ephrin-B3 and and HeV F glycoprotein by about 30 to 40%.
Given the substantial impact of miR-181 on cell-cell fusion (Fig 5B and 5C), it was intriguing that the miRNA did not considerably enhance expression levels of host and viral molecules known to be involved in entry and fusion. This led us to hypothesize that miR-181 supported fusion by down-regulating the expression of one or more novel cellular factor(s) that antagonizes expression and/or activity of the henipavirus entry receptors, ephrin-B2 or–B3. In the host, ephrins serve as native ligands for cellular Eph receptors (Ephs), which are high-affinity cell surface receptors belonging to the receptor tyrosine kinase family. Ephs and ephrins are further subdivided into classes A and B. Co-crystal structures of ephrin-B2 in complex either with its natural Eph [32, 33] or with the henipavirus G glycoproteins [34] have been solved. Intriguingly, these structures reveal the GH binding loop of ephrin-B2 to be the same dynamic region predominantly responsible for mediating the binding of ephrin-B2 to its natural Eph receptors as well as to the viral attachment proteins (S4A Fig.) [35]. Because the affinities of ephrin-B2-HeV-G and known ephrin-B2-Eph receptor interactions are within comparable nanomolar range [34, 36, 37], it implies that the Ephs may compete with the viral glycoprotein for binding to ephrin-B2. If true, one would expect Ephs to exhibit anti-viral properties, and any miRNAs which target these receptors for knockdown would be pro-viral.
Indeed, Bossart et al. previously reported that soluble forms of Ephs (EphB3 and B4) can compete with henipavirus G proteins for binding to ephrin-B2 or ephrin-B3, and can also inhibit virus infection [38]. Additionally, analysis of data from our recently published genome-wide siRNA screen [14] reveal that Ephs are more likely to be inhibitors of HeV infection (S4B Fig and S4 Table). That said, algorithmic analysis by TargetScan [39, 40] of all human Ephs does not predict miR-181 binding sites in the 3’ UTR of the mRNAs of EphB3 or EphB4 (S4 Table). Of the three Ephs which have putative binding sites, all belong to class A receptors. This includes the EphA5 receptor, which is the most anti-viral Eph receptor in our published RNAi screen (S4B Fig). Even though ephrin-B2 or ephrin-B3 tend to preferentially engage with class B Eph receptors (e.g. EphB3 and B4), there is some precedence for crosstalk interaction with class A Eph receptors as well, such as EphA4 [41] and EphA3 [42] (S4A Fig). Thus, a potential model for the pro-viral mechanism of miR-181 posits that the miRNA down-regulates expression of Ephs, increasing the pool of unbound ephrin-B2 or ephrin-B3 for henipavirus G glycoproteins to attach and trigger entry/membrane fusion.
To test this model, we began by assessing the impact of silencing select Ephs on HeV infection. We opted to assess all Ephs with putative miR-181 target sites as predicted by TargetScan [39, 40], namely EphA4, A5 and A7 (S4 Table). Even though it was not predicted to contain any miR-181 binding sites, EphB4 was the most anti-viral hit of the class B receptors in the RNAi screen and was previously shown to compete with HeV G glycoprotein for ephrin-B2 binding [38], so it was incorporated into our study as well.
We first validated the siRNAs used to silence expression of the particular Ephs. Transfecting HeLa cells with siRNAs targeting EphA4, A5, A7 or B4 resulted in a >80% relative decrease in target mRNA expression (Fig 7A). Down-regulating EphA4, A5, A7 or B4 with these siRNA duplexes all caused significant increases in HeV infection (Fig 7B), with EphB4 exhibiting the greatest impact (~200% of siNEG). We next performed a cell-cell fusion assay to directly test the effects of silencing of these receptors on henipavirus fusion. Interestingly, even though EphA4 was inhibitory for virus infection, it did not seem to be repress syncytia formation (Fig 7C), suggesting that EphA4 blocks infection at a step post entry. On the other hand, EphA5, A7 and B4 all suppressed cell fusion, with relative trends comparable to that seen in the infection assay (Fig 7B). We therefore demonstrate, for the first time, that in addition to class B receptors (i.e. EphB3 and B4), class A Eph receptors can also inhibit henipavirus cell-cell fusion.
To evaluate whether these anti-fusion Ephs are suppression targets of miR-181, the mRNA levels of the Ephs in agonist-transfected cells were measured by qRT-PCR. Over-expressing miR-181a and miR-181d caused a significant reduction in EphA5 and A7 mRNA levels, with the latter showing the greatest reduction in gene expression (Fig 7D). In contrast, EphA4 levels were not impacted by miR-181d, and were only modestly (11%) affected by miR-181a, demonstrating some level of specificity in the regulation of Eph receptor expression by miR-181. Interestingly, although it was not computationally predicted to have any target site in its mRNA 3’ UTR, EphB4 levels were reduced by both miRNAs by about 40% (Fig 7D). These data show that endogenous levels of the henipavirus fusion regulators EphA5, A7 and B4 can all be significantly suppressed by miR-181 expression. Simultaneous suppression of all three negative fusion regulators would conceivably result in a cellular state with abundant unbound ephrin molecules, strongly favoring efficient activation of the viral fusion machinery. This provides a coherent mechanistic model for how miR-181 may expedite host entry and virus spread during infection.
Previous studies have reported that members of the miR-181 family are involved in different aspects of immune regulation [43–45]. In particular, miR-181a expression levels have been shown to correlate with pro-inflammatory signals (e.g. IL-1β, IL-6, and TNF-α) in blood tissues of humans with chronic inflammation, as well as blood of LPS-treated mice [46]. Additionally, miR-181 expression in human kidney tissues were found to be associated with increased transcription of genes of inflammation pathways [47]. As the biological relevance of miRNAs is linked to their prevalence [48], we considered examining changes in levels of miR-181 molecules in in vivo infection models for henipaviruses. We hypothesized that, in conjunction with the host pro-inflammatory response during early infection, miR-181 expression might be up-regulated in HeV-infected mammals, but perhaps the virus co-opts this up-regulation to support infection and viral spread in the host.
Ferrets have been established as an animal model for the study of several human respiratory viruses [49], including Hendra and Nipah viruses [50, 51]. Accordingly, sixteen adult ferrets were exposed to HeV via the oronasal route, monitored for clinical signs and play activity, and two or four ferrets were euthanized at 1, 2, 3, 5, 6 and 7 day post-exposure (d.p.e.). After an incubation period of about five days, some ferrets started exhibiting weight loss, which correlated with an increase in rectal temperatures (S5A Fig and S5B Fig). By seven d.p.e., visible clinical signs, including depression, lack of grooming, or a decrease in playfulness, were observed in three out of four remaining ferrets. Establishment of HeV infection was confirmed by performing qRT-PCR for HeV genomic RNA on ten different tissues from the ferrets (Fig 8A) as well as by virus isolations (S5C Fig). By one d.p.e., viral RNA was detected in the retropharyngeal lymph nodes and in lung tissues, and by three d.p.e., HeV RNA was also recovered from the spleen, thymus and brain, suggesting neurotrophic spread characteristic of henipavirus disease. The higher viral RNA loads observed in the lung, lymph nodes and spleen is congruent with previous HeV infection studies performed in ferrets [50]. We next purified total small RNA from serum samples (from all days except for day six), and then performed qRT-PCR using a primer specific for miR-181a and miR-181d. We found that, as early as one d.p.e., miR-181a and miR-181d became elevated in the serum of these ferrets during the course of infection (Fig 8B). Reminiscent of what was observed in mice treated with LPS [46], this early up-regulation appears to be transient, as by day three, levels of miR-181a and miR-181d began dropping to baseline levels.
Natural HeV outbreaks occur in horses, causing severe febrile illness associated with respiratory and neurologic signs. Accordingly, horses serve as another animal model for respiratory and neurologic HeV disease. A HeV infection study involving three experimental horses was previously described [52]. Here, total small RNA was purified from the stored blood samples of these horses, and the relative expression of miR-181a and miR-181d on 0, 1, 3, 5, 7 and 9 d.p.e. were determined by qRT-PCR (Fig 8C). Similar to results observed in ferret, transient yet significant increases in circulating miR-181 molecules were observed during the early stages of infection. Collectively, these observations demonstrate in two different in vivo models that members of the miR-181 family are up-regulated early in the host during HeV infection, implicating a biological role for miR-181 in host immunity as well as in henipavirus pathogenesis.
The development of novel therapeutics for viruses of clinical significance relies on our knowledge of the dynamic interplay between the virus and the human host, and our ability to apply such knowledge to disrupt the viral dependence on host factors. However, progress in our understanding of virus-host interactions of many deadly viruses of significant public health importance (e.g. Ebola, MERS, Nipah virus) is hampered by the high-cost and technical challenges associated with studying these viruses under BSL-3 or BSL-4 conditions. To circumvent these issues, different strategies and approaches have been developed, such as the use of pseudotyped particles [53] or minigenome assays [54]. These approaches, though of much utility, have their shortcomings; in particular, they cannot fully reproduce the entire life cycle of the virus. Therefore, significant progress still needs to be made towards the development and validation of our capabilities to perform technically-challenging experiments in high biocontainment environments.
In recent years functional genomics has become a popular research approach for unbiased discoveries of novel genes and molecular pathways involved in a particular biological process. For infectious diseases, functional genomics has demonstrated much power in its ability to dissect the dynamic interplay between host and viral factors during a virus infection, paving the way for novel drug targets. For instance, a haploid genetic screen resulted in the discovery of the once elusive entry receptor for Ebola virus [55]. That said, methods used in functional genomics, such as high-throughput RNAi screens, are technically challenging and laborious, especially at BSL-4.
In this report, we screened 834 host miRNAs, using both engineered agonists and antagonists, for their ability to enhance or inhibit infection of HeV in human cells. As two complementary screens were performed, we exploited this duality and cross-referenced the two screens to increase our confidence in the top hits. Both complementary screens converged on members of four miRNA families (miR-181, miR-17~93, miR-520h, miR-548d) that strongly promoted henipavirus infection. Since all four members of the miR-181 family were pro-viral hits using this approach, we focused our validation efforts on miR-181. We show that miR-181 promoted infection of both wild-type HeV and NiV infections. Interestingly, this infection enhancement seems to be primarily mediated via the ability of miR-181 to significantly augment henipavirus glycoprotein-mediated cell-cell fusion, implicating miR-181 in the enhancement of henipavirus entry. Congruent with this notion, viral RNA synthesis in a single round of infection is elevated in cells transfected with miR-181 agonists. This pro-fusion effect is specific to the miR-181 family, as transfection with agonists of another strongly pro-viral miRNA (miR-17), did not appreciably alter syncytia formation. Since henipavirus mediated cell-cell fusion is both a surrogate model for virus entry as well as a natural phenomenon during late stages of infection, it is likely that in addition to enhancing henipavirus entry, miR-181 also promotes more efficient cell-to-cell spread of this virus by merging the cytosols of neighbouring cells more rapidly. To our knowledge, this is the first instance of subversion of a host miRNA by a virus to promote entry and membrane fusion.
Investigation into the pro-fusion mechanism of miR-181 led us to hypothesize that Eph receptors, the cellular binding partners of the henipavirus entry mediators ephrin-B2 and -B3, may act as potent anti-fusion regulators. Eph receptors and their ephrin ligands are involved in intracellular signalling via direct cell-cell contact and receptor-ligand interactions [56]. These molecules are divided into two different classes (A and B), and Ephs of a particular class tend to bind to ephrins of the corresponding class [57]. Exceptions to the rule exist, such as the ability of EphB2 to bind to ephrin-A5 [58]. Initial in silico analysis revealed that EphA4, EphA5 and EphA7 possess putative miR-181 binding sites in the 3’ UTRs of their transcripts. Our experiments subsequently showed that EphA5 and EphA7, but not EphA4, are novel suppressors of the fusiogenic effects of henipavirus glycoproteins. Importantly, levels of EphA5 and EphA7, but not EphA4, are reduced by overexpression of both miR-181a and miR-181d, indicating that these class A Ephs are target genes for the miR-181 family, and that the pro-fusion phenotype of miR-181 are, at least in part, due to its downregulation of specific class A Ephs.
We also demonstrated, using an approach different from a previous study [38], that EphB4, an Eph receptor from the same class as the henipavirus entry receptors, has potent anti-fusion characteristics. Indeed, at comparable knockdown levels, siRNAs to EphB4 increased syncytia formation significantly more than siRNAs to EphA5 or EphA7 (Fig 7A and 7C). This data also lends support to the notion originally proposed by Bossart et al. (2008) that class B Ephs can compete with henipavirus glycoproteins for binding to entry receptors to hamper virus entry. Intriguingly, even though the 3’UTR of EphB4 transcripts does not contain any sequence that is complementary to the seed region of miR-181, EphB4 levels were downregulated by miR-181 expression. Though less commonly reported, miRNAs can modulate gene expression by binding to the coding region of mRNAs [59]. Accordingly, human EphB4 does contain a putative miR-181 binding site in its ORF, providing an avenue for miR-181 regulation of its expression.
Since EphA5 and EphA7 do not interact with class B ephrins [56], it is likely the effects of downregulated EphA5 and EphA7 on promoting henipavirus fusion is indirect. One possibility is that since some crosstalk already exists between the two classes of the Eph-ephrin interaction network (e.g. EphB2 with ephrin-A5 [58]) [57], downregulation of some molecules in the network may have broader, indirect effects on the availability of other molecules in the system, including the ephrins utilized by henipaviruses. For instance, repression of EphA5 and EphA7 expression will free up more ephrin-A molecules, including ephrin-A5. The increased level of unbound ephrin-A5 results in its sequestration of EphB2 and thereby make more ephrin-B2 and ephrin-B3 molecules available for binding by the henipavirus glycoproteins.
Alternatively, EphA5 and EphA7 may act directly on ephrin-B2 and B3 to regulate fusion, but via lateral cis interactions on the same cell. It was more recently shown that the canonical intra-class binding rule may only apply for trans interactions [42]. For example, via lateral cis interactions, ephrin-B2 can attenuate the trans ligand-binding capacity of EphA3 and its activation, even though ephrin-B2 does not bind to EphA3 by canonical trans interaction [42]. This cis inhibition of EphA3 by ephrin-B2 implies that cis interactions do not exhibit the same receptor-ligand selectivity as trans interactions, providing a possible non-canonical mechanism for EphA5 and EphA7 to modulate ephrin-B2 activity.
Collectively, our data supports a model where simultaneous inhibition of multiple anti-fusion Ephs from both receptor classes by miR-181 contributes to greatly enhanced membrane fusion and infection. Considering that EphB4 is most antagonistic towards fusion (Fig 7C), it likely makes the most significant contribution to the pro-fusion phenotype of miR-181. Though this is the first report of a host miRNA that promotes virus fusion by suppressing anti-fusion receptors, there is some precedence for host receptor-ligand interactions that negatively regulate virus entry. Natural host ligands (i.e. BTLA and LTα) of the herpes simplex virus entry receptor, HVEM, can inhibit binding of the virus glycoprotein gD to HVEM and suppress infection [60]. Additionally, nectin members can compete with the measles virus hemagglutinin glycoprotein for binding to its exit receptor nectin-4 [61, 62].
Even though the role of miR-181 in inflammation and NKT-cell maturation has been documented [23, 43–46], little has been reported about its role in the infection of other viruses. In stark contrast to henipaviruses, miR-181 has been shown to be inhibit infection of porcine reproductive and respiratory syndrome virus [63]. Interestingly, this inhibition occurs at PRRSV host entry and is achieved by targeting the entry receptor CD163 for downregulation [64].
We found that miR-181 did not affect infection by paramyxoviruses from other genera, indicating specificity in the henipavirus-miR181 virus-host interaction. This supports the model that miR-181 enhances syncytia formation by targeting Ephs that naturally associate with the henipavirus entry receptors ephrin-B2 and B3. On the other hand, miR-17 enhances the infection of HeV as well as RSV, suggesting that the pro-viral effects of miR-17 are broadly applied to the paramyxovirus family, and perhaps beyond this family. For instance, miR-17 has recently been shown to be critical for the replication of pestiviruses, primarily via enhancing viral translation and vRNA stability [12]. Cross-referencing of results from the siRNA screen of host genes associated with HeV infection suggests that miR-181 and miR-17~93 target multiple host genes which are anti-viral for HeV, and that the net outcome of cellular expression of miR-181 or miR-17~93 is likely a host microenvironment that is more conducive for henipavirus infection. These results indicate that, in addition to its role in regulating fusion, miR-181 might act via other anti-viral host mediators to induce a situation that is broadly supportive of henipavirus replication. Consistent with this, we observed that miR-181 is up-regulated in sera of ferrets and blood of horses as early as day 1 during a henipavirus infection. It is tempting to speculate that the host pro-inflammatory response (of which blood miR-181 is correlated with) promotes the early phase of virus spread in the host, thereby contributing to disease progression and pathogenesis [43, 44, 46]. Along similar lines, but in a chronic infection, serum miR-181b is positively correlated with hepatitis B virus (HBV) DNA levels in human patients, and with disease progression of chronic HBV infection [65]. The model of miR-181-mediated immune pathogenesis has potential implications for risk factors associated with susceptibility to henipavirus disease, as well as for the strategic design and development of novel immunotherapy for henipavirus infections.
Direct binding interactions between host miRNAs and viral genomes have been reported for certain RNA viruses like EEEV and pestivirus [11, 12]. These interactions often have functional relevance in terms of supporting viral replication. It was recently first demonstrated that the functional RNAi protein Argonaute preferentially associates with certain subgenomic segments of paramyxoviruses as well [12]. In particular, a higher abundance of Argonaute association with M and N segments of the hMPV RNA was observed relative to the L segment. Due to limited resolution in the AGO-CLIP analysis in that study, whether these physical interactions involve specific host miRNAs remains to be addressed. Considering that our complementary screens identified multiple miRNAs that support henipavirus infection, an interesting subject for future studies would be to investigate whether the pro-viral miRNAs which we identified here bind to the genomes of paramyxoviruses, and whether such interactions have functional roles in supporting infection.
In summary, these dual screens further the understanding of the role of host-derived small noncoding RNAs in the infection cycle of henipaviruses, and provide a miRNA-based resource for the study of viruses from the order mononegavirales, including members of both the filovirus and paramyxovirus families, which presents significant threats to human and animal health. This study implicates miR-181 and certain class A Eph receptors as critical modulators of henipavirus membrane fusion, and highlights how the natural innate immune response of the host can be exploited by a RNA virus to promote cell-to-cell spread.
HeLa cells (ATCC CCL-2) and African green monkey kidney epithelial Vero cells (ATCC CRL-81) were maintained in growth medium comprised of DMEM GlutaMAX supplemented with 10% (v/v) fetal calf serum and 100 U/mL penicillin/streptomycin. Madin-Darby Canine Kidney (MDCK) cells (ATCC CCL-34) and HEK 293T cells (ATCC CRL-3216) were maintained in growth medium comprised of RPMI 1640 medium supplemented with 10% (v/v) fetal calf serum, 10 mM HEPES, 2 mM L-glutamine, 1 mM MEM sodium pyruvate and 100 U/mL penicillin/streptomycin. All cells were incubated at 37°C in a humidified atmosphere containing 5% CO2.
All virology work was conducted at the CSIRO Australian Animal Health Laboratory. Recombinant HeV, wild-type HeV (both Hendra virus/horse/1994/Hendra), NiV (Nipah virus/Malaysia/human/99), MeV (wild type Edmonston strain), MuV (Enders strain) and RSV (strain A2) were passaged in Vero cells. Influenza A/WSN/33 (H1N1) (kind gift, Professor Lorena Brown, University of Melbourne) was passaged in the allantoic fluid of 10-day embryonated specific pathogen-free chicken eggs. HeV and NiV were handled under BSL-4 conditions, MeV and MuV at BSL-3, and RSV and influenza A/WSN/33 at BSL-2. All viruses were aliquoted and stored at −80°C for inoculations.
High throughput miRNA agonist and antagonist screens were performed largely as described [14]. Briefly, HeLa cells were transfected in 384 well plates with miRIDIAN miRNA agonist and antagonist (final concentration 25 nM) libraries using DharmaFECT (DF) 1 lipid transfection reagent (Dharmacon RNAi Technologies, GE, Lafayette, Colorado, USA). Genome-wide miRNA libraries (catalog numbers in S1 Table and S2 Table) were screened at the Victorian Centre for Functional Genomics (VCFG). At 72 h post transfection, in parallel with the point of HeV infection, cell viability for each well was assessed by fixing (4% paraformaldehyde for 10 min) and staining plates with the nuclear stain 4',6-Diamidino-2-Phenylindole, Dihydrochloride (DAPI) (Invitrogen, Carlsbad, CA; 1 μg/ml for 20 min in phosphate buffered saline). HeV infection was quantitated in separate plates suitable for luminescence assays. 72 h post-transfection, cells were infected with recombinant HeV (MOI 0.1 using a BioTek 406 liquid handler housed in a class II biosafety cabinet at BSL-4. At 24 hours post-infection, media was removed and 20 μL of PBS added per well. Luminescence was then measured by addition of 20 μL of Bright-Glo Luciferase reagent (Promega, Madison, WI) and reading on a Synergy H4 multimode microplate reader (BioTek).
Data analysis was performed as described [14]. The experimental robustness was evaluated for each screened plate using the Z’ factor calculation [20], comparing the negative control (siNEG), positive control (siLUC) and death control (siPLK1) for both cell viability and HeV infection. Robust z score = (sample value-sample median)/sample median absolute deviation was used as the hit identification method [20, 66].
For work subsequent to the miRNA screens, HeLa cells were seeded overnight in 96-well plates (8 × 103 cells/well) in growth medium. The following day, growth medium was replaced with antibiotic-free medium (DMEM with GlutaMAX, 10% (v/v) foetal calf serum) (100 μL/well) before cells were transfected with miRNA agonists at a final concentration of 25 nM. Cells were then incubated at 37°C for 72 h (media was changed to growth media at 24 h-post transfection).
TCID50 assays were performed as described [22]. Infectious virus titre was then calculated in accordance with the method described by Reed and Muench [67].
MDCK cells were seeded in 6-well plates (1 x 106 cells/well) in growth media. The following day, cells were washed with PBS and media was changed in order to ensure the removal of any detached cells. When cell confluency reached 100%, 10-fold serial dilutions of influenza virus stocks in FCS-free media (RPMI 1640, 10 mM HEPES, 2 mM L-glutamine, 1 mM MEM sodium pyruvate, 100 U/mL penicillin/streptomycin) were prepared. Cells were then washed with FCS-free media, infected with 500 μL of the appropriate viral dilution and incubated at 37°C for 45 minutes, with gentle shaking every 15 minutes. 2 × L15 medium was then mixed with 1.8% (w/v) pre-autoclaved agarose and added to cells. Cells were incubated at 37°C for 3 days. Following incubation, cells were fixed with 5% (v/v) formaldehyde for 1 h. The agarose overlay was then removed and cells were stained with 0.1% (w/v) crystal violet, diluted in 4% (v/v) ethanol. After 10 minutes of staining, cell monolayers were rinsed with water and visible plaques were counted by eye.
Immunofluorescence microscopy was performed largely as described [14]. Cells were fixed with 4% (w/v) paraformaldehyde, stained with primary antibodies to viral antigens and analysed using an automated Thermo Fisher Scientific CellInsight Imaging System. HeLa cells in 96 well plates were imaged at a magnification of 10 x, 49 fields/well representing the entire well. The percentage of infected cells was quantified using the Target Activation bioapplication of the Cellomics Scan software and was determined by dividing the number of antigen-positive cells by the total cell number, multiplied by 100.
HeV-F and -G mediated fusion assays were performed as described previously [14].
Quantitative RT-PCR for HeV RNA was performed as described previously [14].
For the Eph-ephrin experiments, total RNA was purified from agonist or siRNA-transfected HeLa cells using the RNeasy Plus RNA purification kit from Qiagen (USA) and stored at −80°C. cDNA synthesis was performed using Superscript III reverse transcriptase kit (Invitrogen) according to the manufacturer's guidelines. qRT-PCR was performed using SYBR green (Invitrogen) on a StepOnePlus Real-Time PCR System (Applied Biosystems). PCR cycling for gene detection was at 95°C for 10 min, followed by 40 cycles of 95°C for 15 s and 60°C for 1 min. A melting curve analysis was performed to check for primer-dimer artifacts and to verify assay specificity. PCR primers were purchased from GeneWorks Ltd (Adelaide, Australia). Data were analyzed using the ΔΔCT method and were normalized to GAPDH for mRNA detection. qRT-PCR for HeV RNA during the infection time-course experiments were performed as described previously [14].
All animal studies were approved by the CSIRO Australian Animal Health Laboratory’s Animal Ethics Committee (document #1568) and conducted following the Australian National Health and Medical Research Council Code of Practice for the Care and Use of Animals for Scientific Purposes guidelines for housing and care of laboratory animals.
Sixteen ferrets (aged 12–18 months) were exposed to 5000 TCID50 of Hendra virus/Australia/Horse/2008/Redlands by the oronasal route as previously described [50]. Prior to any manipulations, animals were immobilised with a mixture of ketamine HCl (3 mg/kg) and medetomidine (30 μg/kg); atimepazole was administered for reversal at 50% of the medetomidine dose. After virus exposure, animals were monitored for changes in play activity, other clinical signs of disease, and fever. They were randomly assigned to euthanasia on post-exposure days 1, 2, 3, 5, 6 or 7, when clinical samples including nasal washes, mucosal swabs, blood and urine were collected together with multiple tissue specimens. qRT-PCR and viral loads from the tissues were assessed as described previously [50].
EDTA blood samples from horse infected with HeV were derived from an experimental time-series trial [52]. Ferret sera were collected as described above. Total RNA (including small RNAs) was harvested using Tri-reagent (Sigma, St. Louis, MO) following manufacturer's instructions. Phase separation was achieved by adding 200 μL chloroform to each tube, shaking vigorously and incubating samples for a further 3 min at room temperature, prior to centrifugation at 12,000 x g for 15 min at 4°C. The aqueous upper phase of each sample was then placed in a new tube, before the addition of 10 μg (0.5 μL) of glycogen and 250 μL of isopropanol. Samples were incubated for 10 min at room temperature prior to being centrifuged at 12,000 x g for 10 min at 4°C. The supernatants were then removed, and each RNA pellet was washed with 500 μL 75% (v/v) ethanol before being centrifuged at 7500 x g for 5 min at 4°C. RNA was then resuspended in 20 μL of RNase-free water.
Prior to cDNA synthesis, RNA samples (1 μg) were treated with RNase-free DNase according to manufacturer’s instructions (Promega). DNase-treated RNA was reverse transcribed into cDNA using the Exiqon miRCURY LNA Universal RT miRNA PCR kit, according to manufacturer’s instructions. qRT-PCR preparation was also performed using the Exiqon miRCURY LNA Universal RT miRNA PCR kit, which included miR-181 and U6primers. For each reaction, 5 μL PCR Master Mix and 1 μL PCR Primer Mix were added to 4 μL (5 ng) of the diluted cDNA template. qRT-PCR analysis was performed on the Applied Biosystems 7500 FAST Real-Time PCR System. Cycling conditions began with 10 min at 95°C, followed by 40 cycles of 95°C for 10 seconds and 60°C for 1 minute. Following qRT-PCR, miR-181 expression was analysed using the ΔΔCT method and normalised to U6.
All statistical analyses were performed using GraphPad Prism 5 software. The difference between treatment and control groups was analysed using a two-tailed Student’s t test, with a P value of <0.05 considered to be statistically significant. Error bars represent standard deviations, and all data points are the average of a minimum of 3 replicates.
|
10.1371/journal.pntd.0001936 | Clofazimine Modulates the Expression of Lipid Metabolism Proteins in Mycobacterium leprae-Infected Macrophages | Mycobacterium leprae (M. leprae) lives and replicates within macrophages in a foamy, lipid-laden phagosome. The lipids provide essential nutrition for the mycobacteria, and M. leprae infection modulates expression of important host proteins related to lipid metabolism. Thus, M. leprae infection increases the expression of adipophilin/adipose differentiation-related protein (ADRP) and decreases hormone-sensitive lipase (HSL), facilitating the accumulation and maintenance of lipid-rich environments suitable for the intracellular survival of M. leprae. HSL levels are not detectable in skin smear specimens taken from leprosy patients, but re-appear shortly after multidrug therapy (MDT). This study examined the effect of MDT components on host lipid metabolism in vitro, and the outcome of rifampicin, dapsone and clofazimine treatment on ADRP and HSL expression in THP-1 cells. Clofazimine attenuated the mRNA and protein levels of ADRP in M. leprae-infected cells, while those of HSL were increased. Rifampicin and dapsone did not show any significant effects on ADRP and HSL expression levels. A transient increase of interferon (IFN)-β and IFN-γ mRNA was also observed in cells infected with M. leprae and treated with clofazimine. Lipid droplets accumulated by M. leprae-infection were significantly decreased 48 h after clofazimine treatment. Such effects were not evident in cells without M. leprae infection. In clinical samples, ADRP expression was decreased and HSL expression was increased after treatment. These results suggest that clofazimine modulates lipid metabolism in M. leprae-infected macrophages by modulating the expression of ADRP and HSL. It also induces IFN production in M. leprae-infected cells. The resultant decrease in lipid accumulation, increase in lipolysis, and activation of innate immunity may be some of the key actions of clofazimine.
| Leprosy, caused by Mycobacterium leprae (M. leprae), is an ancient infectious disease that remains the leading infectious cause of disability. After infection, M. leprae lives inside host macrophages that contain a large amount of lipids, which is thought to be an essential microenvironment for M. leprae to survive in host cells. M. leprae infection increases lipid accumulation in macrophages and decreases the metabolic breakdown of lipids (catabolism). In addition, the treatment of leprosy with multidrug therapy (MDT) reverses the effect of infection on the modulation of lipid metabolism. We therefore aimed to use cultured human macrophage cells to determine which of the three MDT drugs (clofazimine, dapsone, or rifampicin) is responsible for this effect. We found that only clofazimine affects lipid accumulation and catabolism in M. leprae-infected cells in vitro. The amounts of lipids accumulated in the cells decreased when clofazimine was added to the cell culture medium. Clofazimine also activated immune responses in M. leprae-infected cells. These results suggest that the effectiveness of clofazimine against leprosy is due to the modulation of lipid metabolism and activation of immune reactions in M. leprae-infected host cells.
| Leprosy is a chronic infectious disease caused by Mycobacterium leprae (M. leprae), which is a typical intracellular pathogen that parasitizes tissue macrophages (histiocytes) and Schwann cells of the peripheral nerves of the dermis. Although its prevalence has declined over the last several decades due to the introduction of multi-drug therapy (MDT) by the World Health Organization (WHO), leprosy remains a major public health problem in many developing countries: In 2010, 228,474 new cases were registered worldwide [1]. Based on their clinical, histological and immunological manifestations, leprosy patients are classified into five groups that comprise one continuous spectrum: Tuberculoid (TT), Borderline Tuberculoid (BT), Borderline (BB), Borderline Lepromatous (BL) and Lepromatous (LL) [2]. LL is characterized by widespread skin lesions containing numerous bacilli that live in the foamy or enlarged lipid-filled phagosome within macrophages. Schwann cells in LL nerves also have the foamy, lipid-laden appearance that favors mycobacterial survival and persistence. In Schwann cells, M. leprae infection-induced biogenesis of lipid droplets correlates with increased prostaglandin E2 (PGE2) and interleukin-10 (IL-10) secretion, which is essential for leprosy pathogenesis [3], [4]. Although lipid-laden macrophages are also observed in other mycobacterial infections, including tuberculosis [5], [6], the amount of lipid and the number of infected macrophages are most prominent in cases of LL [7], [8].
The PAT protein family is named after three of its members: perilipin, adipophilin/adipose differentiation-related protein (ADRP), and tail-interacting protein of 47 kDa (TIP47). PAT family members are responsible for the transportation of lipids and the formation of lipid droplets in a variety of tissues and cultured cell lines, including adipocytes [9]–[12]. ADRP selectively increases the uptake of long chain fatty acids and has an essential role in fatty acid transport [13], [14]. Hormone-sensitive lipase (HSL), as the first enzyme identified in the induction of lipo-catabolic action initiated by hormones, is the predominant lipase effector of catecholamine-stimulated lipolysis in adipocytes [15]. Therefore, ADRP and HSL have opposing functions, i.e., lipid accumulation vs. its degradation. ADRP and HSL also play important roles in lipid accumulation in M. leprae-infected macrophages [8], [16]. M. leprae infection increased the expression of ADRP mRNA and protein, facilitating the accumulation and maintenance of a lipid-rich environment suitable for intracellular survival [8]. Conversely, HSL expression was suppressed in macrophages infected with M. leprae [16]. These results suggest that both ADRP and HSL influence the lipid-rich environment that favors M. leprae parasitization and survival in infected host cells. In our previous study, HSL expression was not detectable in slit-skin smear specimens from non-treated LL and BL patients, but it re-appeared shortly after MDT treatment [16]. However, how treatment modulates HSL expression is not clear. In the present study, we determine the effect of MDT components on host lipid metabolism by investigating the influence of rifampicin, dapsone and clofazimine on the expression of ADRP and HSL in THP-1 cells.
Human specimens were used according to the guidelines approved by the Ethical Committee of the National Institute of Infectious Diseases (Tokyo, Japan). All samples were anonymized before use.
Clofazimine (Sigma-Aldrich Co., St. Louis, MO), rifampicin (Wako Pure Chemical Industries Ltd., Osaka, Japan) and dapsone (Wako Pure Chemical Industries Ltd.) were dissolved in dimethyl sulfoxide (DMSO) and stored at 4°C. The final concentration used in the culture medium was 8.0 µg/ml rifampicin, 5.0 µg/ml dapsone or 2.0 µg/ml clofazimine.
Hypertensive nude rats (SHR/NCrj-rnu), infected with the Thai53 strain of M. leprae [17], [18] were kindly provided by Dr. Y. Yogi of the Leprosy Research Center, National Institute of Infectious Diseases. Japan. The protocol was approved by the Experimental Animal Committee, of the National Institute of Infectious Diseases, Tokyo, Japan (Permit Number: 206055). Animal studies were carried out in strict accordance with the recommendations from Japan's Animal Protection Law. M. leprae was isolated as previously described [19], [20]. The human premonocytic cell line THP-1 was obtained from the American Type Culture Collection (ATCC; Manassas, VA). The cells were cultured in six-well plates in RPMI medium supplemented with 10% charcoal-treated fetal bovine serum (FBS), 2% non-essential amino acids, 100 IU/ml penicillin and 100 µg/ml streptomycin at 37°C in 5% CO2 [7], [8]. Typically, 3×107 bacilli were added to 3×106 THP-1 cells (multiplicity of infection: MOI = 10).
Total RNA from cultured cells was prepared using RNeasy Mini Kits (Qiagen Inc., Valencia, CA) as described previously [7], [8]. Total RNA preparation from slit-skin smear samples was performed as described [8], [16]. Briefly, stainless steel blades (Feather Safety Razor Co., Osaka, Japan) used to obtain slit-skin smear specimens were rinsed in 1 ml of sterile 70% ethanol. The tube was then centrifuged at 20,000×g for 1 min at 4°C. After removing the supernatant, RNA was purified with the same protocol that was used for cultured cells. The RNA was eluted in 20 µl of elution buffer and treated with 0.1 U/µl DNase I (TaKaRa Bio, Kyoto, Japan) at 37°C for 60 min to degrade any contaminating genomic DNA. All RNA samples had an OD260/280 of 1.8–2.0 and an OD260/230 >1.8. RNA sample quality was also confirmed using denaturing agarose gel electrophoresis and the Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA) (Fig. S1). Total RNA from each sample was reverse-transcribed to cDNA using a High Capacity cDNA Reverse Transcription Kit (Applied Biosystems, Foster City, CA) with random primers [8], [16]. The following primers were used to amplify specific cDNAs: ADRP: 5′-TGTGGAGAAGACCAAGTCTGTG-3′ (forward) and 5′-GCTTCTGAACCAGATCAAATCC-3′ (reverse); HSL: 5′-CTCCTCATGGCTCAACTCCTTCC-3′ (forward) and 5′-AGGGGTTCTTGACTATGGGTG-3′ (reverse); interferon (IFN)-β: 5′-TGCTCTCCTGTTGTGCTTCTCCAC-3′ (forward) and 5′-CAATAGTCTCATTCCAGCCAGTGC-3′ (reverse); IFN-γ: 5′-GCAGAGCCAAATTGTCTCCTTTTAC-3′ (forward) and 5′-ATGCTCTTCGACCTCGAAACAGC-3′ (reverse) and actin: 5′-AGCCATGTACGTAGCCATCC-3′ (forward) and 5′-TGTGGTGGTGAAGCTGTAGC-3′ (reverse). Touchdown PCR was performed using a PCR Thermal Cycler DICE (TaKaRa Bio, Tokyo, Japan) [7], [8]. Briefly, the PCR mixture was first denatured for 5 min at 94°C, followed by 20 cycles of three-temperature PCR consisting of a 30-sec denaturation at 94°C, a 30-sec annealing that started at 65°C and decreased 0.5°C every cycle to 55°C, and a 45-sec extension at 72°C. An additional 10 cycles were performed for ADRP and β-actin, and 14 cycles for HSL with a fixed annealing temperature of 55°C. The products were analyzed by 2% agarose gel electrophoresis.
Cellular protein was extracted and analyzed as previously described [16], [21]. Briefly, cells were lysed in a lysis buffer containing 50 mM HEPES, 150 mM NaCl, 5 mM EDTA, 0.1% NP40, 20% glycerol, and protease inhibitor cocktail (Complete Mini, Roche, Indianapolis, IN) for 1 h. After centrifugation, the supernatant was transferred and 10 µg of protein was used for analysis. Cellular proteins were mixed with 4× LDS sample buffer and 10× reducing agent (Invitrogen, Life Technologies, Carlsbad, CA) and incubated for 10 min at 70°C prior to electrophoresis. Proteins were separated on NuPage 4–12% Bis Tris Gels and transferred using an iBlot Gel Transfer Device (Invitrogen). The membrane was washed with PBST (phosphate buffered saline (PBS) with 0.1% Tween 20), blocked in blocking buffer (PBST containing 5% skim milk) overnight, and then incubated with either rabbit anti-ADRP antibody (Santa Cruz Biotechnology Inc., Santa Cruz, CA; 1∶2,000 dilution), rabbit anti-HSL antibody (Cell Signaling Technology, Danvers, MA; 1∶1,000 dilution) or goat anti-β-actin antibody (Santa Cruz; dilution 1∶2,000). After washing with PBST, the membrane was incubated for 1 h with biotinylated donkey anti-rabbit antibody for ADRP and HSL (GE Healthcare, Fairfield, CT; 1∶2,000 dilution) or biotinylated donkey anti-goat antibody for β-actin (Millipore, Billerica, MA; dilution 1∶10,000) followed by streptavidin-HRP (GE Healthcare; 1∶10,000 dilution) for 1 h. The signal was developed using ECL Plus Reagent (GE Healthcare).
THP-1 cells were grown on glass coverslips in 24-well plates for 24 h, before the culture medium was exchanged with RPMI 1640 containing M. leprae and clofazimine. Control and drug-treated THP-1 cells were fixed in 10% formalin for 10 min. They were then washed with Dulbecco's PBS (DPBS) and balanced with 60% isopropanol for 1 min before staining with oil-red-O (Muto Pure Chemicals, Tokyo, Japan) for 10 min. The cells were counterstained with hematoxylin for 5 min followed by ethanol dehydration and coverslip sealing.
Archived formalin-fixed, paraffin-embedded tissue sections were subjected to immunohistochemical staining as described [7]. Briefly, deparaffinized sections were heated in 1 mM NaOH at 120°C for 5 min for antigen retrieval. They were then washed with PBST and blocked in blocking buffer (DAKO, Carpinteria, CA) for 10 min, and then incubated with either anti-ADRP antibody (Santa Cruz Biotechnology Inc.; 1∶200 dilution) or anti-HSL antibody (Cell Signaling Technology; 1∶100 dilution), for 1 h at room temperature. After washing the slides with PBST, peroxidase-labeled streptavidin-biotin method was employed using the LSAB2 kit (DAKO) and 3,3-diaminobenzidine tetrahydrochloride (DAB) for the staining of ADRP. Tyramide signal amplification (TSA)-HRP method was utilized to amplify HSL staining signals using the TSA Biotin System (PerkinElmer, Inc., Waltham, MA) according to the manufacturer's protocol. Sections were then stained using carbol fuchsin to visualize acid-fast mycobacteria and counterstained with hematoxylin.
All experiments were repeated at least three times. Since the replicates produced essentially the same outcomes, representative results from these independent experiments are shown in the figures.
The effect of MDT drugs on lipid metabolism in M. leprae-infected macrophages was examined by infecting human premonocytic THP-1 cells with M. leprae (MOI = 10) in the presence of 8.0 µg/ml rifampicin, 5.0 µg/ml dapsone or 2.0 µg/ml clofazimine for 24 h. Total RNA was isolated and RT-PCR analysis was performed to evaluate possible changes in ADRP and HSL mRNA levels. In our previous studies, M. leprae infection has been shown to increase ADRP and decrease HSL expression, which will in turn increase the lipid accumulation that is thought to contribute to maintaining a phagosome environment which permits M. leprae to parasitize tissue macrophages [8], [16]. However, when M. leprae-infected THP-1 cells were treated with clofazimine, ADRP expression levels decreased and HSL expression increased (Fig. 1). Rifampicin and dapsone did not show significant effects on the mRNA expression of ADRP, while they decreased HSL expression by augmenting the effect of M. leprae infection.
To further evaluate the effect of clofazimine on ADRP and HSL expression, THP-1 cells were treated with clofazimine in the presence or absence of M. leprae infection for 6, 12, 24 and 48 h. Total RNA and cellular protein were extracted and used for RT-PCR analysis and Western blot analysis, respectively. Linearity of the RT-PCR amplifications of ADRP, HSL and β-actin was confirmed by serial dilution of RNA samples and densitometric analysis of the bands (Fig. S2). RT-PCR showed that clofazimine alone had no effect on ADRP and HSL mRNA levels in control THP-1 cells (Fig. 2, left panel). Consistent with previous reports, ADRP mRNA expression was increased and HSL mRNA expression was decreased when THP-1 cells were infected with M. leprae (Fig. 2, middle panel) [8], [16]. However, simultaneous clofazimine treatment and M. leprae infection of THP-1 cells led to decreased ADRP and increased HSL mRNA levels (Fig. 2, right panel). The decrease of ADRP and increase of HSL mRNA expression were further confirmed by quantitative real-time PCR (Fig. S3), which also supports the linearity of our RT-PCR data. Thus, it was shown that clofazimine modulated expression of ADRP and HSL only in M. leprae-infected cells. Similar results were observed for ADRP and HSL protein expression levels in each experiment.
In the above studies, THP-1 cells were simultaneously treated with clofazimine and infected with M. leprae. Therefore, there was a possibility that clofazimine might have modulated the cellular environments of THP-1 cells before engulfing M. leprae. To eliminate this possibility and to imitate clinical situations, THP-1 cells were first infected with M. leprae for 24 h, to allow cells to engulf enough bacilli, before they were treated with clofazimine. M. leprae infection enhanced ADRP expression and suppressed HSL expression for up to 72 h (Fig. 3, left panel), which is consistent with the results shown in Fig. 2, middle panel. However, adding clofazimine 24 h after M. leprae infection produced lower levels of ADRP expression, but increased HSL expression (Fig. 3, right panel). Interestingly, ADRP expression fell even lower than the original level, and HSL rose higher than original levels, following clofazimine treatment. These results suggest that the lipid catabolic activity once suppressed by M. leprae infection was reactivated by clofazimine treatment, which in turn would promote lipolysis in infected macrophages and decrease cellular lipids. Also, these results are consistent with clinical situations in which HSL mRNA levels were recovered following successful treatment with MDT in LL and BL patients [16].
The decrease in ADRP expression and increase in HSL expression produced by clofazimine treatment were also observed when M. leprae-infected cells were further treated with peptidoglycan (PGN), a ligand for Toll-like receptor (TLR)-2, to activate innate immunity [8], [16]. We therefore hypothesized that clofazimine treatment might activate the innate immune response of THP-1 cells, which also confers bactericidal activities. To assess activation of innate immunity, production of interferon IFN-β and IFN-γ mRNA was evaluated in control and M. leprae-infected THP-1 cells treated with clofazimine. A transient increase of IFN-β and induction of IFN-γ were observed only in THP-1 cells infected with M. leprae and treated with clofazimine (Figs. 4A and 4B). Transient induction of IFNs as a result of macrophage activation is consistent with previous reports [22]–[24]. Innate immune activation of infected cells will further contribute to the elimination of intracellular bacilli, which is also consistent with the observation that the active form of vitamin D suppresses CORO1A expression in THP-1 cells [21].
To test whether the decrease in ADRP expression and increase in HSL expression after clofazimine treatment would result in less accumulation of cellular lipids after M. leprae infection, THP-1 cells were infected with M. leprae (MOI = 10) in the presence or absence of 2.0 µg/ml clofazimine for 48 h. Oil-red-O staining clearly demonstrated the accumulation of cellular lipid droplets following M. leprae infection (Fig. 5B vs. Fig. 5A). In M. leprae-infected cells treated with clofazimine, the amount of lipid droplets in the cell had significantly decreased by 48 h (Fig. 5C vs. 5B). The decrease in cellular lipid droplets is in agreement with the results shown in this study in which clofazimine decreased ADRP and increased HSL expression in M. leprae-infected cells.
To confirm the expression pattern of ADRP and HSL in clinical courses of leprosy, ADRP and HSL mRNA levels were evaluated in slit-skin smear specimens by RT-PCR analysis. ADRP mRNA was detected in all LL and most BL cases tested (Fig. 6A, right panel). HSL mRNA was detected in four BL cases; however, ADRP mRNA expression in these cases was absent or weaker than in other BL samples (Fig. 6A, cases 2, 4, 6 and 8). In one case, from which serial samples were obtained, the expression of ADRP mRNA decreased and HSL mRNA levels increased after treatment (Fig. 6B).
To further confirm changes in ADRP and HSL expression following treatment, immunohistochemical and acid-fast staining were performed using formalin-fixed paraffin-embedded skin tissue sections. Consistent with a previous report, ADRP localized to phagosome membranes that contains solid-shaped M. leprae (Fig. 7A) [8]. HSL staining was not evident before treatment (Fig. 7C). Three months after treatment, staining of the bacilli showed a dotted pattern with no solid-staining, indicating degeneration of M. leprae (Fig. 7 B and 7D). At this point, ADRP staining was faint (Fig. 7B), but strong HSL staining was observed along the phagosomal membrane (Fig. 7D). These staining patterns correlate with changes in mRNA levels of ADRP and HSL in the skin smears (Fig. 6B).
In previous studies, we showed that M. leprae infection increases ADRP expression and decreases HSL expression in host macrophages [8], [16]. The results of the present study demonstrate that clofazimine, one of the three major drugs used to treat leprosy, counteracts the effect of M. leprae to reduce ADRP and increase HSL expression of both mRNA and protein levels. These results are consistent with our observations in clinical samples obtained from leprosy patients, in which HSL levels were not detectable in skin smear specimens before treatment, but re-appeared shortly after MDT [8], [16]. The other two MDT drugs, dapsone and rifampicin, revealed no effects on the expression of either ADRP or HSL.
Mycobacteria survive by evading the host immune system and accessing host metabolic pathways to obtain nutrients for growth. M. leprae has undergone reductive evolution and pseudogenes now occupy half of its genome [25]–[27], thus M. leprae is thought to be the mycobacterium most dependent on host metabolic pathways, including host-derived lipids. As we previously reported, PGN can activate TLR2 to increase the expression of HSL [16] and suppress ADRP and perilipin expression [7], [8], [21]. These effects mediated by the TLR-initiated signaling pathway will induce lipid degradation, which makes it difficult for M. leprae to survive within host cells. M. leprae infection not only suppresses HSL expression, but also invalidates all effects of PGN on ADRP and perilipin, thus ensuring a phagosome environment that is favorable for mycobacterial survival [16]. In the present study clofazimine increased HSL expression and decreased ADRP expression only in M. leprae-infected cells. The amounts of lipids accumulated in the cells decreased when clofazimine was added to the cell culture medium. The decrease of the lipid-rich environment against the survival of M. leprae may be one of the key actions of clofazimine.
Clofazimine was the first clinically developed riminophenazine for the treatment of tuberculosis [28]. Its use has been extended to many Gram-positive bacterial infections as well as mycobacterial diseases [28]–[30]. The drug is now widely used for the treatment of leprosy, but its mechanism remains unclear [31]–[33]. The drug is extremely lipophilic and is also active in membrane destabilization and possible promotion of antigen processing. Stimulated phospholipase A2 activity and subsequent accumulation of arachidonic acid and lysophospholipids were confirmed in clofazimine-induced membrane destabilization [29], [34]. Increased major histocompatibility complex (MHC) class II expression in peripheral blood monocytes [35], up-regulated lysosomal enzyme activity of cultured macrophages [36] and decreased suppressor T-cell activity in mycobacteria-infected mice [37] reveal the potential role of clofazimine in facilitating immune recognition.
Although the underlying molecular mechanisms are not clear, clofazimine suppressed ADRP and induced HSL, IFN-β and IFN-γ expression only in cells infected with M. leprae, the same effects products by PGN [8], [16], [21]. Therefore, it is possible that clofazimine revives at least some of the activities of PGN, which is normally shielded by redundant mycolic acid at the M. leprae cell wall. Given the extreme lipophilicity of clofazimine and its activity against many Gram-positive bacteria, clofazimine may interact with the mycolic acid in the M. leprae cell wall that facilitates the exposure of PGN, which in turn activates TLR2-mediated signaling cascades, subsequently decreasing ADRP and increasing HSL [8], [16], [21]. Furthermore, since most lepra reactions, a cell-mediated, delayed-type hypersensitivity immune response, occur during or after MDT [38], [39], the prospect that clofazimine rescues shielded PGN activities, promoting lysosomal fusion and antigen processing, would be a plausible explanation for the trigger of lepra reactions.
The results from present and previous studies may explain the underlying mechanisms, at least in part, of successful parasitization of M. leprae and the effects of MDT treatment observed in patients. In conclusion, we have shown that clofazimine devastates the lipid-rich environment in M. leprae-infected host macrophages by modulating the expression of ADRP and HSL and activates the innate immune response of infected cells, both of which would be important in fighting mycobacterial infection.
|
10.1371/journal.pgen.1003226 | TDP2–Dependent Non-Homologous End-Joining Protects against Topoisomerase II–Induced DNA Breaks and Genome Instability in Cells and In Vivo | Anticancer topoisomerase “poisons” exploit the break-and-rejoining mechanism of topoisomerase II (TOP2) to generate TOP2-linked DNA double-strand breaks (DSBs). This characteristic underlies the clinical efficacy of TOP2 poisons, but is also implicated in chromosomal translocations and genome instability associated with secondary, treatment-related, haematological malignancy. Despite this relevance for cancer therapy, the mechanistic aspects governing repair of TOP2-induced DSBs and the physiological consequences that absent or aberrant repair can have are still poorly understood. To address these deficits, we employed cells and mice lacking tyrosyl DNA phosphodiesterase 2 (TDP2), an enzyme that hydrolyses 5′-phosphotyrosyl bonds at TOP2-associated DSBs, and studied their response to TOP2 poisons. Our results demonstrate that TDP2 functions in non-homologous end-joining (NHEJ) and liberates DSB termini that are competent for ligation. Moreover, we show that the absence of TDP2 in cells impairs not only the capacity to repair TOP2-induced DSBs but also the accuracy of the process, thus compromising genome integrity. Most importantly, we find this TDP2-dependent NHEJ mechanism to be physiologically relevant, as Tdp2-deleted mice are sensitive to TOP2-induced damage, displaying marked lymphoid toxicity, severe intestinal damage, and increased genome instability in the bone marrow. Collectively, our data reveal TDP2-mediated error-free NHEJ as an efficient and accurate mechanism to repair TOP2-induced DSBs. Given the widespread use of TOP2 poisons in cancer chemotherapy, this raises the possibility of TDP2 being an important etiological factor in the response of tumours to this type of agent and in the development of treatment-related malignancy.
| DNA double-strand breaks (DSBs) are dangerous because they can lead to cellular death and tissue degeneration if not repaired, or to genome rearrangements, which are a common hallmark of cancer, if repaired incorrectly. Although required for all chromosomal transitions in cells, transient DNA cleavage by topoisomerase II (TOP2) is a potential endogenous source of DSBs, which are characteristic in that TOP2 remains covalently bound to the DNA termini. In addition, numerous chemotherapeutic regimes rely on compounds that “poison” TOP2 activity, stimulating the formation of DSBs that target tumour cells. However, these compounds also affect healthy tissue and confer undesirable side effects, including the stimulation of genome rearrangements that can trigger secondary malignancies (mainly acute leukemia). Identifying the factors that participate in the repair of TOP2-induced DSBs and fully understanding their mechanism of action are therefore important for the design of chemotherapeutic regimes that are more effective and safer. Here we demonstrate that TDP2, a recently identified protein that can liberate DSB termini from blocked TOP2, functions as part of established cellular DSB repair processes and is required to safeguard genome integrity upon treatment with TOP2 poisons, both in cells and in mice. These results can therefore have important implications in cancer treatment.
| The double-stranded helical structure of DNA creates topological problems in all processes that involve opening of the double helix and accessing the genetic information [1], [2]. In particular, the transcription and duplication of DNA and its condensation into chromosomes generates knots and tangles that need to be resolved to avoid interference with diverse cellular processes and to ensure faithful chromosome segregation during mitosis. DNA topoisomerases are enzymes that introduce transient breaks in DNA to solve these topological problems. Type II topoisomerases, such as topoisomerase II in eukaryotes (TOP2) are essential homodimeric enzymes that relax, unknot and decatenate DNA molecules by catalyzing the passage of duplex DNA through a transient DNA double strand break (DSB) created by the enzyme [3]. Two isoforms of TOP2, α and β, exist in higher eukaryotes, with primary roles in replication and chromosome segregation and in transcription, respectively.
A key intermediate of TOP2 activity is the cleavage complex, in which each of two topoisomerase subunits is covalently linked to the 5′-terminus of an enzyme-generated DSB via a phosphodiester bond between the active-site tyrosine and the 5′-phosphate. The cleavage complex is normally a very short-lived intermediate, because the topoisomerase rapidly re-ligates the DSB once DNA strand passage through the DSB has occurred. However, under certain circumstances, such as the presence of nearby DNA lesions, cleavage complexes can be stabilized resulting in an increased likelihood of collision with RNA or DNA polymerases [4]. Such collisions can convert cleavage complexes into potentially clastogenic or lethal DSBs that require cellular DNA repair pathways for their removal.
Cleavage complexes are the target of a widely used class of anti-tumor agents that ‘poison’ topoisomerase activity, thereby prolonging the half-life of the intermediate and increasing the possibility of DSB formation [4], [5]. Thus, these drugs kill tumor cells by inducing high levels of TOP2-associated DSBs. Consequently, TOP2 poisons are commonly used antineoplastic drugs in the treatment of a broad range of tumor types including malignant lymphomas, sarcomas, leukemias, and lung, ovarian, breast and testicular cancers [5]. However, similar to other chemotherapeutic agents, TOP2-targeting drugs are only partially selective for tumour cells, resulting in unwanted toxicity in normal tissues and in therapy-associated chromosome translocations and secondary leukemias [6]–[14]. Moreover, some breakpoints in such translocations have actually been correlated with preferential sites of cleavage by TOP2 [13]–[17].
A characteristic feature of TOP2-induced DNA breaks is covalent attachment of the enzyme to 5′ ends of the DNA, which must be removed by cellular end-processing enzymes if DSB repair is to occur [18]. Until recently, the only known mechanism for removal TOP2 peptide from DNA 5′-termini in mammalian cells involved excision of the DNA fragment linked to the peptide using nucleases such as the MRN complex, CtIP or Artemis [19]–[21]. Recently, however, we identified a human 5′-tyrosyl DNA phosphodiesterase (5′-TDP) that can cleave 5′-phosphotyrosyl bonds and thereby release TOP2 from DSB termini without the need to also remove DNA sequence [22]. Consequently, this enzyme, which was previously known as signalling protein and transcription cofactor TTRAP/EAPII [23], [24], is now denoted tyrosyl DNA phopshodiesterase-2 (TDP2; Human Gene Nomeclature Organisation). Notably, consistent with its enzyme activity, TDP2 is required for cellular resistance to the anti-cancer TOP2 poison etoposide, but is not required for cellular resistance to ionizing radiation or methylmethane sulphonate [22], [25]; agents that induce DNA damage independently of TOP2 activity.
Following DNA end processing, DSBs can be repaired either by homologous recombination (HR) or by non-homologous end joining (NHEJ) [26]. However, these pathways utilize fundamentally different mechanisms for rejoining DSBs and consequently differ in their accuracy. In particular, HR utilizes undamaged sister chromatids to replace any nucleotides removed from DNA termini during DNA end processing and consequently is normally ‘error-free’. However, this process is available only during S phase or G2, when sister chromatids are available. In contrast, NHEJ is a ‘cut-and-splice’ process in which DSB termini are ligated together following DNA end processing without accurate replacement of missing nucleotides, and thus is potentially ‘error-prone’.
Here, we employ avian and murine experimental models to show that TDP2/Tdp2 deletion results in hypersensitivity to a structurally diverse range of anti-cancer TOP2 poisons. Moreover, we present genetic, biochemical and cellular evidence for TDP2 functioning in a mechanism of NHEJ that protects genome integrity in response to TOP2-induced damage. Finally, we show that this TDP2 dependent pathway also operates in vivo, as, upon exposure to TOP2 poisons, it is required for normal adult mouse lymphopoiesis, intestinal mucosa homeostasis and the maintenance of genome stability in the bone marrow. Collectively, our results suggest that TDP2 defines an error-free mechanism of NHEJ in mammals, which is specialized in the repair of TOP2-induced DSBs and reduces both tissue toxicity and genome instability in response to this particular type of DNA damage. These findings suggest the possibility of TDP2 being a significant etiological factor in the clinical tolerance and response to widely used TOP2 poisons.
The discovery of TDP2 as the first 5′-TDP activity raised the possibility of it being an important factor in the clinical response to TOP2 poisons [22], [25]. Indeed, TDP2 deleted avian DT40 cells are hypersensitive to etoposide [22], [25]. To address this question further, we examined the sensitivity of TDP2−/−/− cells to two additional, structurally diverse, TOP2 poisons. These drugs, denoted doxorubicin and amsacrine (m-AMSA), are employed widely during cancer therapy but in contrast to etoposide, ‘poison’ TOP2 by intercalating into DNA [5]. Nevertheless, similarly to etoposide, TDP2−/−/− cells displayed significant hypersensitivity to both doxorubicin and m-AMSA (Figure 1A). Moreover, a functional TDP2 phosphodiesterase domain was required for cellular resistance to this type of drug, because expression of wild-type human TDP2 (hTDP2) rescued the sensitivity of TDP2−/−/− DT40 cells to m-AMSA, whereas hTDP2D262A harbouring an inactivating mutation in the catalytic active site [5] did not (Figure 1A). These results show that TDP2 is required for cellular resistance to a range clinically relevant and structurally diverse TOP2 poisons, and support our contention that this requirement reflects the 5′-TDP activity of this enzyme.
To determine the impact of TDP2 on TOP2-induced DNA damage in mammals, and thus its possible relevance to anti-cancer therapy, we adopted a mouse model in which the first three exons of Tdp2, plus the 5′-UTR, were deleted by Cre-mediated excision (Figure 1B; see Materials and Methods). Mice homozygous for the deleted allele (Tdp2flΔ, from here-on denoted Tdp2Δ1–3) are viable, and so far we have not detected any abnormal pathology (unpublished observations). However, transformed Tdp2Δ1–3 mouse embryonic fibloblasts (MEFs) were hypersensitive to etoposide (Figure 1C, left, and Figure S1), but were not hypersensitive to DNA damage induced independently of TOP2 by γ-irradiation (Figure 1C, right).
Protein extracts from spleen, thymus, and bone marrow from wild type mice possess robust 5′-TDP activity, but, importantly, this activity was absent in analogous protein extracts from Tdp2Δ1–3 mice, confirming successful inactivation of the enzyme (Figure 2A). Cell extracts prepared from primary Tdp2Δ1–3 MEFs also lacked detectable 5′-TDP activity (Figure 2B). This was true not only for blunt-ended DSB substrates, but also for DSB substrates harbouring a 4-bp 5′-overhang (Figure 2C), characteristic of TOP2-induced DSBs. Additionally, EDTA-mediated chelation of Mg2+, which is essential for TDP2 function, completely eliminates 5′-TDP activity in wild type MEF extracts. These observations are significant because the related enzyme TDP1, whose activity is Mg2+ independent, was recently reported to possess weak activity on this type of substrate [27]. Our data therefore suggest that TDP2 is the primary, if not only, source of 5′-TDP activity in MEF extracts (Figure 2C).
Based on the mechanism of TOP2 cleavage, we anticipated that TDP2 activity would reconstitute ‘clean’ DSBs (5′ phosphate and 3′ hydroxyl termini) with 4-bp overhangs, which would be an ideal substrate for ligation by NHEJ. Interestingly, these ligation events would accurately preserve the DNA sequence, suggesting the possibility of an error-free NHEJ mechanism that specifically acts on TOP2-induced DSBs. To test this hypothesis, we examined whether TDP2 action at DSBs typical of those induced by TOP2 creates termini that can be ligated by T4 DNA ligase. Indeed, inclusion of T4 DNA ligase in reactions containing wild type MEF extract resulted in the additional appearance of a product of 46-nt, indicative of the completion of DSB repair by DNA ligation. However, this product was not observed if reactions contained cell extract from Tdp2Δ1–3 MEFs, confirming that DNA ligation was dependent on TDP2 activity (Figure 2D). Interestingly, the length of the product is consistent with a ligation event in which DNA sequence is preserved. To analyse ligation events directly catalysed by cell extracts, we generated linear plasmids harbouring 5′ phosphate or 5′ phosphotyrosine ends by PCR amplification with the corresponding modified primers. The incubation of these substrates with NHEJ-competent nuclear extracts [28] results in plasmid circularization events that can be scored as colonies following bacterial transformation. As can be seen in Figure 2E, nuclear extracts from Tdp2Δ1–3 MEFs efficiently circularized linear plasmids with 5′ phosphate ends but not linear plasmids harbouring 5′-phosphotyrosine. This difference was lost upon addition of recombinant TDP2 to the reaction, confirming the TDP2–dependent nature of the repair reaction. Collectively, our data suggest that TDP2 activity facilitates NHEJ of 5′ tyrosine-blocked ends by generating DSBs with ligatable termini, consistent with our hypothesis that this enzyme can support error-free NHEJ of TOP2-induced DNA damage.
To genetically test whether TDP2 functions indeed during NHEJ, we generated TDP2−/−/− DT40 cells harboring a targeted deletion of Ku70, a core component of the NHEJ pathway (Figure S2). Whilst both TDP2−/−/− and KU70−/− cells were hypersensitive to etoposide, cells in which both genes were deleted (TDP2−/−/−/KU70−/−) were no more hypersensitive than cells in which Ku70 alone was deleted (Figure 3A). In contrast to this epistatic relationship with a core NHEJ factor, transient knockdown of TDP2 further enhances etoposide sensitivity of HR defective (BRCA2 mutated) human fibroblasts (Figure 3B). Based on these genetic relationships, we conclude that TDP2 functions in a NHEJ-mediated and HR-independent pathway for the repair of TOP2-induced DSBs.
To further assign a role for TDP2 in the NHEJ pathway for DSB repair, we measured DSB repair rates in primary Tdp2Δ1–3 MEFs by immunodetection of γH2AX, a phosphorylated derivative of histone H2AX that arises at sites of chromosomal DSBs [29]. We measured DSB repair in specific phases of the cell cycle, because whilst NHEJ is operative throughout, HR-mediated DSB repair is operative only in S/G2 [30]. Notably, DSB repair rates were markedly reduced in Tdp2Δ1–3 MEFs following etoposide treatment, both in G0/G1 (Figure 3C) and G2 (Figure 3D), consistent with TDP2 functioning, as NHEJ, independently of cell cycle. These results were not specific to murine cells, since similar results were observed in TDP2-depleted human A549 cells (Figure S4). In contrast to treatment with etoposide, the rate of DSB repair was normal in Tdp2Δ1–3 MEFs following γ-irradiation, consistent with a role for TDP2 specifically at TOP2-induced DSBs (Figure 3E). Collectively, these data demonstrate that TDP2 is required in mammalian cells for rapid repair of TOP2-induced DSBs by NHEJ, and for cellular resistance to these lesions.
We hypothesized that this TDP2-mediated error-free NHEJ mechanism would be important to maintain genome integrity upon exposure to TOP2 poisons. To address this possibility, we quantified the frequency of micronuclei (MN), nucleoplasmic bridges (NB), and chromosomal aberrations following etoposide treatment. These events constitute well-established indicators of genome instability caused by misrepair of DSBs in which acentric, dicentric and aberrant chromosomes or chromosome fragments can be formed. As expected, etoposide increased the number of micronuclei and nucleoplasmic bridges in both transformed Tdp2+/+ and Tdp2Δ1–3 MEFs, but this increase was significantly higher (up to three-fold) in Tdp2Δ1–3 cells (Figure 4A). Primary Tdp2Δ1–3 MEFs at low passage (P3–4) similarly displayed elevated levels of micronuclei and nucleoplasmic bridges following etoposide treatment, compared to wild type primary MEFs (Figure 4B), although in the case of nucleoplasmic bridges the low number of cells displaying these structures prevented the difference from reaching statistical significance.
An additional indicator of genome instability is elevated frequencies of chromosome aberrations. Consequently, we quantified the frequency of chromosome breaks and exchanges in metaphase spreads of transformed Tdp2+/+ and Tdp2Δ1–3 MEFs. In agreement with the increased cell cycle arrest of TDP2−/−/− DT40 cells in G2 following etoposide treatment [25], we noted an etoposide-dependent reduction in metaphase cells that was particularly severe in Tdp2Δ1–3 MEFs (unpublished observations). However, of those metaphases observed and scored, both chromosome exchanges and breaks were significantly higher (2 to 5-fold) in Tdp2Δ1–3 MEFs than in Tdp2+/+ MEFs (Figure 4C). A similar increase in these events in Tdp2Δ1–3 MEFs, compared to wild type cells, was observed if low-passage primary MEFs were employed, ruling out the possibility that the elevated genome instability in Tdp2Δ1–3 MEFs was an artefact of cellular transformation (Figure 4D). In the latter case, etoposide treatment almost ablated the appearance of mitotic cells in populations of both wild type and Tdp2Δ1–3 MEFs, necessitating the use of caffeine to prevent G2 arrest. Taken together these results demonstrate that loss of TDP2 results in increased genome instability following TOP2-induced DNA strand breakage.
The above results demonstrate increased genome instability in Tdp2Δ1–3 MEFs, consistent with a role for TDP2 in error-free NHEJ-mediated repair of TOP2-induced DSBs. In this scenario, we considered the possibility that loss of TDP2 might also result in channelling of DSB repair towards HR. To address this question, we analyzed the formation of RAD51 foci, a well-established indicator of repair by HR. Following treatment with etoposide, the average number of Rad51 foci per cell was ∼3-fold higher in Tdp2Δ1–3 than in wild-type MEFs (Figure 5A), in agreement with an increase in the use of HR to repair TOP2-induced DSBs when TDP2 is not present. Furthermore, we compared the frequency of etoposide-induced sister chromatid exchanges (SCEs), a molecular hallmark of HR [31], in wild type and Tdp2Δ1–3 MEFs (Figure 5B). Notably, SCE levels increased substantially in transformed MEFs following acute etoposide exposure, being significantly higher in Tdp2Δ1–3 cells at two etoposide concentrations tested (1 and 2.5 µM). These data confirm that, upon etoposide treatment, the frequency of HR is elevated in Tdp2Δ1–3 MEFs, consistent with TDP2 functioning in NHEJ.
To address the relevance of TDP2-mediated repair of TOP2-induced DSBs in vivo, we compared the impact of etoposide on adult (8 wk) wild type and Tdp2Δ1–3 mice. A single intraperitonal injection of etoposide (75 mg/kg) caused a decrease in body weight in the initial 4 days post-treatment both in wild type and Tdp2Δ1–3 animals (Figure 6A). However, whereas Tdp2+/+ mice exhibited relatively mild and transient weight loss, Tdp2Δ1–3 littermates lost weight progressively and were sacrificed at day 6 to prevent suffering. No differences in body weight were observed between mock-treated (with DMSO) wild type and Tdp2Δ1–3 mice. Histopathological analysis of Tdp2Δ1–3 mice sacrificed 6 days after etoposide treatment revealed marked villous atrophy in the small intestinal mucosa as the likely cause of the drastic weight loss (Figure 6B). This was not observed in either wild-type and/or DMSO treated animals (data not shown), suggesting a protective role for TDP2 against adverse effects of etoposide in vivo.
In addition to severe intestinal damage, etoposide administration resulted in elevated splenic and thymic atrophy in Tdp2Δ1–3 mice, compared to wild type mice (Figure 6C), consistent with the known hypersensitivity of these organs to this drug [32]. Histological analysis of these tissues revealed a marked reduction in the cellular content in Tdp2Δ1–3 animals (Figure 6C, right, note the low density of dark-stained nuclei). In light of these results, we analysed B-cell and T-cell maturation in wild type and Tdp2Δ1–3 mice (Figure 6D and Figure S5). In the case of B-cell precursors in bone marrow, treatment with etoposide resulted in a decrease of 30–50% in the fraction of cells that were CD43+ B220+ progenitors (Pro-B cells) and a decrease of >95% in the fraction of cells that were CD43− B220low (Pre-B cells) or CD43− B220high (immature B cells) precursors. In all cases the reduction in B-cell precursors was greater in Tdp2Δ1–3 mice, but the differences were not statistically significant at the administered dose. In contrast, in the case of T-cell maturation, whereas etoposide treatment reduced the fraction of CD4+ CD8+ immature T cells by 30–40% in wild type mice, these cells were almost completely eliminated in Tdp2Δ1–3 mice (Figure 6A, bottom right). No effect was observed in CD11b/Mac-1+ myeloid cells in the bone marrow (Figure S6). Taken together, these results suggest that loss of TDP2 increases cellular attrition in the lymphoid system, particularly in the T-cell lineage, in response to TOP2-induced DNA damage.
A major side-effect of cancer therapy employing TOP2 poisons is secondary hematological malignancy, and in particular acute leukemia, resulting most likely from error prone/erroneous repair of TOP2-induced DSBs and chromosome translocations [4], [7]. Given our findings that TDP2 limits genome rearrangements induced by etoposide in cells, we examined whether TDP2 also promotes genome stability in bone marrow in vivo. We quantified the fraction of micronucleated polychromatic erythrocytes (PCEs) in bone marrow smears from Tdp2Δ1–3 and Tdp2+/+ mice 24 hour after intraperitoneal injection of etoposide (1 mg/kg). The rodent erythrocyte micronucleus test is a standard procedure to detect cytogenetic damage in toxicological studies and is based on the detection of micronuclei in erythrocyte precursors (Hayashi et al 1994). As expected, etoposide increased the fraction of PCEs that were micronucleated in both wild type and Tdp2Δ1–3 animals (Figure 7). However, this increase was ∼2-fold higher in Tdp2Δ1–3 mice than in wild type mice, suggesting that TDP2 protects heamatopoietic cells from genome instability induced by anti-cancer TOP2 poisons.
In the current study we observe that Tdp2 deletion ablates detectable 5′-TDP activity in different mouse tissues and MEFs, consistent with our previous observations in DT40 cells [25]. It is worth noting that other roles have been assigned to this protein, in other cellular processes such as signal transduction and transcriptional regulation [33]. So far, however, we have been unable to detect any spontaneous phenotype caused by TDP2 loss, either at cellular level or in vivo, while dramatic effects are observed upon etoposide treatment. This suggests that the most important function of TDP2, following Top2 induced DNA damage at least, is related to the 5′ TDP activity of this enzyme. Additionally, our data suggest that alternative, TDP2–independent, mechanisms of DSB repair are sufficient to cope with the endogenous level of TOP2 damage arising during normal mouse development and life.
A role for human TDP1 in repairing TOP2-induced DSBs was recently suggested by a weak 5′-TDP activity of human recombinant protein on DSBs possessing 4-bp 5′-overhangs, and on a mild sensitivity of TDP1−/− DT40 cells to etoposide [27]. This is also consistent with the increased resistance to etoposide reported in cells highly overexpressing TDP1 [34], and with the reported 5′-TDP activity of Tdp1 in Saccharomyces cerevisiae [35]. However, while our standard activity assays employs DSBs with blunt-ended 5′-phosphotyrosyl termini, in the current study we similarly failed to detect residual 5′-TDP activity in Tdp2Δ1–3 MEF extracts on DSB substrates with 4-bp 5′-overhangs (Figure 2C). In addition, in our hands, TDP1−/− DT40 cells are not hypersensitive to etoposide, and deletion of TDP1 in TDP2−/−/− DT40 cells does not increase sensitivity to etoposide above that observed by TDP2 deletion alone [36]. Consequently, we conclude that TDP2 is the major if not only 5′-TDP activity in mammals (as in DT40 chicken cells), at physiologically relevant enzyme concentrations at least.
We have shown that Tdp2-deleted mouse cells are hypersensitive to TOP2-induced DNA damage, but not to ionizing radiation, in agreement with previous results with TDP2−/−/− DT40 cells [25]. Moreover, we demonstrate that this hypersensitivity correlates with a defect in the repair of etoposide-induced DSBs, as measured by immunostaining for sites of γH2AX, which suggests that TDP2-mediated repair promotes tolerance to TOP2-induced DNA damage in mammalian cells. Remarkably, we observed that TDP2 is required for resistance to TOP2-induced DNA damage not only at the cellular level, but also at the whole-organism level. Indeed, etoposide administration in Tdp2Δ1–3 mice resulted in both increased mortality due to intestinal damage and in elevated toxicity in lymphoid tissue, established in vivo targets of etoposide [32]. TDP2 is therefore a critical factor in the cellular and physiological response to TOP2 poisons.
One important result of our study was to uncover the relationship between TDP2 and the major DSB-repair pathways, NHEJ and HR. We have shown that TDP2 can convert DSBs with 5′-phosphotyrosyl termini into DSBs that are directly ligatable, and might thus be of particular utility in facilitating an error-free NHEJ pathway for repair of TOP2-induced DSBs. Several of our observations support the idea that TDP2 is a component of NHEJ. First, the contribution of TDP2 to cellular resistance to TOP2 induced DNA damage is dependent on the NHEJ machinery and independent on HR, as, with regards to etoposide sensitivity, KU70 is epistatic over TDP2 deletion in DT40 cells while an additive effect is observed when TDP2 is depleted in BRCA2-deficient human fibroblasts. Second, loss of TDP2 results in a DSB repair defect not only in G2 but also in G0/G1, cell cycle stages in which NHEJ is the main if not only DSB repair mechanism available [26], [30], [37], [38]. Third, Tdp2Δ1–3 MEFs exhibit increased levels of HR-mediated DSB repair, as measured by elevated frequencies of RAD51 foci and sister chromatid exchange in response to etoposide treatment, which is a phenotype observed in other cell lines in which NHEJ is defective [39]–[41]. Additionally, we have been unable to generate DT40 cells in which both TDP2 and XRCC3 are deleted, suggesting that loss of both TDP2 and HR-mediated DSB repair is cell lethal (unpublished observations).
Whilst the above observations argue strongly that TDP2 is a component of NHEJ, it is important to note that TDP-independent NHEJ mechanisms to process TOP2-linked termini most likely also exist and employ nucleases such as MRN complex, CtIP or Artemis [5], [18]–[21]. This explains why KU70−/− DT40 cells exhibit much greater hypersensitivity to etoposide than TDP2−/−/− DT40 cells, and why Tdp2Δ1–3 MEFs still repair a significant fraction of etoposide-induced DSBs in G0/G1 (when NHEJ is the only DSB repair pathway available). Whilst nuclease-mediated NHEJ can support cell survival in response to TOP2-induced DNA damage, they most likely do so at the expense of increased genetic instability. This is because the removal of sequence from 4-bp complementary 5′-overhang during NHEJ will, on the one hand, likely result in chromosome deletions, and on the other hand, increase the propensity for DSB misjoining and chromosome translocation. In contrast, HR provides an error-free pathway to repair TOP2-induced DSBs that have been processed by nucleases, by restoring any missing DNA sequence from and intact sister chromatid in S and G2 [30], [37], [42]. In this scenario, the increased etoposide-induced genome instability in Tdp2Δ1–3 mice, both in cultured cells from these animals and in bone marrow in vivo, likely reflects the use of TDP2–independent NHEJ in cellular contexts in which HR-mediated DSB repair is unavailable (e.g. in cells in G0/G1), or is saturated by the number of etoposide-induced DSBs.
In summary, based on these and our previously published data, we suggest that TDP2 defines a novel error-free NHEJ sub-pathway that converts TOP2-linked 5′-termini into ligatable DNA termini. We suggest that this may be particularly important during G1 and in post-mitotic cells, which lack HR-mediated repair, and thus in which it may be the only mechanism for error-free DSB repair of TOP2-induced DSBs (Figure 8).
The results presented here can have important implications in the treatment of cancer. Given the widespread use of TOP2 poisons in cancer therapy, and the observed hypersensitivity to TOP2 poisons of cells lacking TDP2, our findings suggest that TDP2 could affect the response of tumour cells to chemotherapy. In this context, TDP2 expression is reportedly elevated in the majority of non-small cell lung cancer cells [43], and mutant-p53-dependent over-expression of TDP2 has been implicated in cellular resistance to etoposide in lung cancer cells [44]. TDP2 might therefore be a valid target for overcoming tumour resistance to TOP2 poisons and/or a useful predictive biomarker for clinical response to these agents.
In addition, our toxicity assays in mice and the increased genome instability in cells and in mouse bone marrow correlate well with known side effects of treatment with TOP2 poisons during cancer therapy. This raises the possibility that heterogeneity in expression levels or activity of TDP2 could be an important etiological factor both in the toxicity that accompanies chemotherapy involving TOP2 poisons [45] and on the incidence of treatment-related hematological malignancy, typically acute leukemia occurring in a relatively high proportion of patients [4], [7], [8]. Like other acute leukemias, therapy-related malignancies are linked to specific translocations that result in the expression of fusion proteins and contribute in some way to disease development. Intriguingly, in some cases, these translocations map to regions of preferential TOP2 cleavage, supporting a model in which the translocations arise via erroneous repair of TOP2-induced DSBs. These translocations are also surprisingly similar to those found in infant leukemia [46], suggesting that erroneous repair of TOP2-induced DSBs may also be a source of primary malignancy. Consistent with this idea, TOP2-induced DSBs are implicated in translocations commonly associated with prostate cancer [47]. In the light of our findings, it is tempting to speculate that TDP2 activity reduces the likelihood of oncogenic translocations, by ensuring rapid and accurate repair of TOP2-induced DSBs. It is possible, however, that TDP2 might occasionally promote a translocation, by liberating a DSB that engages in erroneous DNA ligation, as might be the case in some extremely conservative rearrangements that have been reported [12], [13].
We have shown that TDP2 protects mouse cells from the cytotoxic and clastogenic effects of TOP2 poisons, most likely by functioning in error-free pathway for NHEJ. These results have important implications in the treatment of cancer. For example, development of small molecule inhibitors for TDP2 may provide a way of sensitizing particular types of tumor to chemotherapy, though precaution is necessary to consider the possible consequences of TDP2 inhibition on normal cells and on the generation of secondary malignancies.
All animal procedures were performed in accordance with European Union legislation and with the approval of the Ethical Committee for Animal Experimentation of the University of Leuven and the University of Seville, respectively.
Chicken DT40 B lymphoma cells were cultured at 39°C, 5% CO2 in RPMI 1640 medium supplemented with 10−5 M β-mercaptoethanol, penicillin, streptomycin, 10% fetal calf serum (FCS), and 1% chicken serum (Sigma). TDP2−/−/− cell line was previously described [25]. To generate KU70 deletion constructs, Hygromycin (HygroR) or Neomycin (NeoR) resistance cassettes were inserted between sequences of 1.6 kb and 3.3 kb in length from the KU70 locus [48]. KU70-HygroR and KU70-NeoR deletion constructs were sequentially transfected into wild-type and TDP2−/−/− cells. The gene targeting events were confirmed by Southern blot analysis of EcoRI -digested genomic DNA hybridized to an external probe (Figure S1).
Transformed human fibroblast lines 1BR (wild-type) and HSC62 (BRCA2-mutant) were described previously [49]. Cells were cultured in DMEM supplemented with penicillin, streptomycin and 15% FCS.
Primary MEFs were isolated from littermate embryos at day 13 p.c. and cultured at 37°C, 5% CO2, 3% O2 in Dubelcco's Modified Eagle's Medium (DMEM) supplemented with penicillin, streptomycin, 10% FCS and non-essential aminoacids. All experiments were carried out between P2 and P4. MEFs were transformed by retroviral delivery of T121, a fragment of the SV40 large T antigen that antagonizes the three Rb family members but not p53 [50]. Transformed MEFs were maintained at 37°C, 5% CO2 in DMEM supplemented with penicillin, streptomycin and 10% FCS.
A targeting construct was generated for Tdp2 in which the first three exons were flanked by loxP sites, followed by an FRT- and loxP-flanked neomycin-resistance (neo) cassette. These three exons encode for the N-terminal half of TDP2 and contain mapped interaction domains for e.g. TDP2 itself, CD40 and TRAF6 [23]. The Tdp2flEx1–3,neo targeting construct was electroporated in E14 (129Ola) ES cells and correctly recombined ES cell clones were confirmed by Southern blot analysis. The functionality of the loxP sites was shown in vitro by electroporation of a correctly targeted ES cell clone with a Cre-expressing vector. Several correctly targeted ES cell clones were used for aggregation with CD1 morulae and transferred into pseudo-pregnant recipient females to obtain chimaeric mice. Three chimaeric males produced heterozygous offspring after breeding with CD1 wild-type females. The obtained offspring was genotyped with both a loxP-specific and a neo-specific PCR. Intercrosses between Tdp2flEx1–3,neo/+ mice led to the generation of homozygous floxed Tdp2 mice which were viable and fertile. To delete the critical exons we crossed the heterozygous Tdp2 mice with an EIIa-Cre mouse (Adenovirus EIIa-promoter driven Cre) and obtained Tdp2flΔ/+ mice. Intercrosses of the latter mice resulted in viable homozygous knockout mice (from now on denoted Tdp2Δ1–3) at the normal 25% Mendelian distribution. Southern blot analysis confirmed the complete recombination of the loxP-flanked sequences in the homozygous mice and hence the generation of Tdp2 knockout mice.
Labelled double-stranded 5′-phosphotyrosyl substrates were generated essentially as previously described [22], [22,25]. For 5′ overhang substrates 5′-Y-P-AATTCTTCTCTTTCCAGGGCTATGT-3′ (Midland Certified Reagents) and 5′-AGACATAGCCCTGGAAAGAGAAG-3′ (Sigma) oligonucleotides were annealed. Cell and tissue extracts were prepared by mild sonication in Lysis Buffer (40 mM Tris–HCl, pH 7.5, 100 mM NaCl, 0.1% Tween-20, 1 mM DTT) supplemented with 1 mM PMSF and protease inhibitor cocktail (Sigma) and clarification by centrifugation 10 min 16500 g 4°C. Protein concentration was measured with Bradford reagent (Sigma). 5′-TDP reactions contained 50 nM substrate, 80 µM competitor single-stranded oligonucleotide and the indicated amount of protein extract in a total volume of 6 µl Reaction Buffer (50 mM Tris-Cl, pH 7.5, 50 mM KCl, 1 mM MgCl2, 1 mM DTT, 100 µg/ml BSA). Ligation reactions with oligos contained in addition 5 units of T4 DNA ligase (Fermentas) and 1 mM ATP (Sigma) Reactions were stopped by the addition of 3 µl 3× Formamide Loading Buffer and 5 min 95°C incubation. Samples were resolved by denaturing polyacrylamide gel electrophoresis and analysed by phosphorimaging in a Fujifilm FLA5100 device (GE Healthcare).
Substrates were generated by PCR-mediated amplification of plasmid pEGFP-Pem1-Ad2 [51] with primers 5′-AATTCTTCTCTTTCCAGGGCTATGT-3′ and 5′- AATTCATCCCCAGAAATGTAACTTG-3′ harbouring phosphate (Sigma) or phosphotyrosine (Midland Certified Reagents) moieties at 5′ ends. NHEJ-competent nuclear extracts were prepared as previously described [28]. Reactions were performed by incubating (6 h at 16°C) 100 ng of each substrate with 7 µg of nuclear extracts in NHEJ Buffer (50 mM Tris-HCl pH 7.5, 50 mM KCl, 1 mM DTT, 2 mM MgCl2, 1 mM ATP, 100 µg/ml BSA) in the presence or absence of 50 nM hTDP2 recombinant protein (purified as previously described [22]). Reactions were stopped by addition of EDTA (to a final concentration of 100 mM) and treated 30′ with Proteinase K (0.2 mg/ml). DNA was purified using FavorPrep GEL/PCR Purification Mini Kit (Favorgen) and transformed into MegaX DH10B T1 Electrocompetent Cells (Invitrogen). Positive transformants were selected by plating on LB agar plates containing kanamycin (25 µg/ml).
To determine sensitivity in DT40, cells were plated in 5 ml of medium containing 1.5% (by weight) methylcellulose (Sigma) in 6-well plates at 50, 500, and 5000 cells/well per treatment condition. Media also contained the indicated concentration of doxorubicin (Sigma), mAMSA (Sigma) or etoposide (Sigma). In all experiments, cells were incubated for 7–11 days and visible colonies were counted.
Survival assays in MEFs were carried out seeding 2000 cells in 100 mm dishes, in duplicate for each experimental condition. After 6 hours, cells were irradiated or treated with the indicated concentrations of etoposide for 3 hours, washed with PBS and fresh medium was added. Cells were incubated for 10–14 days and fixed and stained for colony scoring in Crystal violet solution (0.5% Crystal violet in 20% ethanol). The surviving fraction at each dose was calculated by dividing the average number of visible colonies in treated versus untreated dishes.
Human fibroblasts were transfected with non-targeting Negative Control and TDP2 smartpool siRNAs (Thermo Scientific) using HyperFect transfection reagent (Invitrogen). Cells were transfected twice in two consecutive days and used for survival 48 hours after second transfection. Other details as described above.
MEFs grown on coverslips for the required time, 7 days for confluency-arrested cells and 2 days for cycling cells, were treated as indicated and fixed (10 min in PBS-4% paraformaldehyde), permeabilized (2 min in PBS-0.2% Triton X-100), blocked (30 min in PBS-5% BSA) and incubated with the required primary antibodies (1–3 h in PBS-1% BSA). Cells were then washed (3×5 min in PBS-0.1% Tween 20), incubated for 30 min with the corresponding AlexaFluor-conjugated secondary antibodies (1/1000 dilution in 1% BSA-PBS) and washed again as described above. Finally, they were counterstained with DAPI (Sigma) and mounted in Vectashield (Vector Labs). Rad51 foci scoring requires 30 sec. pre-extraction in PBS-0.1% Triton X-100 prior to fixation. γH2AX and Rad51 foci were manually counted (double-blind) in 40 cells from each experimental condition. When necessary to identify replicating cells, 5-ethynyl-2′-deoxyuridine (EdU, Invitrogen) was added throughout treatment and repair at a final concentration of 10 µM. Click chemistry reaction was performed before DAPI staining by incubating (30 min r.t.) with 1 µM AlexaFluor-conjugated azide (Invitrogen) in reaction cocktail (100 mM TrisHCl pH 8.5, 1 mM CuSO4, 100 mM ascorbic acid). For the analysis of G0/G1 confluency-arrested cells only Cyclin A negative cells were scored. For G2, as EdU was present (10 µM) during and after treatment, only Cyclin A positive cells without EdU incorporation were scored (see Figure S3). Primary antibodies were used at the indicated dilution: γH2AX (Millipore, 05-636) 1/1000, Cyclin A (Santa Cruz, sc-751) 1/500, Rad51 (Abcam, ab213) 1/200 and Tubulin (Santa Cruz, sc-5286) 1/2500.
Micronuclei and nucleoplasmic bridges were analysed in transformed and low passage primary MEFs previously seeded onto coverslips. Following treatment, cytochalasin B (Sigma) was added at 4 µg/ml to transformed but not to primary MEFs. 22 h (transformed) or 30 h (primary) post-treatment, cells were fixed and subject to DAPI staining as described above. In transformed cells only binucleated cells were scored, which was confirmed by visualization of the cytoplasm with anti Tubulin immunofluorescence (performed as described above).
Chromosomal aberrations were scored in Giemsa stained metaphase spreads. Following treatment, recovery in fresh medium was allowed for 2 h (transformed MEFs) or 4 h (primary MEFs) and demecolcine (Sigma) was added at a final concentration of 0.2 µg/ml. Caffeine (Sigma) was also added at a final concentration of 0.1 µg/ml but only to primary cells. 4 h later cells were collected by trypsinisation, subject to hypotonic shock for 1 hour at 37°C in 0.3 M sodium citrate and fixed in 3∶1 methanol∶acetic acid solution. Cells were dropped onto acetic acid humidified slides and stained 20 minutes in Giemsa-modified (Sigma) solution (5% v/v in H2O).
For SCEs 10 µM BrdU (Sigma) was added to the medium for two complete cycles (approximately 48 hours) before collection. Drug treatment was applied for 30 minutes 6–8 hours before cell collection. Metaphase spreads were obtained as described above. Before Giemsa staining, slides were incubated in Hoescht 33258 solution (10 µg/ml) for 20 minutes, exposed to UV light (355 nm) for 1 hour and washed for 1 hour at 60°C in 20× SCC.
The mouse colony was maintained in an outbred 129Ola, CD1 and C57BL/6 background under standard housing conditions, at 21±1°C with a photoperiod of 12∶12 h (lights on at 8:00). They were housed in isolated cages with controlled ventilation trough HEPA-filters and were in flow cabins. Sterile food pellets and water were available ad libitum. Breeding pairs between heterozygotes (Tdp2+/flΔ×Tdp2+/flΔ) were set to obtain wild-type (Tdp2+/+) and knock-out (Tdp2Δ1–3) littermates for analysis. Mice were genotyped with Phire Animal Tissue Direct PCR Kit (Thermo) following manufacturer instructions and using primers 5′-CCTTCATTACTTCTCGTAGGTTCTGGGTC-3′, 5′-ACCCGCTCTTCACGCTGCTTCC-3′ and 5′-TACACCGTGCCATAATGACCAAC-3′. This results in amplification of a 429 bp fragment from the wild-type allele or 561 bp fragment from the mutant allele.
At 8 weeks of age, Tdp2+/+ and Tdp2Δ1–3 mice underwent intraperitoneal injection with 3 µl/g of body weight of either DMSO (vehicle control) or etoposide at 25 mg/ml in DMSO for a final dose of 75 mg/kg. Weight and general health status was monitored daily from the day of injection (inclusive). 6 days post-treatment mice were sacrificed by cervical dislocation and dissected for histopathological analysis. Weight of spleen and thymus was recorded prior to their histological or cell content analysis. Bone marrow (BM) from femurs and tibias of each mouse was also obtained.
For histological analysis organs were fixed in 4% paraformaldehyde for 2 days, embedded in paraffin, cut in 6 µm slices by microtome, stained with Hematoxylin-Eosin and visualized under the microscope. For cell content analysis by FACS, BM and thymus were homogenized in EDTA Buffer (140 mM NaCL, 1.5 mM KH2PO4, 2.7 mM KCl, 8.1 mM Na2HPO4, 0.6 mM EDTA). Cells from both tissues were immunolabelled with the appropriate fluorescently-labelled antibodies according to manufacture's recommendations and analyzed using a FACScalibur flow cytometer (Becton Dickinson): B220-APC (17-0452), CD43 FITC (11-0431), CD8 APC (17-0081) and CD4 FITC (11-0043) (eBiosciences); CD11b/Mac-1 PE (550019) (Becton Dickinson). Data was compiled and analysed using CellQuest software (Becton Dickinson).
At 8 weeks of age, Tdp2+/+ and Tdp2Δ1–3 mice underwent intraperitoneal injection with 2.5 µl/g of body weight of either 10% DMSO (vehicle control) or etoposide at 400 µg/ml in 10% DMSO for a final dose of 1 mg/kg. Mice were sacrificed by cervical dislocation 24 h after injection and BM from one femur and tibia was extracted and homogenized in 3 ml FBS. Cellular content was concentrated in 150 µl FBS by centrifugation and smears were prepared on glass slides. Following 5 min fixation in methanol, slides were stained 30 min in Giemsa-modified (Sigma) solution (5% v/v in 100 mM Tris-HCl pH 6.8) and visualized under the microscope. 2000 polychromatic erythrocytes (PCE) were scored for the presence of micronuclei (MN-PCE) in each slide.
|
10.1371/journal.pgen.1004696 | Fast Evolution from Precast Bricks: Genomics of Young Freshwater Populations of Threespine Stickleback Gasterosteus aculeatus | Adaptation is driven by natural selection; however, many adaptations are caused by weak selection acting over large timescales, complicating its study. Therefore, it is rarely possible to study selection comprehensively in natural environments. The threespine stickleback (Gasterosteus aculeatus) is a well-studied model organism with a short generation time, small genome size, and many genetic and genomic tools available. Within this originally marine species, populations have recurrently adapted to freshwater all over its range. This evolution involved extensive parallelism: pre-existing alleles that adapt sticklebacks to freshwater habitats, but are also present at low frequencies in marine populations, have been recruited repeatedly. While a number of genomic regions responsible for this adaptation have been identified, the details of selection remain poorly understood. Using whole-genome resequencing, we compare pooled genomic samples from marine and freshwater populations of the White Sea basin, and identify 19 short genomic regions that are highly divergent between them, including three known inversions. 17 of these regions overlap protein-coding genes, including a number of genes with predicted functions that are relevant for adaptation to the freshwater environment. We then analyze four additional independently derived young freshwater populations of known ages, two natural and two artificially established, and use the observed shifts of allelic frequencies to estimate the strength of positive selection. Adaptation turns out to be quite rapid, indicating strong selection acting simultaneously at multiple regions of the genome, with selection coefficients of up to 0.27. High divergence between marine and freshwater genotypes, lack of reduction in polymorphism in regions responsible for adaptation, and high frequencies of freshwater alleles observed even in young freshwater populations are all consistent with rapid assembly of G. aculeatus freshwater genotypes from pre-existing genomic regions of adaptive variation, with strong selection that favors this assembly acting simultaneously at multiple loci.
| Adaptation to novel environments is a keystone of evolution. There is only a handful of natural and experimental systems in which the process of adaptation has been studied in detail, and each studied system brings its own surprises with regard to the number of loci involved, dynamics of adaptation, extent of interactions between loci and of parallelism between different adapting populations. The threespine stickleback is an excellent model organism for evolutionary studies. Marine-derived freshwater populations of this species have consistently acquired a specific set of morphological, physiological and behavioral traits allowing them to reside in freshwater for their whole lifespan. Previous studies identified several genomic regions responsible for this adaptation. Here, using whole-genome sequencing, we compare the allele frequencies at such regions in four derived freshwater populations of known ages: two natural, and two artificially established in 1978. Knowledge of population ages allows us to infer the strength of selection that acted at these loci. Adaptation of threespine stickleback to freshwater is typically fast, and is driven by strong selection favoring pre-existing alleles that are likely present in the ancestral marine population at low frequencies; however, some of the adaptation may also be due to young population-specific alleles.
| Studies of adaptation in nature are complicated by the typically long timescales at which evolution proceeds, and therefore are rather rare (e.g. [1]–[3]). Positive selection, the hallmark of adaptation, can be inferred from patterns of divergence and/or polymorphism in genome comparisons. While experimental evolution coupled with searches for patterns consistent with positive selection is becoming an accepted tool for “real time” studies of adaptation in microbes [4], [5], it is rarely possible to use genomic data to observe the adaptation process in higher animals such as vertebrates [6], [7]. Furthermore, the mechanisms of adaptation at the genomic level are still poorly understood [8]–[10].
The study of the genomics of adaptation has experienced a recent upheaval since the advent of population-level next-generation sequencing, which enables identification of selected loci and detailed studies of divergence and polymorphism within them in a wide range of model systems [11]–[13]. The data reveal that the number of loci responsible for adaptation, the ratio of coding and regulatory changes, the proportions of parallel to non-parallel genetic changes vary between systems [13]–[15]. The reasons for such variation are still unclear, making further genomic studies of adaptation a priority.
Threespine stickleback (Gasterosteus aculeatus) has become a widely used model organism for studying adaptation and speciation [16], [17]. The species is very variable, and is represented by a number of morphs [18], [19]. The ancestral populations of G. aculeatus likely lived in the sea, and colonization of new freshwater habitats, followed by evolution of freshwater populations, occurred repeatedly all over the Northern Hemisphere. While fish from marine populations utilize freshwater lakes and streams only as temporary spawning grounds, thousands of isolated freshwater resident populations have been independently established, and they have diverged in morphological, physiological and behavioral traits allowing them to survive in the freshwater for their entire lifespan. Independent origin of freshwater populations of G. aculeatus in different locations in the Northern Hemisphere provides an opportunity to study adaptive evolution under similar environments [15], [20]–[24]. Much of this adaptive evolution has been shown to be parallel, involving repeated recruitment for adaptation at different freshwater populations of the same pre-existing alleles that are presumably carried at low frequencies by marine populations. However, some of the adaptations are specific to individual populations [15],[22]; the relative importance of adaptations by new mutations vs. standing variation, and of population-specific vs. parallel adaptations, is not known.
Freshwater and marine forms of G. aculeatus possess a number of phenotypic differences. One of the most obvious is their armor plates: while the marine form has a complete set of lateral plates covering their body from pelvic girdle to the caudal peduncle, there are usually just a few lateral plates in the freshwater form [25],[26]. Genetic differences responsible for the number of the armor plates have been identified, pointing to the EDA gene on chromosome IV [27]. Later, sequencing of the G. aculeatus genome (available at http://genome.ucsc.edu) facilitated studying the genetic basis of the differences between the two forms, and several large genomic regions with high concentrations of nucleotide substitutions between the forms were found by comparing individuals from marine and freshwater habitats in a RAD-Seq analysis [22]. A recent study of G. aculeatus from Atlantic and Pacific basins used whole genome sequencing to reveal more than two hundred small genetic regions throughout the stickleback genome that differ between the forms [15].
Populations of G. aculeatus adapted to freshwater inhabit lakes and streams that originated after the retreat of the Pleistocene glaciers, indicating that adaptation can be fast [16]. Although rather different phenotypically, the freshwater and the marine forms often can hybridize and produce fertile offspring [18],[28],[29]. However, in some populations, there can also be a nearly-complete reproductive isolation in natural habitats between freshwater residential populations and anadromous marine forms spawning in the same lake [30],[31]. Reproductive isolation is mediated by phenotypic traits [32], and generally, there is not a clear cut relationship between the age of freshwater populations and reproductive isolation between marine and freshwater morphs.
Studies of G. aculeatus in the White Sea and its basin were initiated by Valery Ziuganov in the 1970s [33],[34]. The upper boundary for the age of the marine population in the White Sea is 15,000–18,000 years, because earlier, this area was covered in ice sheets during the Last Glacial Maximum [35]. After the end of the glaciation, the White Sea region experienced eustatic raising, giving rise to a unique system of young lakes as bays gradually separated from the sea by lift of the coast [36]. The rate of this process, which is still ongoing, has been estimated as 3.8 mm/year [37]; this allows inferring the age of a freshwater population from the elevation of the lake above sea level. Furthermore, in 1978, Ziuganov established several independent artificial sticklebacks populations in abandoned mica and spar quarries filled with ground water, by seeding each quarry with controlled numbers of marine and freshwater individuals [33]. Sampling these populations in 2011 allowed us to study two evolutionary trajectories with known points of departure. Thus, the availability of a wide range of young lakes of known ages in the White Sea basin provides an opportunity to trace the dynamics of adaptation to freshwater environments.
Here, we use whole-genome sequencing to study eight populations of G. aculeatus from the White Sea basin, including two artificial populations. We aimed to detect the genetic differences between the ancestral marine and the derived freshwater populations, and to measure the rate of adaptation, and the strength of positive selection which drives it, at divergent genomic loci. Whole-genome comparisons of multiple artificial and natural derived populations allow detailed analysis of selection acting simultaneously at multiple loci. Using multiple populations of different ages allows studying the process of adaptation at these loci at a range of time points, from tens to hundreds of years, and the uniformity of the process of selection. Finally, whole-genome analysis reveals the detailed patterns of divergence and polymorphism within the selected loci and in their vicinity.
We searched for the genetic markers of differences between the ancestral population of G. aculeatus in the White Sea and the derived freshwater populations in its vicinity. To identify such markers, we compared the genome sequences of two samples of marine individuals with two samples of freshwater individuals (Figure 1, Table 1). Phenotypically marine individuals were collected in Nilma bay and among the anadromous (marine) fish in Lake Ershovskoye where they came to spawn. Phenotypically freshwater individuals were collected from Lake Lobaneshskoye on the Island Velikiy and Lake Mashinnoye on the mainland. Their ages since desalination, inferred from their current elevations above the sea level, are ∼600 and ∼700 years, respectively [37]. By using these two freshwater populations of independent origins which are the oldest in the area, we aimed to identify those genetic changes that occurred in parallel in both freshwater populations, and therefore likely include sites responsible for adaptation of G. aculeatus to freshwater.
We estimated allele frequencies from pooled samples of individuals; these allele frequencies were confirmed using allele-specific PCR for specific loci (see below). We defined “marker” single-nucleotide polymorphisms (SNPs) as polymorphic nucleotide sites where both marine samples contained a particular allele at frequencies above 80%, while both freshwater samples contained another allele at frequencies above 80% (“strong criterion”) or above 50% (“weak criterion”). For comparison, we also identified marker SNPs according to the strong criterion using only one marine-freshwater pair of populations (Nilma vs. Mashinnoye).
Identified marker SNPs were distributed unevenly along the reference genome, clearly consisting of dense aggregations (“divergence islands”, DIs; [11],[15]) of markers in short genomic regions. A strong (weak) DI was defined as a continuous region where each 10 Kb window carried at least 10 strong (20 weak) markers, after merging any two such regions that are closer than 40 Kb to each other, because recombination is not likely to occur on such short distance [38],[39]. This definition, which seems to describe our data well (Figure 2), leads to a smaller number of wider DIs than the definitions used in [15] (see Methods for an alternative approach). Among the 6,107 marker SNPs obtained under the strong criterion, 5,801 (95.0%) were concentrated in DIs. By overlapping the strong and the weak criteria, we identified 19 DIs, which were located on ten out of the 21 G. aculeatus chromosomes (Figure 2) and covered a total of 3,301,948 nucleotides, or 0.74% of the genome.
In the majority of the DIs, the number of weak markers is only slightly above the number of strong markers, indicating that the freshwater-specific alleles have usually reached the frequency of 80% in both lake populations. Two exceptions are DIs IV-5 and XI-1 (i.e., the fifth DI on the 4th chromosome, and the first DI on the 11th chromosome), which, although identifiable both by the strong and the weak criteria, contains twenty times as many weak markers as strong markers. Furthermore, five DIs (II-1, VII-1, IX-1, IX-2 and XXI-1) could be identified only by the weak criterion, indicating that the frequency of freshwater-specific alleles in one or both freshwater populations is below 0.8.
The overall nucleotide diversity within the two freshwater samples was reduced by 22% relative to that in the two marine samples (Table 1), consistent with lower effective sizes of freshwater populations and/or moderate bottlenecks in the course of their origin. Both strong and weak DIs in marine population carried higher levels of nucleotide diversity (0.0049 and 0.0034, respectively), compared to the genomic background (0.0020). In the freshwater populations, diversity was reduced in the strong DIs (0.0012), but elevated in the weak DIs (0.0048), compared to the genomic background (0.0016).
Seventeen out of the 19 detected DIs overlap protein-coding genes, for a total of 170 genes (Table 2, Table S1); among the 285 marker SNPs identified under the strong marker criteria and covered by these genes, 139 were nonsynonymous, while 146 were synonymous (Table 3). The ratio of nonsynonymous to synonymous marker SNPs was much higher than that within non-marker SNPs in protein-coding genes segregating within the marine population (290 nonsynonymous to 528 synonymous; Table 3), implying positive selection favoring the preferential fixation of amino-acid changing marker SNPs between marine and freshwater populations [40]. However, 7 of the 17 DIs do not include any nonsynonymous marker SNPs. The remaining two DIs do not overlap any known protein-coding or miRNA genes. The DIs encompass genes that might affect several traits responsible for phenotypic difference between the marine and freshwater forms, including the well-known EDA gene responsible for body armor ([27],[41], DI IV-1), as well as genes likely to be important for adaptation to freshwater through their effects on osmoregulation, immunity, or morphology: Na+/K+ transporting ATPase (ATP1A1 [42], DI I-1), neurotransmitter and hormone binding (SULT4A1 [43], DI IV-4), and immunity response to viral infection (NLRC5 [44], DI XIX-1). Other genes might be involved in several important aspects of metabolism and behavior (INHA [45], DI I-1), responsible for growth and development of nerve cells (SLITRK2 [46], DI IV-5), adhesion and differentiation of nerve cells (CTNNA2 [47], DI IX-3), calcium/phosphate homeostasis (STC2 [48], DI IV-2), and mediation of functions in the central and peripheral nervous systems (HTR3A [49], DI V-1).
Only 5.0% of marker SNPs identified under the strong marker criteria were not located within any of the DIs. This amounted to a total of 306 marker SNPs, located on 19 out of the 21 chromosomes. 21 of the 306 SNPs were located within protein-coding regions; of these 21 SNPs, 17 were amino-acid changing, while only 4 were synonymous. Again, this ratio of nonsynonymous to synonymous substitutions is higher than that observed within a single (marine) population (28,616 and 41,902, respectively; Table 3), consistent with positive selection favoring amino acid-changing mutations even outside the DIs. Notably, among the 12 genes that carried amino-acid changes, a number could be plausibly held responsible for adaptation and speciation in G. aculeatus. For example, GCNT3 gene, which plays an important role in biosynthesis of mucin which is used for nest building [50], carries 4 marker SNPs, all nonsynonymous. MUC-like gene on the chromosome II carries 2 non-synonymous substitutions. Two nonsynonymous non-DI marker SNPs are positioned within 300 bp of the EPX-like gene on chromosome XIII; EPX is a gene contributing to the activity of eosinophils which are responsible for lysis of parasites [51]. Another nonsynonymous SNP is in INSR gene on the chromosome III; this gene is known to play a crucial role in early development and growth, and in the development of the neural system [52],[53].
In addition to lakes Mashinnoye and Lobaneshskoye, we also sampled G. aculeatus from four bodies of freshwater of more recent origin: quarries Goluboy and Malysh, and lakes Martsy and Ershovskoye (Table 1), all located near the White Sea. Populations in Goluboy and Malysh were established by Ziuganov in 1978 in isolated pools that developed in quarries after they had been abandoned, but were, before the start of the experiments, devoid of fish; therefore, they both were 34 years old at the time of sampling. The Quarry Goluboy population (area ∼70,000 m2, carrying capacity over 1,000 fish) was started from 20 marine (10 females and 10 males) and 20 freshwater (10 females and 10 males) individuals. The Quarry Malysh population (∼75 m2, carrying capacity about 100 fish, but the number of reproducing males may be limited by the very low number of nesting sites) was started from 1 marine female and 1 freshwater male [34]. Founding marine individuals were taken from the White Sea, and founding freshwater individuals were taken from Lake Mashinnoye. Lakes Martsy and Ershovskoye originated through isolation of marine bays due to the steady glacio-isostatic rise of the coast at the rate of ∼3.8 mm per year [37]. The age of the freshwater population in Lake Martsy can be estimated as ∼250 years, because the surface of the lake is currently at about 1 meter above the sea level. As recently as in 1978, Lake Ershovskoye (now ∼14 cm above spring-tide level) was a typical meromictic lake, inhabited by only anadromous fish with the typical marine phenotype [33]. This lake became fresh soon afterwards, and now contains an abundant residential population, which can be easily distinguished from anadromous individuals both morphologically and by a different parasite load (Schistocephalus solidus are dominant in residential individuals, and nematodes in anadromous individuals [33]); therefore, we estimate the age of Lake Ershovskoye also as 34 years. Thus, at these four bodies of water, the adaptation of G. aculeatus populations to freshwater is likely to still be ongoing.
In all four young populations, frequencies of freshwater alleles at marker SNPs within DIs have been increasing rapidly (Figure 3, see Table S2 for allele frequencies data on all populations). For five of the DIs, the estimates of allele frequency obtained by Illumina sequencing of DNA pools were also validated by genotyping individual fish from each population with allele-specific primers (Table S3, Table S4). These increases imply that in each of the freshwater populations, selection favors the identified freshwater alleles. The initial frequencies of freshwater alleles in natural lakes Ershovskoye and Martsy were likely the same as the frequency observed in the marine population, i.e., ∼0.1 (Table S2). In the two quarry populations, the initial frequencies were assumed to equal 0.5. The mean frequency of freshwater alleles over all DIs in the artificial populations was 0.56 at Quarry Malysh, and 0.73 at Quarry Goluboy (Figure 3B). A lower average and a higher variance of freshwater allele frequencies at Quarry Malysh population are likely due to its small effective population size, and therefore, stronger genetic drift. The frequencies of freshwater alleles in Lake Martsy are higher than in Lake Ershovskoye, consistent with the former being older than the latter.
Freshwater alleles also increased in frequency at the marker SNPs located outside DIs. This was observed for all chromosomes in the two natural populations, except chromosomes XIV and XVI, each carrying only one marker SNP (Figure S1A); and for most of the chromosomes in the Quarry Goluboy artificial population (Figure S1B). Overall, however, the marker SNPs located within DIs have reached substantially higher frequencies in all freshwater populations, compared with the marker SNPs outside DIs (Figure 4). An increase of the frequency of freshwater alleles was also observed when only one marine-freshwater pair of populations (Nilma vs. Mashinnoye), rather than two pairs, was used to define the marker SNPs, although it was even less pronounced (Figure S3).
Knowing the rate of increase of an allele frequency makes it possible to estimate the strength of positive selection favoring this allele [54]. Because the age of the Lake Martsy is known only approximately, we cannot reliably estimate the rate of allele frequency change in it, while for the Lake Ershovskoye, the age of the population is known rather precisely. Furthermore, among the two artificial populations, selection in Quarry Goluboy was more pronounced than in Quarry Malysh, probably due to a stronger contribution of drift in the latter (see above). Therefore, for estimation of selection strengths, we used Lake Ershovskoye and Quarry Goluboy populations. We made such estimates assuming that the generation time of G. aculeatus is two years ([1] and our data; see Methods). We also made the simplifying assumptions that selection remained constant, and that freshwater alleles possess intermediate dominance (h = 0.5), as has been recently shown to be true, in particular, for most skeletal quantitative trait loci (QTL) in G. aculeatus [55].
We estimated the selection coefficient s for each DI (no estimates were made for non-DI marker SNPs, as estimates based on individual SNPs are unlikely to be robust). Only one selection coefficient s was ascribed to a DI, on the basis of the average frequency of freshwater marker alleles in it. As the allele frequencies within a DI are non-independent, the strength of selection cannot be estimated with precision, and the estimates for individual DIs should be treated with caution; still, this analysis allows us to appreciate the selection in effect.
Overall, estimated values of s were high: out of the 19 DIs, fifteen in Lake Ershovskoye, and twelve in Quarry Goluboy, had s>0.1 (Table 2, Figure 5). The selection coefficients estimated for each DI were not significantly correlated between the two freshwater populations (Figure 5; Spearman's rho = 0.30, p = 0.27), possibly due to the low number of DIs and high variance in estimation of s for individual DIs; however, the mean values of s inferred from Lake Ershovskoye (s = 0.16) and Quarry Goluboy (s = 0.13) were rather similar. In particular, DI IV-1, carrying the EDA allele, experienced s = 0.19 in Lake Ershovskoye, and s = 0.09 in Quarry Goluboy, consistent with the previously reported data on dynamics of armor phenotype over an even shorter time period [1]. The mean shift in allele frequency observed in Lake Ershovskoye was the largest in DI V-1, where freshwater allele frequency has changed from 0.1 to 0.56, corresponding to s = 0.255. This DI is centered on several nonsynonymous substitutions in gene HTR3A, a subunit of serotonin ligand-gated ion channel receptor with a wide spectrum of physiological functions. The close second was DI I-1 (sE = 0.247, sG = 0.212), a 470 kb long chromosomal inversion overlapping 27 genes, including the ATP1A1 gene which encodes a well-studied Na+/K+ transporting ATPase; the differences between the freshwater and the marine allele of this gene include 6 amino acid substitutions, and may be responsible for the differences in osmoregulation between the two forms [56]. In Quarry Goluboy, the most radical change in allele frequency was observed at DI XII-2 (the freshwater allele has reached near-fixation here, so s is hard to estimate). This region overlaps the upstream region of the gene OVGP1 (estrogen-dependent oviduct protein or mucin-9). This gene is involved in reproduction [57], and therefore is a likely target for positive selection; whether the strong selection associated with formation of freshwater phenotype is associated with arising reproductive isolation between the two forms should be the subject of a further study.
We have identified more than 18,000 marker SNPs that distinguish the two freshwater populations of G. aculeatus from the White Sea basin from the ancestral marine population. The great majority of these markers cluster within 19 short genomic regions, or DIs (Figure 2). All the DIs we found overlap with regions found in previous studies [15],[22], indicating a substantial degree of genetic parallelism in the origin of geographically distant freshwater populations and supporting the significance of the corresponding genomic regions for the process of adaptation.
Out of the 19 DIs, 12 overlap the regions reported both in [22] and [15] as responsible for adaptation to freshwater (Table 2), and the remaining 7 overlap loci reported in [15] only. Out of the 81 top-scoring regions described in [15], 71 regions are located within our DIs. 7 out of our 19 DIs (I-1, IV-1, IV-3, IV-4, VII-1, XIX-1, and XIX-2) overlap the top-scoring regions identified in [15] by both SOM/HMM and CSS analyses, and 12 more overlap the regions identified in [15] by less stringent criteria.
Extensive overlap between the adaptive regions described here for the White Sea populations and the ones reported from other regions confirms the previously described widespread parallelism in stickleback evolution [15],[22],[24],[27],[58]. A caveat is that both our study and the previous studies are specifically focused on finding genomic signatures of parallel evolution, and it is hard to infer the extent of parallelism from them directly. When only one, rather than two, marine-freshwater pairs of populations is used to identify marker SNPs, their number increases radically, likely due to addition of some true loci of population-specific positive selection, but also due to an increased number of false positives (Figure S3). Future studies involving many populations, and comparing loci involved in adaptation in some of the populations but not in others, are necessary to reveal the exact extent of parallelism.
The aggregated distribution of marker SNPs in the genome is probably determined by recombination patterns. For each DI, the observed increase of allele frequencies could be driven by positive selection favoring the “freshwater” allele at only one crucial polymorphic site, accompanied by hitch-hiking at surrounding neutral SNPs [59]. Alternatively, several distinct adaptive loci may be clustered together in regions of low recombination, with selection acting simultaneously at several such loci [15],[55]. Both these explanations are consistent with the fact that the three longest DIs found in our study all correspond to known inversions (DIs I-1, XI-1, XXI-1; Figure 2), i.e., the regions where recombination is minimal and hitch-hiking is most pronounced. In any case, our data do not imply that freshwater alleles at all the markers within a DI are favored by selection, even when a DI encompasses several protein-coding genes; the vast majority of the marker SNPs are probably neutral, and get fixed between the marine and freshwater populations due to hitch-hiking. Still, among the SNPs within the protein-coding genes, the nonsynonymous-to-synonymous ratio is higher for marker SNPs than for SNPs segregating within the ancestral marine population (Table 3), consistent with positive selection favoring marine-freshwater divergence at nonsynonymous sites [40]. Nine out of the 19 DIs do not contain any genes with nonsynonymous substitutions between freshwater and marine populations (Table 2), suggesting that evolution of regulatory mechanisms played a major role in the process of adaptation [15],[60],[61].
We observed a gradual increase of freshwater allele frequencies inside DIs (Figure 3 and Table S2) by exploring two young lake populations as well as two artificial populations established in 1978 from equal numbers of founder individuals of different forms. The selection coefficients estimated from the rates of increase of freshwater allele frequencies in Quarry Goluboy and Lake Ershovskoye are generally high (mean s = 0.13 and 0.16, respectively), implying rapid adaptation of G. aculeatus to the lacustrine environment (Table 2). The selection coefficient inferred for the EDA allele (s = 0.19 and 0.09) is somewhat lower than that inferred previously from short-term experimental populations (s∼0.5 [6]). Outside sticklebacks, the observed range of selection coefficients is comparable to that acting in the course of adaptation of Biston betularia peppered moth butterflies to predation (0.05–0.16 [2]), or estimated for the lactose-persistence allele in humans (0.014–0.19 [62]), and exceeds those estimated for PDYN promoter (<0.01 [63]) or genes involved in pigmentation (0.02–0.10, [64]) in humans. The variance of the frequencies of freshwater alleles among DIs is larger in the Quarry Malysh population than in the Quarry Goluboy population. The population of Malysh is much smaller than that of Goluboy and, therefore, this difference is possibly due to stronger genetic drift [65] in Malysh.
Within-DI densities of marker SNPs are rather high, and the marine and freshwater haplotypes differ from each other at over 1% of nucleotide sites within some DIs. Because most of marker SNPs are probably selectively neutral by themselves, this implies that these haplotypes diverged ∼106 generations ago, assuming the mutation rate of 10−8 per nucleotide site per generation [66]. Such high divergence times are obviously inconsistent with de novo origin of freshwater alleles in each of the freshwater populations. Instead, they are consistent with repeated recruitment of the same ancient alleles in the course of establishment of different freshwater populations.
Only 5.0% of the marker SNPs identified under the strong marker criteria is located outside DIs. Some of these markers are in fact clustered, although they do not form DIs under our formal criteria. For example, at chromosome XIII, which carries no DIs, 11 out of the 18 markers are located within 0.5 Mb from each other. Notably, these marker SNPs also have the nonsynonymous-to-synonymous ratio substantially exceeding that in non-marker SNPs segregating within the marine population (Table 3). The fraction of nonsynonymous SNPs among coding marker SNPs is even higher outside DIs (17 out of 21) than within DIs (139 out of 285; Fisher's test, two-tailed P = 0.0056), suggesting that the fraction of marker SNPs under selection may be even higher outside than within DIs. Arguably, the differences in allelic frequencies of non-DI SNPs could be due to hitch-hiking with the DI marker SNPs; however, we see no difference in the rate of increase when the non-DI marker is located on a chromosome with a DI vs. without a DI (Figure S1A – natural populations; Figure S1B – artificial populations), and for markers at chromosomes with DIs, there is little correlation with the distance from the DI (Figure S2 – two natural lakes), suggesting that at least some of the non-island markers are also targets for selection. Plausibly, these SNPs are young and not yet surrounded by as many adjacent neutral hitchhikers accompanying them; the higher fraction of selected SNPs among them is therefore as expected. Still, each individual marker SNP located outside DIs is probably more likely to be indeed neutral, and to result from genetic drift and/or sampling bias, than a DI. Indeed, although the non-DI marker SNPs also increase their frequency in freshwater, this increase is less pronounced than within DIs (Figure 4), probably due to a higher fraction of neutral loci and/or weaker selection in the former.
7 out of the 19 detected DIs are “weak”. Analysis of such DIs reveals a number of patterns. In some of the weak DIs (VII-1 and IV-5, Figure 3A), the frequencies of freshwater alleles are close to 80%, but do not reach this threshold in either of the two populations from lakes Mashinnoye and Lobaneshskoye, so that the strong marker criteria do not hold. In others, although the frequencies of freshwater alleles seemingly increased gradually with the age of the population, they have not reached 80% in both lakes, perhaps due to weakness of selection favoring them (DIs II-1, IX-1, IX-2, and XI-1). In still others, freshwater alleles rapidly reach rather high frequencies in the young population, but these frequencies remain at the same level when the population's age increases (DI XXI-1, Figure 3A). The latter scenario could conceivably be due to the some form of balancing selection, perhaps due to interactions between the genes linked within the long inversion which constitutes DI XXI-1.
According to the “transporter hypothesis” [23], freshwater alleles are constantly present at low frequencies in the marine population, probably due to rare emigration from freshwater populations, and are recruited when a new freshwater population is established. The fact that alleles recruited in different freshwater populations tend to coincide, and that freshwater and marine haplotypes are highly divergent within DIs, support this hypothesis. Thus, a sort of balancing selection acts on the sites directly involved in adaptation to freshwater at the level of the global metapopulation [67] of G. aculeatus, keeping freshwater alleles from extinction. This metapopulation consists of the ancestral anadromous marine population and many derived residential freshwater populations. While the individual derived populations are often short-lived, the metapopulation has probably existed during much of the history of the species.
Although the ancestral alleles favored in the sea and the derived alleles favored in freshwater have coexisted for a long time, they have had only occasional opportunities to be separated by recombination from adjacent neutral polymorphisms. Indeed, this recombination can happen only in new residential populations, before fixation of alleles favored in freshwater together with linked neutral markers around them, or during the presumably short periods of time when such alleles exist in the sea before being recruited for formation of a new residential population. Thus, the width of a DI must be determined by the strength of selection favoring the freshwater alleles in freshwater populations and disfavoring them in the sea, by the recombination rate, and by the number of generations between the time when the freshwater allele has escaped from one freshwater population into the sea and the time when it has become recruited in another emerging lake residential population. Nucleotide diversity within strong DIs in our freshwater populations is somewhat lower than in the marine population, but not radically, indicating that adaptation to each freshwater lake has involved soft, rather than hard, selective sweeps [68]. Indeed, soft sweeps involve recruitment of multiple simultaneous sweeping haplotypes, and thus do not lead to a significant reduction in the nucleotide diversity around the selected site. Soft sweeps are likely when the sweeping alleles arise from pre-existing genetic variation rather than de novo mutations, and thus the lack of major reduction in diversity also supports the transporter hypothesis [23]. When a freshwater allele is brought into a new lake population by several individuals, nucleotide diversity is expected to increase on both sides of the DIs due to unequal length of freshwater DIs in founder individuals [69]; however, we see no such effect, probably due to the young age of our populations.
The very high evolutionary rate observed at several of the DIs during transition from marine to residential form could be attributed to pre-existing genomic regions, recruited from the standing variation of the marine population. Such “precast bricks” allow emerging freshwater population of sticklebacks to build rapidly a phenotype adapted to various challenges (salinity, parasites, energy metabolism, etc.) which it faces in the new environment. Plausibly, this form of evolution may be widespread beyond the stickleback model. Rapid emergence of parallel well-differentiated autochthonous flocks in the genus Eubosmina (Cladocera: Crustacea) in European lakes [70], flocks of Labeobarbus (Cyprinidae: Teleostei) in lakes and rivers of Ethiopia [71],[72], and genus Salvelinus (Salmonidae: Teleostei) [73],[74] could be a few out of many examples of this kind of evolution from precast bricks, during which new adaptive phenotypes are repeatedly created by rearrangement of ancient genetic elements, which were formed during earlier adaptive radiations and retained in ancestral population as standing variation.
Fish were collected in June–August 2011 by scoop-net or landing-net, anaesthetized and euthanized with a tricaine methane sulphonate solution (MS222), and then immediately fixed in 96% ethanol on site. Fish euthanasia was conducted under the supervision of the Ethics Committee for Animal Research of the Koltzov Institute of Developmental Biology Russian Academy of Sciences. Location of lakes and quarries, estimated age of population, and sample size are presented in Table 1.
Total genomic DNA was extracted from each individual using Wizard genomic DNA purification kit (Promega). Prior to library preparation, DNA samples of between 8 and 20 (Table 1) fish from the same population were pooled in equal proportions. Resulting pooled DNA samples were processed as described in the TruSeq DNA Sample Preparation Guide (Illumina). Library lengths were estimated using 2100 Bioanalyzer (Agilent). Libraries were quantified using fluorimetry with Qubit (Invitrogen) and real-time PCR (primers I-qPCR-1.1: AATGATACGGCGACCACCGAGAT and I-qPCR-2.1: CAAGCAGAAGACGGCATACGA) and diluted to final concentration of 9 pM. Diluted libraries were clustered on a paired-end flow cell using cBot instrument and sequenced using HiSeq2000 sequencer with TruSeq SBS Kit v3-HS (Illumina), read length 101 from each end. Sequences for each population are available at the NCBI Short Read Archive (http://www.ncbi.nlm.nih.gov/Traces/sra; accession number of the project SRP023197).
The reads were mapped onto the reference genome of G. aculeatus downloaded from the UCSC (http://genome.ucsc.edu/) using bwa aln program of the BWA (Burrows-Wheeler Alignment Tool) package (http://bio-bwa.sourceforge.net/). Output was then converted to SAM format using bwa sampe. Next, data were processed with picard (http://picard.sourceforge.net/) in order to remove duplicated reads. We identified SNPs in all populations using program mpileup of the samtools package (http://samtools.sourceforge.net/). For SNP calling, different depth cutoffs were used for different populations due to differences in read coverage among populations: >10 for Nilma, Malysh, Goluboy and residential fish from Ershovskoye, and >5 for Mashinnoye, Lobaneshskoye, Martsy, and anadromous fish from Ershovskoye. To minimize sequencing errors, positions with base qualities lower than 40 within a population were excluded from the analysis. As an alternative approach to SNP calling, we used UnifiedGenotyper program from GATK package (http://www.broadinstitute.org/gatk/) with the same coverage cutoffs. This led to a larger pool of marker SNPs, but similar clustering patterns: all the DIs identified with the program mpileup were observed, as well as several new clusters (Table S5). The patterns of allele frequency dynamics were qualitatively similar under mpileup and GATK SNP calling (Table S6). Overall, a higher proportion of marker SNP were located outside DIs using GATK (Table S5); therefore, we chose to use mpileup for the results in the main text.
Positions of genes were derived from Ensembl database release 72 (http://www.ensembl.org/)
To define DIs, we merged any two regions with above-threshold numbers of marker SNPs that were closer than 40 Kb to each other. Generally, this merging procedure described our data well: for example, it prevented splitting several DIs all corresponding to a single known inversion, or division of one DI into several due to the lack of coverage in some regions. Not merging adjacent DIs led to a drastic increase in their number [71]; the qualitative patterns of allele dynamics remain the same (Table S7). This is as expected, because recombination (average recombination rate in threespine stickleback is 3.11 cM/Mb [39]) is not likely to occur between regions located so close to each other (less than 40 Kb) over the considered timescales.
To validate our estimates of allelic frequencies based on high-throughput sequencing data, we also genotyped each fish used for pooled DNA sample with an allele-specific set of primers for markers located within several of the DIs. For this purpose, we designed 8 allele-specific sets of primers for 7 of the DIs (one DI was genotyped with two sets of primers). Each set consisted of three primers: two allele-specific, and one anchor primer. Additionally, we used previously published primers Stn382 [27] to genotype DI IV-1. Primers, positions of target SNPs, and PCR annealing temperature for each pair of primers are presented in Table S3. Two allele-specific PCR reactions (each with one allele-specific and one common anchor primers) were set for each individual, and the second PCR product was applied in the same well of agarose gel as the first PCR reaction after 5 min of running the gel. Individuals with one or both PCR products were categorized as homo- or heterozygotes, respectively. The obtained allele frequencies matched well those estimated from high-throughput sequencing data (Table S4).
As a proxy for the time of formation of natural residential populations, we use the time of desalination. Before complete desalination, a lake is meromictic, and contains two water layers – a higher freshwater layer, and a lower saltwater layer, forming a halocline. This halocline prevents proper oxygenation; as a result, the lake becomes anoxic every winter, causing extirpation of residential populations [75].
For each DI, we calculated the average frequency of a freshwater allele over all marker SNPs (Table S2). We estimated generation time using length-cohort analysis, which revealed two cohorts present in each lake: immature one-year old fish, and a second-year class which participated in reproduction. Presence of three year-old and older fish in the lake population was negligible. Therefore, we assumed generation length of two years; reproduction at older ages will lead to underestimation of s. The ages of Goluboy and Malysh populations are known to be 34 years (17 generations). The Ershovskoye freshwater population is known to be 34 years old or younger [33]; we assumed its age to be 34 years (17 generations), and younger age will again lead to underestimation of the true s.
The initial frequencies of freshwater alleles in the two artificial populations, Goluboy and Malysh, were assumed to equal 0.5. In 1978, all fish in Lake Ershovskoye had phenotypic composition similar to that of a typical marine population [33]. We assume that the initial allelic frequencies in the Lake Ershovskoye matched the frequencies in marine populations, i.e., p0∼0.1. Selection coefficient s was calculated from the per generation change in allele frequency under the assumption that this change is driven by selection alone (Eqn. 3.2 in [76], assuming h = 0.5).
|
10.1371/journal.pgen.1005948 | Regulation of Gap Junction Dynamics by UNC-44/ankyrin and UNC-33/CRMP through VAB-8 in C. elegans Neurons | Gap junctions are present in both vertebrates and invertebrates from nematodes to mammals. Although the importance of gap junctions has been documented in many biological processes, the molecular mechanisms underlying gap junction dynamics remain unclear. Here, using the C. elegans PLM neurons as a model, we show that UNC-44/ankyrin acts upstream of UNC-33/CRMP in regulation of a potential kinesin VAB-8 to control gap junction dynamics, and loss-of-function in the UNC-44/UNC-33/VAB-8 pathway suppresses the turnover of gap junction channels. Therefore, we first show a signal pathway including ankyrin, CRMP, and kinesin in regulating gap junctions.
| The nervous system is made up of individual neurons connected by junction structures called synapses. There are two fundamentally different types of synapses: chemical synapses and electrical synapses (also called gap junctions). Through studies in different model organisms, we have gained rich knowledge about the development and dynamics of chemical synapses. However, we still know little about how gap junctions are formed during development and what regulates the dynamic of gap junctions in functional circuits. Using C. elegans PLM neurons as a model, we carried out an unbiased genetic screen and isolated mutants altering gap junctions. In this study, we focus on two of them, unc-44/ankyrin and unc-33/CRMP. Through genetic analysis in combination with live imaging, we find that UNC-44/ankyrin and UNC-33/CRMP play important roles in gap junction turnover and demonstrate that UNC-44/ankyrin acts upstream of UNC-33/CRMP and VAB-8/ kinesin to regulate the removal of gap junction channels from gap junctions.
| Gap junctions were first discovered in the myocardium and nerves for their properties of electrical transmission between two adjacent cells [1,2], and they are clusters of channels connecting two cells to allow direct transfer of ions and small molecules [3–5]. Gap junctions play essential roles in many biological processes, such as embryo development, cell differentiation, cell growth, metabolic coordination of avascular organs, and neural development [6]. In excitable cells, the presence of gap junctions provides them with abilities to generate synchronized electrical and mechanical outputs [3–5]. Gap junction channels form polymorphic maculae or plaques with a few to thousands of units [4] and are composed of connexins in chordates and innexins in prechordates [7,8]. Although connexins and innexins are not homologs in terms of their primary sequences, they share similar structures with four transmembrane domains, two extracellular and one intracellular loop, and intracellular amino- and carboxy- termini [5]. Vertebrates have innexin-related proteins, called pannexins, however, their roles in forming gap junctions are still under debate [9,10].
Regulation of gap junctions is observed at two levels: fast regulation involving change of channel conduction and open probability, and slow regulation including alternation of composition and turnover of channels [4]. Studies show that voltage changes and phosphorylation of channel proteins are important for fast regulation of gap junctions [11–14]. In terms of slow regulation, the turnover of gap junctions plays an important role [15]. Gap junctions have remarkably rapid turnover rate, for example connexin/Cx 43 has a half-life of only 1–3 hours [16–20]. The rapid turnover rate allows cells to quickly eliminate and rebuild their gap junctions to adapt to environmental conditions [4]. Gap junction channel proteins can form gap junctions in homomeric or heteromeric manners, and different combinations have distinct permselectivity [4]. The rapid turnover rate of gap junctions provides cells with the ability to change the composition of gap junctions in a timely manner. During gap junction turnover, the addition of new channels is at the edge of gap junction plaques, and the removal of channels happens at the center of the plaques [21,22]. Although recent studies show that the phosphorylation of gap junction channels and channel binding proteins is involved in regulating gap junction turnover, the molecular mechanisms orchestrating the removal of gap junctions are still largely unknown [4].
Transient gap junctions are important for mammalian brain development [3,23–25]. In invertebrates, transient gap junctions can regulate the formation of chemical synapses in leeches [26] and are required for asymmetry development of sensory neurons in C. elegans [27]. To avoid interruption of neuronal functions, those transient gap junctions need to be eliminated during development, but it remains unknown what regulates their elimination. Understanding the molecular mechanisms underlying gap junction dynamics may answer this question.
As scaffolding proteins, ankyrins can organize membrane proteins into discrete domains and integrate them with the cytoskeleton [28]. In neurons, ankyrin-G is essential for the assembly of axon initial segment (AIS) and nodes of Ranvier [29,30] and is important for synapse formation [31,32]. Collapsin response mediator proteins (CRMP) are conserved microtubule interaction proteins that regulate neuronal polarity and axon growth [33–35]. C. elegans has only one ankyrin homolog unc-44 and one CRMP homolog unc-33. Loss of function of unc-44 and unc-33 generate similar defects in locomotion, axon growth and axon-dendrite differentiation, suggesting they may function in the same pathway in regulating neuronal development [36–41]. Indeed, a recent study by Maniar et al. has shown that unc-44 acts upstream and regulates the localization of unc-33 in organization of microtubules in C. elegans neurons [41].
Although the important roles of ankyrin and CRMP in neuronal development have been documented in many organisms, their functions in regulating gap junctions have not been explored. Here, we show that in gap junction turnover, UNC-44/ankyrin acts upstream of UNC-33/CRMP and VAB-8/ kinesin to regulate the removal of UNC-9/innexin from gap junctions.
To study molecular mechanisms underlying gap junction regulation, we used C. elegans PLM neurons as a model. PLM neurons are a pair of mechanosensory neurons with simple morphology, that cell bodies are located at the tail region with a long axon growing to the middle part of body and a short posterior process toward the end of tail (Fig 1A) [42]. Electron microscope studies showed that PLM formed gap junctions at two regions along the axon: at zone 1, PLM neurons form gap junctions with PVC, LUA and PVR neurons [43]; at zone 2, PLM neurons form gap junctions with BDU neurons [44] (Fig 1A). Three innexins, unc-9, inx-7 and inx-3, are expressed in PLM neurons [45]. To visualize PLM gap junctions in vivo, we used GFP labeled UNC-9 as marker. UNC-9 has four transmembrane domains and cytoplasmic N-terminus and C-terminus (S1A Fig). Tagging GFP to the UNC-9 N-terminus affects its interaction with the gap junction regulator UNC-1, but UNC-9::GFP retains its endogenous localization and forms functional gap junctions [46]. Since UNC-9 has a rather short intracellular N-terminus (28 amino acids), we tested whether tagging GFP to its N-terminus will affect its function. We found that expression of GFP::UNC-9 under its own promoter (a 2.5kb fragment upstream of the start codon) rescued the uncoordinated phenotypes of unc-9 mutants to a similar level as untagged UNC-9 (S1 Movie (unc-9), S2 Movie (Punc-9::GFP::unc-9(5ng/ul);unc-9), S3 Movie (Punc-9:: unc-9(5ng/ul);unc-9)), supporting the idea that GFP::UNC-9 fusion protein may function in the same manner as untagged UNC-9. Therefore, we examined UNC-9 localization using transgenes expressing GFP::UNC-9 in PLM neurons. We found that GFP::UNC-9 formed stereotypical patterns at two gap junction zones with 2–3 GFP puncta at zone 1 and one GFP puncta at zone 2 (Fig 1A), and the same expression pattern was also observed in UNC-9::GFP transgene (Fig 1A). Using an UNC-9 specific antibody [46], we confirmed that endogenous UNC-9 formed similar punctate structures as those observed transgenes (Fig 1B). We further confirmed those puncta localized at the region where PLM neurons meet their gap junction partners. As shown in S1B Fig, we expressed mCherry in PVC neurons (Pglr-1::mCherry) and consistently observed that one GFP::UNC-9 punctum localized to the region where PVC axons crossed PLM axons. C. elegans stomatin protein UNC-1 is co-localized with and functionally important for UNC-9 containing gap junctions in muscle cells [46]. Consistent with this observation, we found that UNC-1 formed similar punctate patterns in PLM neurons (Fig 1A). In addition to forming gap junctions, UNC-9 and UNC-7 could also function as hemichannels in C. elegans motor neurons [47]. The conserved cysteines (Cys) at extracellular loops are essential for formation of UNC-7/UNC-9 gap junctions but not hemichannels [47]. We found that mutating these cysteines (Cys) to alanines (Ala) blocked the formation of GFP::UNC-9 puncta, supporting the conclusion that those puncta were gap junctions but not hemichannels (S1C Fig). All together, we believe that the GFP puncta in GFP::UNC-9 transgenes represent the localization of PLM gap junctions.
Since UNC-1 is co-localized and functionally important for UNC-9 containing gap junctions [46], it might be involved in gap junction assembly. To verify this possibility, we tested the formation of GFP::UNC-9 puncta in unc-1(lf) background and found that loss of function of unc-1 did not affect UNC-9 puncta, suggesting UNC-1 was not required for the formation of UNC-9 puncta in PLM neurons (Fig 1C). This observation is consistent with previous findings in muscle cells [46]. We also tested whether UNC-9 localization depended on inx-7, an innexin co-expressed in PLM neurons with unc-9. As shown in Fig 1C, we did not observe any defects of UNC-9 distribution in inx-7 mutants, suggesting that the assembly of UNC-9 puncta did not rely on other gap junction proteins. Previous studies showed that Netrin and its receptor Frazzled could regulate the formation of gap junctions between Drosophila interneurons and motor neurons [48]. We tested whether the Netrin signaling pathway was also involved in formation of PLM gap junctions. We found that loss of function of unc-6, the only Netrin homolog in C. elegans, did not affect the formation of UNC-9 puncta. There are two Netrin receptors in C. elegans, UNC-5 and UNC-40/DCC. Loss of function of unc-5 did not affect UNC-9 puncta, but about 3–5% unc-40/DCC(e271) mutant animals lost UNC-9 puncta (S1D Fig). Further studies will be necessary to confirm the function of UNC-40/DCC in gap junction regulation.
Gap junctions are required for the formation of chemical synapses in leeches [26]. The conserved Neurobeachin is involved in development of both chemical synapses and gap junctions in zebrafish [49]. It seems that the formation of gap junctions and chemical synapses could share some common mechanisms. To test this idea in C. elegans, we examined GFP::UNC-9 puncta in loss-of—function mutants of rpm-1 and syd-2, two genes playing important roles in C. elegans chemical synapse formation [50–53] and found that neither of them was required for the organization of UNC-9 puncta (Fig 1D). Double mutants of rpm-1 and syd-2 suppress the formation of chemical synapses [54], but we did not observe any defects of UNC-9 puncta in double mutant animals (Fig 1D). These results are in support of different mechanisms regulating gap junction and chemical synapse formation in C. elegans neurons.
To uncover the molecular mechanisms underlying gap junction regulation, we carried out an unbiased genetic screen using yadIs12 (Pmec-4::GFP::unc-9) as a starting strain and isolated mutants with three types of phenotypes: 1, type one mutants lost UNC-9 puncta; 2, type two mutants had more UNC-9 puncta close to the original gap junctions; 3, type three mutants had more UNC-9 puncta along the axon (Fig 1E). In this study, we focused on two mutants with type two phenotypes. As shown in Fig 2A and 2B, about 65% of yad21 and 35% of yad26 animals had more UNC-9 puncta at both zone 1 and zone 2, and the length of gap junction zone 1 in those animals was enlarged from 2–3 μm to about 25 μm (Fig 2C). Neither yad21 nor yad26 changed the overall expression level of GFP::UNC-9 (S1F Fig). Both yad21 and yad26 were linked to yadIs12 marker on Chromosome IV and had strong uncoordinated (unc) phenotypes. After testing some unc genes on chromosome IV, we found yad21 failed to complement the loss-of-function allele of unc-44(e362), and yad26 failed to complement the loss-of-function allele of unc-33(mn407). unc-44(e362) had the same UNC-9 punctate defects as that seen in yad21, and unc-33(mn407) and unc-33(e204) had identical phenotypes with yad26 (Fig 2A and 2B). Sequencing results showed that yad21 introduced a point mutation (P4813S) and a premature stop codon (Q6827*) in the neuronal specific long isoform of unc-44, and yad26 had two point mutations (E168K and G328D) in unc-33 (S2A and S2B Fig). These evidences supported the conclusion that yad21 was a loss-of-function allele of unc-44, and yad26 was a loss-of-function allele of unc-33.
unc-44 and unc-33 are homologs of ankyrins and CRMPs, respectively, and loss of function of them causes defects in axon development [36,39], raising a concern that the UNC-9 distribution defects could be an indirect result of misregulation of axon development. In majority of yad21 (83% n = 220 animals) and yad26 (94%, n = 175 animals) animals, PLM axons grew straight from cell bodies to the middle section of the body similar to wild type animals, but about 70% of yad21(lf) and yad26(lf) animals had either shorter or longer PLM axons (S3A Fig). However, we did not notice any correlation between the UNC-9 distribution defects and axon phenotypes. We quantified UNC-9 distribution defects in animals with normal axon length and found that the percentage of animals with UNC-9 distribution defects (yad21: 62% n = 71, yad26: 32.3% n = 62) was same as those seen in all animals. Loss of function of unc-34/Enabled/VASP affected PLM development. Loss-of-function mutation in rpm-1 caused overextension of PLM axons, but we did not observe any defects of UNC-9 distribution in these mutants (Fig 1D and S1E Fig) [41]. These results showed that UNC-9 distribution defects in unc-44(lf) and unc-33(lf) were not results of axon growth defects.
The UNC-44/UNC-33 pathway regulates neuronal polarity without changing the overall morphology of neurons [41]. There are three isoforms of unc-33 in C. elegans, named L (long), M (middle) and S (short) isoforms based on the length of cDNA, and only the long isoform could rescue unc-33(lf) polarity defects [41]. We found that expression of any of these three isoforms in PLM neurons rescued UNC-9 distribution phenotypes in unc-33(lf) animals, supporting the phenotypes we observed were not results of mis-regulation of neuronal polarity (Fig 2A and 2B). The successful rescue of unc-33(lf) phenotypes by expressing unc-33 cDNA in PLM neurons also showed that unc-33 cell autonomously regulated UNC-9 puncta (Fig 2A and 2B).
Using the UNC-9 specific antibody, we observed similar mis-accumulation of endogenous UNC-9 in unc-44 and unc-33 mutants (S3B Fig). We also noticed that UNC-1 was mis-localized (Fig 2D) and lost its co-localization with UNC-9 in unc-33(lf) and unc-44(lf) mutants (Fig 2E). Since the interaction between UNC-1 and UNC-9 was important for gap junction function [46], these results suggested that unc-44 and unc-33 might regulate gap junction functions. During C. elegans development, sensory neurons form NSY-5(INX-19) contained transient gap junctions, and most of these gap junctions are eliminated in adults [27]. In day one adults, about 13% of control animals did not have any NSY-5 puncta, and 87% of control animals had 30–40 NSY-5 puncta (Fig 2F). In unc-33(lf) and unc-44(lf) mutants, we observed significantly more NSY-5 puncta (50–90 punca/animal) in all examined animals (Fig 2F). These results support that UNC-44 and UNC-33 are involved in regulating multiple gap junction channels in different neuronal types.
To determine whether unc-44 and unc-33 could regulate each other, we carried out immunostaining experiments. As shown in Fig 3A, UNC-33(S) accumulated at the nerve ring of control animals, and unc-44 mutants induced more diffuse distribution of UNC-33(S). Using an antibody specifically recognizing the neuronal specific long isoform of UNC-44 [39], we found that UNC-44(L) was present in the nerve ring and neuronal processes, and unc-33 loss-of-function mutants did not affect UNC-44 distribution (Fig 3B). These results were consistent with a previous report [41] and supported the hypothesis that UNC-44 acted upstream and regulated UNC-33 localization. We further confirmed this conclusion by testing the suppression ability of overexpressing unc-33(S) on unc-44(lf) phenotypes. As shown in Fig 3C, expressing unc-33(S) at high level suppressed unc-44(lf) phenotypes. In conclusion, by analyzing mutants with abnormal accumulation of UNC-9 in PLM neurons, we uncovered an important role of the UNC-44/UNC-33 pathway in the regulation of gap junctions.
The striking phenotypes of unc-44 and unc-33 suggested that this pathway might regulate gap junction dynamics. To test this possibility, we analyzed GFP::UNC-9 movement at the gap junction zone one. In control animals, we observed bidirectional movement of UNC-9 both anterior and posterior to gap junctions, and each animal had almost equivalent numbers of UNC-9 particles moving toward and away from gap junctions (Fig 4A and 4B and S4 Movie), implicating the stable number of UNC-9 puncta at zone 1 was due to the balance of bidirectional movement. Loss-of-function mutation in unc-33 or unc-44 decreased the number of UNC-9 particles moving away from gap junctions and induced an imbalance of gap junction dynamics (Fig 4A and 4B, S5 and S6 Movies). These results suggested that unc-44(lf) and unc-33(lf) phenotypes might be due to suppression of gap junction turnover. To further examine this hypothesis, we used a transgene expressing photoactivatable GFP (PAGFP) tagged UNC-9 [55]. In this experiment, we first photoactivated PAGFP::UNC-9 in both cell bodies and axons, and we found that the fluorescence intensity at cell bodies decreased 35% at 3 hours after photoactivation, but the fluorescence intensity at gap junction zone 1 did not change in the same time period (Fig 4C and 4D). The decrease of fluorescence intensity at cell bodies was likely due to continuous transport of PAGFP::UNC-9 out of cell bodies. At gap junction zone 1, the stable fluorescence intensity indicated rapid turnover and replacement of PAGFP::UNC-9. In support of this conclusion, in animals that we locally photoactivated PAGFP::UNC-9 only at zone 1, we found that the fluorescence intensity decreased 43% after 3 hours in control (Fig 4E and 4H), and loss of function of unc-44 and unc-33 suppressed the decrease of fluorescence intensity (Fig 4F, 4G and 4H). We also tested whether unc-44 and unc-33 were involved in the assembly of UNC-9 puncta using the Fluorescence Recovery After Photobleaching (FRAP) assay. Briefly, we photobleached the GFP signal in zone 1 and measured the recovery of GFP signal after 3 hours. As shown in S4 Fig, loss of function of neither unc-44 nor unc-33 affected the recovery of GFP signal, supporting the UNC-44/UNC-33 pathway did not regulate the assembly of UNC-9 puncta. We believed that the UNC-44/UNC-33 pathway regulated gap junction turnover through suppression of transport of UNC-9 out of gap junctions.
UNC-33 homolog CRMP-2 has been shown to bind to the kinesin light chain subunit kinesin-1 in transport of neurotrophin receptors into growth cones [56,57]. The effect of mutating unc-44/unc-33 on UNC-9 dynamics suggested that they might regulate some motor proteins. We first tested three classic motor proteins, UNC-104, UNC-116 and DHC-1, that are known to be important for neuronal development. UNC-104 is the kinesin transporting synaptic vesicles from cell bodies to chemical synapses [58]. UNC-116 is a kinesin that is required for organization of presynaptic buttons and axonal mitochondria [59,60]. DHC-1 is the major dynein mediating retrograde transportation in C. elegans neurons [61]. However, we did not observe any UNC-9 defects caused by mutations in these three genes (S5 Fig). In testing the function of other motor proteins in UNC-9 regulation, we found that loss of function of a potential kinesin vab-8 induced similar phenotypes as those seen in unc-44(lf) and unc-33(lf) (Fig 5A, 5B and 5C). Further genetic analysis showed that double mutants of vab-8 unc-33 enhanced single mutant phenotypes to a degree similar to that seen in unc-44 mutants, indicating that VAB-8 might work together with UNC-33 downstream of UNC-44 in regulating gap junction turnover (Fig 5B and 5C). Indeed, using the local photoactivation assay, we confirmed that loss of function of vab-8 suppressed UNC-9 turnover (Fig 5D). vab-8 has two isoforms, vab-8(L) and vab-8(S), and the major difference between these two isoforms is that vab-8(S) lacks the kinesin domain present in vab-8(L) [62]. We found that only expression of vab-8(L) in PLM neurons was able to rescue vab-8(lf) phenotypes, suggesting the potential kinesin function of VAB-8 was essential for gap junction regulation (Fig 5B). Since UNC-33 homolog CRMP-2 has been shown to directly bind to kinesin, we examined whether UNC-33(S) could bind to VAB-8(L). As shown in Fig 5E, using a transgene pan-neurally expressing FLAG::VAB-8(L) and HA::UNC-33(S), we detected co-immunoprecipitation of UNC-33(S) and VAB-8(L), supporting that UNC-33(S) could bind to VAB-8(L) in vivo. This observation raised the possibility that UNC-33 might regulate the activity of VAB-8 in control of UNC-9 dynamics, such that, adding additional VAB-8L would restore UNC-9 distribution in unc-33 and unc-44 mutants. Indeed, we found that overexpression of VAB-8L in PLM neurons was able to suppress both unc-33(lf) and unc-44(lf) phenotypes (Fig 5F). These results showed that VAB-8L might function together with UNC-33 at the downstream of UNC-44 to regulate UNC-9 dynamics.
Using C. elegans PLM neurons as a model, we investigated the molecular mechanisms underlying gap junction regulation. In adult animals, PLM neurons form gap junctions at two regions, at zone 1 with PVC, LUA and PVR neurons and at zone 2 with BDU neurons [42–44]. Using transgenes expressing GFP tagged UNC-9, we were able to observe punctate structures that represented the localization of gap junctions in PLM neurons. In a genetic screen targeting isolation of mutants affecting gap junctions, we uncovered the important roles of C. elegans ankyrin/unc-44 and CRMP/unc-33 in regulating gap junctions. By imaging GFP::UNC-9 dynamics and PAGFP::UNC-9 decay, we further demonstrated that the UNC-44/UNC-33 pathway regulated UNC-9 dynamics and turnover. In searching for motor proteins contributing to unc-33/unc-44 phenotypes, we uncovered a role for a potential kinesin VAB-8 in UNC-9 turnover. VAB-8(L) directly bind to UNC-33(S), and overexpression of VAB-8(L) suppressed both unc-33(lf) and unc-44(lf) phenotypes, supporting the hypothesis that VAB-8 is the downstream target of the UNC-44 pathway in gap junction regulation. Double mutants of vab-8 and unc-33 had stronger phenotypes than each of single mutant and showed similar phenotypes as unc-44(lf), suggesting that UNC-33 and VAB-8L function in parallel downstream of UNC-44. The suppression of unc-33(lf) phenotypes by VAB-8L overexpression was likely due to the enhanced contribution from VAB-8L (Fig 6A). The other possibility is that VAB-8 is regulated by both UNC-33 and an unknown factor that is in parallel with UNC-33 and downstream of UNC-44 (Fig 6B).
CRMPs bind to tubulin heterodimers and microtubules to promote the assembly of microtubules during neuronal polarity and to regulate Numb-mediated endocytosis at growth cones [33–35,63,64]. In all these cases, the phosphorylation of CRMPs by Rho kinase or GSK-3β is important for its regulation [33,34]. However, UNC-33(S), the minimal rescue fragment of UNC-33, does not have those phosphorylation sites, suggesting other mechanisms may be involved in regulating UNC-33/CRMP. Besides its roles in regulating microtubules, CRMP-2 has been shown to bind with the kinesin light chain subunit, kinesin-1, in transport of neurotrophin receptors into growth cones [56,57]. In cultured neurons, CRMP-2 forms a complex with Slp1 and Rad27B to directly link TrkB to kinesin-1, and to mediate the antrograde transportation of TrKB receptor upon stimulation of BDNF [56]. CRMP-2 can also mediate interactions between kinesin-1 and Sra-1/WAVE1 complex to control axon growth [57]. In our study, we found UNC-33/CRMP could bind to VAB-8L/Kinesin to regulate the dynamic and turnover of gap junction channels. These results suggest that the interactions between CRMPs and kinesins may play multiple roles in control of membrane proteins.
In our experiments, we found that the localization of UNC-33(S) was regulated by unc-44. Together with the suppression of unc-44(lf) phenotypes by overexpressing unc-33(s), we concluded that unc-44 acted upstream of unc-33 in gap junction regulation. However, the localization of UNC-33(S) and UNC-44 appeared to be different, that UNC-33(S) was concentrated at the nerve ring, but UNC-44 had a broader distribution in neurons. One concern is that the transgene used to label UNC-33(S) (Prgef-1::FLAG:: unc-33(S)) may not reflect its in vivo localization. We compared our results of UNC-33(S) with the previous description of UNC-33(L) localization [41] and found that their localization were largely overlap, suggesting our transgene likely showed the in vivo localization of UNC-33(S). Since UNC-44 plays important roles in many aspects of neuronal development, the broad distribution of UNC-44 seems consistent with its multiple functions in neurons.
The UNC-44 homolog ankyrins have the ability to bind to channels and to integrate them to cytoskeleton, and the interactions between ankyrins and channels are dynamically regulated by intracellular signals [65]. It is possible that UNC-44 may play a role in holding UNC-9 at gap junctions. During gap junction turnover, the conformational change of UNC-44 could recruit UNC-33/CRMP and VAB-8L to UNC-9, and the activated UNC-33 could promote the transport of UNC-9 away from gap junctions through VAB-8L or an unknown factor.
We maintained C. elegans strains on NGM plates at 20–22.5°C. All transgenes and strains are described in the S1 Table. We use yadIs12 (Pmec-4::GFP::UNC-9) to visualize gap junctions in PLM neurons. Using yadIs12 as a starting strain, we performed a clonal recessive screen following standard ethyl methane sulfonate (EMS) mutagenesis protocol. In 1000 mutagenized haploid genomes we examined, 12 mutants were isolated with three type phenotypes as mentioned in the manuscript.
All DNA expression constructs were made using Gateway cloning technology (Invitrogen). Sequences of the final clones were confirmed. The S1 Table lists the genotypes and DNA constructs for the transgenes. unc-33 (L) cDNA was obtained from Dr. Cori. Bargmann, and unc-33(M) and unc-33(S) was amplified from unc-33(L) cDNA. vab-8(L) and vab-8(S) cDNA was amplified from yk clones from Dr. Yuji Kohara lab. Photoactivatable GFP (PAGFP) plasmid was obtained from Addgene. PAGFP coding sequence was inserted in an ASCI enzyme site before the start codon of UNC-9. All primer sequences are available upon request. Transgenic animals were generated following standard procedures. In general, plasmid DNAs of interest were used at 1–50 ng/ml with the co-injection marker Pttx-3::rfp/Pttx-3::gfp at 50 ng/μl.
Representative images and immunostaining results were acquired with a Zeiss LSM700 confocal microscope using a Plan-Apochroma 40x/1.4 objective. Worms were immobilized in 1% 1-phenoxy-2-propanol (TCI America, Potland, OR) in M9 buffer. For quantification of the percentage of animals with gap junction defects in PLM neurons, we used a Zeiss Axion Imager 2 microscope equipped with Chroma HQ filters. Each analyzed data performed takes at least three independent experiments and total 200–300 1-day old adults. For quantification of the length of gap junction zone, Images were acquired with LSM700 confocal microscope using a Plan-Apochroma 40x/1.4 objective, and the length of gap junction zone was measured using Zeiss Zen Black software.
The photoactive experiments were carried out with a Zeiss LSM700 confocal microscope using a Plan-Apochroma 40x/1.4 objective. 1-day young adult transgenes expressing mCherry and PAGFP::UNC-9 in PLM neurons were immobilized in 1% 1-phenoxy-2-propanol. We first use mCherry signal to localize PLM neurons and use 405 laser to globally or locally photoactivate PAGFP::UNC-9. After photoactive animals were recovered on the NGM plates for 3 hours before the second images were taken. The fluorescent intensity was analyzed using Image J software.
GFP::UNC-9 dynamic experiments were preformed using an Andor revolution microscopy with a 60 x /1.46 Plan-Apochromat objective controlled by MetaMorph software. All videos were acquired by an Andor EM-CCD camera (DU897). 1-day adult animals were immobilized in 5mM levamisole and on 5% agar pads for imaging. Videos for GFP::UNC-9 dynamic analysis were roughly 30–40s with 8 frames per second. Kymographs were generated using ImageJ, and the direction of GFP::UNC-9 movement was judged by the direction of the black lines in the kymograph pictures. In general, dynamic puncta were defined as their velocities >0.1 μm/s for last at least 5s. For puncta that change directions during experiments, we trace them for the overall direction.
1-day adult animals were immobilized in 5mM levamisole and on 5% agar pads. FRAP was performed using Zeiss LSM700 confocal microscope using a Plan-Apochromat 40x/1.4 objective. A high-powered laser (at 100% energy, 488 nm) was used to photobleach the region of interest. Worms were recovered on NGM plates with food for 3 hours before taking the second image. Quantification of GFP::UNC-9 was carried out using Image J. Percentage Recovery = (I3h-Ibleach)/(Ipreblach -Ibleach)x100. I3h: the intensity at the region of interest (ROI) three hours after photobleaching; Ibleach: the intensity at the ROI after photobleaching; Iprebleach: the intensity in the ROI before photobleaching. Background (the intensity in the non-bleached part of ROI) was subtracted, respectively.
All immunostaining experiments were carried out following the standard protocol using 1-day young adults. The Rabbit anti-UNC-9 antibody was a gift from Dr. Zhao-Wen Wang. The Rabbit anti-UNC-44(L) antibody was a gift from Dr. Anthony Otsuka. Mouse anti-FLAG M2 antibody (Cat# F1804), mouse anti- acetylated tubulin antibody (Cat# T7451) and rabbit anti-GFP (Cat# G1544) were purchased from Sigma. The dilutions for each antibodies are: anti-UNC-9(1:100), anti-UNC-44(1:100); anti-FLAG (1:300), anti-acetylated tubulin(1:300) and anti-GFP (1:150). Alexa Fluor 488 Donkey-anti-rabbit IgG (H+L) antibody (Cat# A-11008) and Alexa Fluor 594 Goat Anti-Mouse IgG (H+L) Antibody (Cat# A-11005) from Molecular Probes were used as secondary antibody in 1:500 dilution.
For the immunoprecipitation experiment, we generated an transgene (yadEx421) expressing FLAG-VAB-8(L) and HA-UNC-33(S) under the control of Pan-neuronal promoter Prgef-1. Proteins from mixed stages animals were first extracted using RIPA buffer by frozen-throw about 50 times in ethanol with dry ice, and protein lysis was incubated with mouse anti-FLAG M2 antibody (Cat# F1804) in room temperature for 5 hours and then precipitated using Protein A/G PLUS-Agarose (Santa Cruz Bio. esc-2003). Heated Protein samples were separated using SDS-PAGE Gradient Gels (4–20%), and then transferred to nitrocellulose. Blots were probed with mouse anti-Flag antibodies (sigma, F1804) and rabbit anti-HA(Sigma H6908), and then visualized with Amerisham HRP-conjugated anti-rabbit secondary antibodied at 1:5000 using the SuperSignal West Femto kit (Pierce, Rockford, 1L).
We analyzed our data using one-tailed Student’s t test, one way ANOVA or Fisher exact test in Graphpad Prism (GraphPad Software, La Jolla, CA).
|
10.1371/journal.pmed.1002512 | Prevalence of sexually transmitted infections among young people in South Africa: A nested survey in a health and demographic surveillance site | Sexually transmitted infections (STIs) and bacterial vaginosis (BV) are associated with increased transmission of HIV, and poor reproductive and sexual health. The burden of STIs/BV among young people is unknown in many high HIV prevalence settings. We conducted an acceptability, feasibility, and prevalence study of home-based sampling for STIs/BV among young men and women aged 15–24 years old in a health and demographic surveillance site (HDSS) in rural KwaZulu-Natal, South Africa.
A total of 1,342 young people, stratified by age (15–19 and 20–24 years) and sex were selected from the HDSS sampling frame; 1,171/1,342 (87%) individuals had ≥1 attempted home visit between 4 October 2016 and 31 January 2017, of whom 790 (67%) were successfully contacted. Among the 645 who were contacted and eligible, 447 (69%) enrolled. Consenting/assenting participants were interviewed, and blood, self-collected urine (men), and vaginal swabs (women) were tested for herpes simplex virus type 2 (HSV-2), chlamydia, gonorrhoea, syphilis, trichomoniasis, and BV. Both men and women reported that sample collection was easy. Participants disagreed that sampling was painful; more than half of the participants disagreed that they felt anxious or embarrassed. The weighted prevalence of STIs/BV among men and women, respectively, was 5.3% and 11.2% for chlamydia, 1.5% and 1.8% for gonorrhoea, 0% and 0.4% for active syphilis, 0.6% and 4.6% for trichomoniasis, 16.8% and 28.7% for HSV-2, and 42.1% for BV (women only). Of the women with ≥1 curable STI, 75% reported no symptoms. Factors associated with STIs/BV included having older age, being female, and not being in school or working. Among those who participated in the 2016 HIV serosurvey, the prevalence of HIV was 5.6% among men and 19% among women. Feasibility was impacted by the short study duration and the difficulty finding men at home.
A high prevalence of STIs/BV was found in this rural setting with high HIV prevalence in South Africa. Most STIs and HIV infections were asymptomatic and would not have been identified or treated under national syndromic management guidelines. A nested STI/BV survey within a HDSS proved acceptable and feasible. This is a proof of concept for population-based STI surveillance in low- and middle-income countries that could be utilised in the evaluation of STI/HIV prevention and control programmes.
| Adolescents and young adults are particularly vulnerable to sexually transmitted infections (STIs).
The first strategic direction of the WHO Global Health Sector Strategy on Sexually Transmitted Infections 2016–2021 is to collect information on STI prevalence and incidence across representative populations.
There is evidence that bacterial vaginosis (BV) is a risk factor for poor birth outcomes and STIs including HIV. The collection of BV prevalence may therefore also be important.
Developing new cohorts for dedicated STI/BV prevalence studies may not be realistic, particularly in sub-Saharan Africa, where the impact of STIs/BV and their consequences may be greatest.
Nesting STI/BV surveys within networks of health and demographic surveillance sites (HDSSs) would be an efficient way of providing data to better understand STI epidemiology among adolescents and young people in high HIV prevalence settings.
We carried out a nested STI/BV survey among 1,342 adolescent and young people in an HDSS in KwaZulu-Natal, South Africa, between October 2016 and January 2017.
Potential participants were contacted at home and invited to participate.
Participants were interviewed, and samples were collected for STI/BV testing.
We showed that this study was feasible within the 3.5-month time period: 1,171/1,342 (87%) individuals had ≥1 attempted home visit, of whom 790 (67%) were successfully contacted.
The study was acceptable: among those contacted and eligible, 447/645 (69%) enrolled. Both men and women reported few problems with sample collection.
We report a high burden of STIs/BV in this population, particularly of chlamydia (5% in men and 11% in women), herpes simplex virus type 2 (17% in men and 29% in women), and BV (42% in women).
Nested STI/BV surveys in HDSSs can be feasible and acceptable; however, more survey time is needed to ensure that all potential participants are visited and contacted.
These studies should be carried out in conjunction with studies to measure STI/BV prevalence in high-risk populations (e.g., female sex workers) to provide robust prevalence estimates.
These data are essential to advocate, fund, plan, implement, and evaluate interventions for STI prevention and control among adolescents and young people.
Strategies for the prevention and control of chlamydia, herpes simplex virus type 2, and BV are needed in this population.
| In 2012, 286 million people aged 12–24 years lived in Africa, accounting for 18% of the global youth population. By 2040, the number of young people in Africa is projected to increase by 60% to 466 million [1]. Health interventions targeted at this age group are important for current and future adult health and for the health of the next generation. This is particularly true for sexually transmitted infections (STIs), which, when acquired in adolescence, can jeopardise sexual and reproductive health later in life and, for women, the health of their babies. In low- and middle-income countries (LMICs), symptomatic STIs are treated by syndromic management (presumptive treatment for symptomatic people without the use of laboratory tests) [2], but most STIs are asymptomatic and go unnoticed and untreated. Both symptomatic and asymptomatic STIs can cause serious morbidity, including pregnancy complications, cancer, infertility, and enhanced HIV transmission. Many of these sequelae are preventable if STI testing and treatment is implemented. Moreover, there is growing evidence that the common reproductive tract condition bacterial vaginosis (BV) is an independent risk factor for HIV [3,4], and BV-associated microbiota may decrease the efficacy of topical microbicides [5].
High STI prevalence among young people has been observed worldwide and highlights the critical need for global efforts to improve sexual and reproductive health in this population. In an individual participant data meta-analysis of 18 HIV prevention studies among women in sub-Saharan Africa, STI prevalence was higher among young women aged 15–24 years than among older women for all STIs except herpes simplex virus type 2 (HSV-2) [6]; in this age group, the estimated range of prevalence of STIs in South Africa among clinic/community populations was 8.0% to 20.6% for chlamydia, 1.4% to 8.9% for gonorrhoea, 3.1% to 20.0% for trichomoniasis, 31.9% to 53.7% for HSV-2, and 35.8% to 52.4% for BV. In addition, viral STIs such as HSV-2 and human papillomavirus (HPV) infection are often acquired soon after sexual debut, which usually occurs in adolescence, and both are common among young people in sub-Saharan Africa [7–10]. However, many of the studies yielding these results are conducted in urban areas and/or clinical cohorts of adolescents and young adults known to be at high risk of infection. To date, there have been few population estimates of the burden of STIs among adolescent girls and young women and no studies among men [6].
The WHO Global Health Sector Strategy on Sexually Transmitted Infections 2016–2021 has outlined the goals and targets for global STI prevention and control. The first strategic direction is to collect information on STI prevalence and incidence across representative populations [11]. Understanding regional and national STI epidemics is essential to advocate, fund, plan, and implement interventions for STI prevention and control. The strategy also urges LMICs to move from syndromic to aetiologic surveillance of STIs, and to conduct routine surveillance in key populations most at risk for STIs including adolescents. Yet, in resource-limited settings, developing new cohorts for dedicated STI prevalence studies may not be realistic, particularly in sub-Saharan Africa, where the impact of STIs and their consequences may be greatest.
Networks of health and demographic surveillance sites (HDSSs) conducting longitudinal population-based research such as the International Network for the Demographic Evaluation of Populations and their Health (INDEPTH Network) may provide opportunities to obtain representative STI/BV prevalence estimates for adolescents and young people and facilitate community entry and engagement with sensitive topics such as sexual health [12]. However, population-based surveys can be challenging to conduct. Key requirements include the acceptability of being approached at home and home sampling, the feasibility of finding young people at home and a parent available to consent, the receipt of results while maintaining confidentiality, and establishing clinical pathways for the treatment of cases. We conducted a study in the Africa Health Research Institute (AHRI; formerly the Africa Centre for Health and Population Studies) HDSS, a member of the INDEPTH Network, to investigate the acceptability and feasibility of home-based sampling of STIs/BV among young people aged 15–24 years, and to measure prevalence and factors associated with STIs/BV. The background 2011 HIV prevalence in women aged 15–19 years and 20–24 years was 14.7% and 26.5%, respectively, and in men aged 15–19 years and 20–24 years was 7.0% and 10.2%, respectively [13].
This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines (S1 STROBE Checklist) [14].
The AHRI HDSS is located in the rural uMkhanyakude district of KwaZulu-Natal, covering an area of 438 km2, with a 2016 population of approximately 100,000 people who are members of 12,000 households [15]. Since 2000, annual household-based surveys have been used to collect information on births, deaths, and migration patterns from all household members, including non-residents. In addition, resident household members aged ≥15 years are invited to participate in an annual HIV serosurvey, and to complete a questionnaire on general health and sexual behaviour.
For the STI survey, young men and women who were resident in the HDSS, based on the data collected in the routine household surveillance, and aged 15–24 years as of 19 July 2016 were eligible for inclusion. A random sample of 1,342 young people was selected to obtain a target sample size of 800, allowing for 40% non-contact/refusals. This sample size would have provided acceptable precision for estimating the prevalence of an STI with a prevalence as low as 1.5%. Sampling was stratified by age group (15–19 years and 20–24 years) and sex. The HDSS is divided into 14 subareas; within each stratum, a fixed proportion was sampled from each subarea to reflect the population distribution across the HDSS.
The University of KwaZulu-Natal Biomedical Research Ethics Committee, the London School of Hygiene & Tropical Medicine Research Ethics Committee, the Southampton General Hospital Faculty of Medicine Ethics Committee, Hlabisa District Hospital, and the AHRI Somkhele Community Advisory Board approved the study protocol. The STI survey was called Ukuvikela impilo yetho yokuzalana eyigugu, isiZulu for ‘protecting our precious reproductive health’. The AHRI Community Engagement Team disseminated information about the study in community dialogues and road shows. Potential participants were contacted at home and invited to participate. Written parental consent was required for participants <18 years old, with participant written assent. Participants aged 18 years or older proved written consent. Participants consented separately for each sample type (vaginal swab [women only], urine [men only], and blood); participants who did not consent for a sample could still enrol in the study. Participants were asked for permission to link their STI survey data with the data collected in the annual routine household and individual surveillance.
The study team consisted of 2 field workers (1 male and 1 female), 2 female licensed practical nurses, 1 male licensed practical nurse, and 1 male registered nurse team leader, with an intention to match a same-sex nurse to participants whenever possible. The field work was conducted Tuesday to Saturday from 11 AM to 7 PM to maximise the chances of finding participants at home.
After informed consent/assent, the participant had a short computer-assisted personal interview by the study nurse [16]. The interview obtained data on demographics, substance use, sexual behaviour, violence, circumcision (men only), family planning (women only), genital hygiene, and genital symptoms. For questions about sexual behaviour and violence, the participant was asked to self-interview using a tablet device; however, the study nurse was available to support the participant if needed. If a participant reported genital complaints, they were referred to our study nurse in a local primary health clinic for syndromic management as per 2015 South African STI management guidelines [17].
All participants had 8.5 ml of blood drawn for syphilis and HSV-2 testing. For women, the research nurse explained the procedure to self-collect a total of 5 vaginal swabs for testing for chlamydia, gonorrhoea, trichomoniasis, and BV (an additional swab was collected for storage). Swab collection took place in a private setting identified by the participant. Men collected a urine sample for testing for chlamydia, gonorrhoea, and trichomoniasis.
After the sample collection, participants were asked to rate their agreement with 10 statements using a visual analogue scale (VAS) ranging from 0 (easy/agree) to 100 (difficult/disagree) to assess the ease of understanding of consent for the study, the instructions for collecting the sample, and the participant’s experience of participation.
All participants were asked to provide contact information for test results, including their preferred mode of contact for both positive and negative results (e.g., telephone call, SMS message, or WhatsApp message), and ideal hours for contact. We attempted to contact all participants with the results for laboratory-diagnosed curable STIs (chlamydia, gonorrhoea, trichomoniasis, and syphilis). All participants with mobile phones were given 5 South African rand (US$0.37) of air time to contact the study nurses with questions if needed. Participants who had a positive test for a curable STI were referred for free treatment; reimbursement for travel was provided. We traced all cases who were not contactable or did not come to clinic for treatment. We used British Association for Sexual Health and HIV guidelines for the treatment of laboratory-diagnosed chlamydia, gonorrhoea, and trichomoniasis [18–20], and South African STI management guidelines for the treatment of syphilis [17].
Laboratory testing was performed according to manufacturers’ instructions and standard operating procedures in the central AHRI laboratory and Global Clinical and Viral Laboratory in Durban, South Africa. Serum samples were used to test for IgG antibodies for HSV-2 by a type-specific ELISA (Kalon Biological, Guildford, UK). Syphilis infection was determined by the Determine Syphilis TP rapid test (Alere, Waltham, MA, US) in the central AHRI laboratory. All positives were confirmed at the Global Clinical and Viral Laboratory with Treponema pallidum haemagglutination (TPHA) (Randox Laboratories, Crumlin, UK) and tested with the Venereal Disease Research Laboratory (VDRL) test (Omega Diagnostics, Alva, UK) using a reverse algorithm as per South African STI management guidelines [17] due to the young age of participants (i.e., unlikely to have treated past infections). Syphilis infection was defined as follows: negative, TPHA−/VDRL−; early or previously treated infection, TPHA+/VDRL−; and active syphilis, TPHA+/VDRL+ low titre [<1:8] or TPHA+/VDRL+ high titre [≥1:8].
Vaginal swabs were used to prepare a slide at the home and air dried. Slides were transported to the central AHRI laboratory, methanol-affixed, Gram stained, and examined for BV using the Nugent score [21]. A Nugent score of 0–3 indicated normal microbiota, 4–6 indicated intermediate microbiota, and 7–10 indicated BV. Vaginal swabs (women) and urine (men) were sent to Global Clinical and Viral Laboratory for testing by real-time PCR for Neisseria gonorrhoeae, Chlamydia trachomatis, and Trichomonas vaginalis. Detection was carried out using the Lightmix Kit Neisseria gonorrhoeae, the Lightmix Kit Chlamydia trachomatis, and the Lightmix Kit Trichomonas vaginalis (TIB MOLBIOL, Berlin, Germany) following the manufacturer’s instructions. All positive tests for N. gonorrhoeae were confirmed using GeneXpert (Cepheid, Sunnyvale, CA, US). The confirmation test should have a higher specificity than the first test; the GeneXpert N. gonorrhoeae detection probe has 2 primer sets that increase the specificity needed for the N. gonorrhoeae confirmation [22,23]. External quality controls were carried out quarterly for real-time PCR with the College of American Pathologists.
Data were captured electronically using REDCap software [24]. Range and consistency checks were done automatically during data capture; further data cleaning and analysis was done using Stata 14 (College Station, TX, US). All questions required a response to minimise missing data, although participants could reply ‘don’t know’ or ‘prefer not to say’.
The statistical analysis plan was prepared prior to the statistical analysis (S1 Analysis). Changes in response to peer review of this paper included the inclusion of other STIs in the BV risk factor analysis, and the inclusion of transactional sex in each risk factor analysis. Continuous variables were summarised using means and standard deviations or medians and interquartile ranges; categorical data were summarised using frequency counts and percentages. Missing data were not imputed.
The acceptability and feasibility of our survey were measured by the following outcomes: proportion of participants who were selected and contactable, the proportion of those contacted who agreed to participate, the proportion who agreed to each sample collection (e.g., blood, vaginal swabs, and urine), median and interquartile range of responses to a VAS measuring acceptability post-sampling, and proportion of cases who presented for treatment. We also estimated STI/BV prevalence and explored factors associated with any curable STI (chlamydia, gonorrhoea, syphilis, and trichomoniasis), HSV-2, and BV.
The number of individuals who were successfully contacted and who consented to participate were tabulated by sex, age group, residence location (urban/peri-urban/rural), household socioeconomic status, education level, and HIV status using linked data from the HDSS. Characteristics of individuals who participated and the remainder in the eligibility list were compared using chi-squared tests.
The prevalence estimate of each STI or BV, and its 95% confidence interval, was calculated overall and by sex; prevalence estimates were weighted to account for the stratified sample design and non-response, calculated as the inverse probability of study participation in strata defined by age group, sex, and residence location (urban/peri-urban/rural). We compared these results to unweighted prevalence and prevalence weighted for the stratified sample design only.
Logistic regression was used to estimate odds ratios and 95% CIs for factors associated with the presence of any curable STI (chlamydia, gonorrhoea, syphilis, or trichomoniasis), of HSV-2, and of BV; separate models were developed for each outcome. Potential factors associated with curable STIs, HSV-2, and BV were examined using a conceptual framework with 3 levels: sociodemographic factors, modifiable behavioural factors (including genital hygiene), and sexual behaviour and violence. For each outcome, age and sex (except for BV, which was in women only) were considered a priori confounders and were included in all models. Sociodemographic factors whose age- and sex-adjusted associations with the outcome were significant at P < 0.10 were included in a multivariable model; those remaining associated at P < 0.10 were retained in a core model. Behavioural factors were then added to this core model one by one; those that were associated with the outcome at P < 0.10, after adjusting for sociodemographic factors, were included in a multivariable model and retained if they remained associated at P < 0.10. Associations with sexual behavioural and violence factors were subsequently determined in a similar way. Many of the questions about sexual relationships were asked only if participants reported having ever had sex, so analyses of these variables were restricted to that subgroup.
The field work took place from 4 October 2016 to 31 January 2017. Due to unexpected time limitations, only 1 visit attempt per selected individual was carried out from November to January to attempt coverage in subareas (a total of 14 subareas), but not all selected young people were visited. Among the 1,342 individuals selected, 1,171 (87%) had ≥1 attempted home visit, of whom 781 (67%) were successfully contacted (Fig 1). Of those who were contacted, 645 (83%) were still eligible. Among those contacted and eligible, 447 (69%) enrolled. Individuals aged 20–24 years were less likely to be contacted than those aged 15–19 years (63% versus 70% of those with an attempted visit, P = 0.01) and less likely to be eligible after contact was made (mostly due to migration). Men were less likely to be contacted than women (58% versus 75%, P < 0.001). Overall, there was strong evidence that individuals who were sampled but did not enrol were more likely to be older, male, and from rural or urban areas, and to have completed secondary education or above, compared with those who enrolled (S1 Table).
Of those enrolled, 96% of women provided all vaginal swabs and 93% provided blood samples; all men provided urine samples and 98% provided blood samples. Both men and women reported that it was easy to understand how to collect urine and vaginal swabs, respectively (Fig 2A). Participants agreed they felt comfortable, in control, relaxed, and confident of their ability to collect the sample correctly (Fig 2B). Participants disagreed that sampling was painful. Most men disagreed that they felt anxious or embarrassed, and over half of women disagreed that they were anxious or embarrassed (Fig 2C).
Of those who provided samples, 206/245 (84%) of individuals aged 15–19 years and 184/192 (96%) of individuals aged 20–24 years enrolled had access to a telephone to receive results. Of those, the majority preferred a telephone call for both positive and negative results (59% and 57%, respectively), followed by an SMS message (37% and 39%, respectively). Few chose to receive their results (positive or negative) by WhatsApp message (4% and 4%, respectively). These results were similar by sex and age, although a higher proportion of males than females preferred to receive their results by telephone (S2 Table).
Fifty-five participants had ≥1 curable STI and were invited to the clinic for management: 52 (95%) came on their own, and 3 had to be traced.
Most participants were currently enrolled in school (Table 1). Few participants were working (11% of men and 5% of women). Proportionally, more men aged 20–24 years reported having ever smoked a cigarette compared to women of the same age (17% versus 8%, respectively). Conversely, proportionally fewer men compared to women reported having ever had at least 1 drink of alcohol (23% versus 51%, respectively). A small proportion of participants reported cannabis use: 8% among men aged 20–24 years and 4% among women of the same age. Few participants reported using other drugs (1%).
In all, 51% of men reported circumcision: younger men were more likely to be circumcised (Table 1). Sixteen percent of women reported using intravaginal cleansing; older women were more likely to report intravaginal cleansing. In all, 23% of women aged 15–19 years and 50% of women aged 20–24 years reported use of any type of contraception.
In all, 61 (33%) men and 155 (63%) women reported having had sexual intercourse; of these, the median (IQR) number of lifetime partners was 4 (2–6) for men and 3 (2–3) for women. A larger proportion of men than women reported having used a condom at last intercourse (58% versus 41%, respectively). A smaller proportion of men than women reported knowing their last partner’s HIV status (37% versus 55%, respectively), and a smaller proportion of men than women reported discussing their own HIV status with their last partner (48% of men versus 63% of women).
More women aged 20–24 years reported providing (32%) or receiving oral sex (46%) than men of the same age group (12% and 13%, respectively). Few participants reported ever having anal sex, and fewer men than women (1% versus 5%). Among men who participated in the 2016 HIV serosurvey, the prevalence of laboratory-diagnosed HIV among those aged 15–19 years was 4% and among those aged 20–24 years was 19%. Among women, the prevalence of laboratory-diagnosed HIV among those aged 15–19 years was 9% and among those aged 20–24 years was 30%.
Weighted prevalence from Table 2 shows a high prevalence of chlamydia in men aged 20–24 years (12.6%; 95% CI 6.4%–23.3%) and women in both age groups (15–19 years: 11.7%; 95% CI 6.8%–19.3%; 20–24 years: 10.2%; 95% CI 6.0%–16.9%). The prevalence of gonorrhoea was low, from 0 cases among men aged 20–24 years to 3.2% (95% CI 1.2%–8.2%) in women of the same age group. There was 1 case of active syphilis—the overall prevalence of active syphilis was 0.1%. There were 5 TPHA−/VDRL+ samples. The prevalence of trichomoniasis was lower in men compared with women (0.6% [95% CI 0.1%–4.0%] versus 4.6% [95% CI 2.6%–7.9%]); the highest prevalence was among women aged 20–24 years. In all, 14% of individuals had a curable STI (chlamydia, gonorrhoea, syphilis, or trichomoniasis). Of these, 75% reported no symptoms. The prevalence of HSV-2 was lower in men compared with women (16.8% [95% CI 11.3%–24.1%] versus 28.7% [95% CI 23.3%–34.7%]), with the highest prevalence among women aged 20–24 years. The prevalence of BV was 41.1% (95% CI 32.3%–50.5%) among women aged 15–19 years and 44.2% (95% CI 35.5%–53.2%) among women aged 20–24 years. Prevalence weighted for sampling and non-response (Table 2) was similar to unweighted prevalence and prevalence using sampling weights only (S3 Table).
In the adjusted analysis of factors associated with curable STIs (chlamydia, gonorrhoea, syphilis, and trichomoniasis), participants aged 20–24 years and women had more than twice the odds of having a curable STI compared to participants aged 15–19 years and men, respectively (Table 3). Having a higher number of lifetime sexual partners was associated with having a curable STI (P = 0.038). Reporting having had sexual intercourse was strongly associated with having a curable STI.
In the adjusted analysis of factors associated with HSV-2, participants aged 20–24 years and women had twice the odds of HSV-2 infection compared to participants aged 15–19 years and men, respectively (Table 4). Participants currently enrolled in school or working had less than half the odds of HSV-2 infection compared to those who were neither in school nor working.
In the adjusted analysis of factors associated with BV, there was weak evidence that being currently enrolled in school or working was associated with a diagnosis of BV (Table 5). Those having ever drunk alcohol had twice the odds of a diagnosis of BV, and there was weak evidence that having ever smoked a cigarette was associated with a diagnosis of BV. Independently, those reporting genital touching and having ever had sex had twice the odds of a diagnosis of BV. Participants who were HSV-2 seropositive had 4 times the odds of a diagnosis of BV.
In the subgroup analysis among participants who reported having had sex, there was some evidence that discussing the last partner’s HIV status was associated with not having a curable STI (adjusted OR 0.48; 95% CI 0.23–1.00; S4 Table). There was no evidence that factors included in this subgroup analysis were associated with either HSV-2 infection or diagnosis of BV (S5 and S6 Tables).
We conducted a nested STI survey among young people aged 15 to 24 years in a rural HDSS in KwaZulu-Natal, and found it to be feasible and acceptable. The HDSS provided infrastructure and a sampling frame to carry out a population-based cross-sectional study of STI/BV prevalence. There was a high burden of STIs/BV in this high HIV prevalence setting. Most of the infections were asymptomatic and would not have been identified or treated using national syndromic management guidelines.
This study is a proof of concept that STI surveys can be successfully conducted within HDSS networks such as the INDEPTH Network [12], the Network for Analysing Longitudinal Population-based HIV/AIDS data on Africa (ALPHA Network) [25], the Department of Science and Technology (DST), and the South African Medical Research Council (SAMRC) South African Population Research Infrastructure Network (SAPRIN) [26] (Fig 3). STI surveys can be conducted within the infrastructure of HDSSs with 2 important advantages. First, STI surveys can be carried out in LMICs intermittently to contribute to estimates of the global burden of STIs and to evaluate local implementation of global STI control programmes at a population level. Second, STI surveys can be carried out more frequently in settings with high HIV/STI prevalence to monitor and evaluate enhanced STI/HIV control programmes. HDSS networks could provide a strategic platform to strengthen STI surveillance and control in LMICs, especially in sub-Saharan Africa, where HIV and STI/BV prevalence are high. Importantly, while population-based data are crucial for an effective STI prevention and control programme, these data must be complemented by robust data from high-risk groups (e.g., female sex workers) to account for STI transmission dynamics that depend on high rates of partner change [27].
The prevalence of chlamydia was high in this STI survey among women of both age groups and among men aged 20–24 years. Several studies report high prevalence of chlamydia in South Africa [6,28–31]; both Microbicide Trials Unit (MTN)–003 (VOICE) and HIV Prevention Trials Network (HPTN) 055 showed higher baseline chlamydia prevalence and incidence among women in South Africa compared with other sites in the multi-site studies. Sub-regional or national differences in STI epidemics among young people could be further elucidated in STI surveys in a network of HDSSs. Aetiological diagnosis of STIs is unaffordable and inaccessible for most LMICs. Rapid, accurate, and affordable point-of-care tests might bridge this gap in future [32]. The development of tools such as these must be carried out in parallel with population-based STI surveys and analyses of risk factors.
Although HSV-2 and BV are not curable STIs, better control tools are needed for them, and we recommend continued integration of HSV-2 and BV in STI prevalence surveys. The prevalence of HSV-2 in our study was almost twice as high for young women as for young men, and almost 50% in women aged 20–24 years. Rapid acquisition of HSV-2 after sexual debut has been reported in several studies [8,10], suggesting that HSV-2 seropositivity could be used as a biological proxy for sexual activity. Over 40% of women in this study had BV, consistent with other studies in sub-Saharan Africa [32]. Factors associated with BV in our study (sexual debut, currently having more than 1 sex partner, and HSV-2 infection) are consistent with the literature [32]. Despite BV not being considered a traditional STI, there is an accumulating body of evidence suggesting that sexual transmission is an integral part of its pathogenesis [32]. In addition, BV is associated with serious sequelae, including preterm delivery and increased risk of STI and HIV acquisition and transmission of HIV [3,33–37].
Population-based demographic and behavioural data are also important for planning and evaluating STI prevention and control programmes [38]. In this HIV hyperendemic setting, it is reassuring that there was a higher prevalence of self-reported circumcision among the younger men than among the older men—suggesting the population impact of male medical circumcision programmes. However, the extremely low self-reported condom use at last sex is a tremendous concern. In addition, few participants knew their last partner’s HIV status. In this STI survey, current enrolment in school or working was protective for HSV-2 and BV. These data mirror findings from the AHRI HDSS, which showed that out-of-school youth reported earlier sexual debut and more high-risk sex than in-school youth [39], suggesting that interventions to keep adolescents in school may be just as relevant for other STIs as they are for HIV [40,41].
Strengths of this study include a high rate of acceptability for participation and sample collection, the success in treating those with a curable STI, and the use of a population-based platform as a sampling frame. There are several challenges for carrying out home-based studies, including contacting young people during school hours and the provision of confidential results to participants; however, we maximised contact by modifying the field work hours from 11:00 to 19:00 from Tuesday to Saturday, and provided participants with a choice of mode for receiving results. Once contacted, enrolment into a population-based study of STI/BV testing was acceptable among young people, as was the home-based collection of samples, including the self-collection of genital samples. An additional strength of this study is that it was conducted in an area with persistently high HIV incidence and prevalence. Results of this study could help to inform co‐strategies to address both HIV and STIs that synergise the transmission of HIV.
This study was not without limitations. The sample collection period was limited to 3.5 months by the start of the next HDSS surveillance round, and we did not reach our target of 800 young people. The smaller sample size of 447 provided less precision for prevalence estimates and less power to investigate factors associated with STIs/BV. In addition, the overall coverage in the survey was low, increasing the potential for selection bias. It was challenging to find young men aged 20–24 years at home. HPTN 017 (PopART), a cluster-randomised controlled trial offering home-based HIV counselling and testing in South Africa and Zambia, also reported that young men (32.7%) more often than young women (20.2%) were not at home at the time of visits [42]. Furthermore, many young people were not at home due to migration. The AHRI individual surveys of residents aged 17–49 years indicate that approximately one-fifth of men and women in any survey round have migrated at least once in the last 2 years, and persons with a recent migration history have a higher risk of HIV infection [43]; thus, those with a recent migration history are likely to have a different risk profile. The AHRI HDSS was established in a highly mobile population with a severe HIV epidemic, in which characterisation of migration and mobility was central to its conceptual and data model [44]. Indeed, nesting STI surveys in HDSSs may offer another advantage over one-off de novo STI prevalence surveys: the HDSS sampling frame has information about those who are not enrolled into the study. Additionally, while a one-time survey will miss some of those who have migrated; annual repeat cross-sectional surveys ensure that most age-eligible household members contribute data over time. Reassuringly, STI/BV prevalence weighted for both sampling and non-response data was very similar to the unweighted prevalence or prevalence weighted for sampling only.
Another limitation was that there was evidence of underreporting of sexual behaviours: 6% of participants with a curable STI and 15% of participants with HSV-2 reported never having had sex. Underreporting of sexual behaviour is common, especially among adolescents [45]. We used a computer-assisted survey instrument, study nurses were sex-matched, and interviews were conducted in a private location to improve the completeness and accuracy of self-reported sexual behaviour [16,46], but underreporting was still a challenge. Further research is needed to assess factors affecting the validity of self-reported behaviours among adolescents [47,48]. Importantly, underreporting of sexual behaviour highlights the need to have more robust biological measures of sexual risk, such as STI prevalence.
Finally, this survey is limited to the STIs we tested for—future surveys should consider surveillance of Mycoplasma genitalium infection and N. gonorrhoeae resistance in this population. In addition, surveillance of HPV infection and receipt of vaccination may be important to evaluate implementation of HPV vaccination programmes.
In conclusion, the global population of adolescents and young people is increasing, particularly in sub-Saharan Africa. STIs, including incident HIV, cluster in this population, especially among women. The principles of ‘epidemiology synergy’ between STIs and HIV strongly suggest that STI control must be addressed if HIV is to be brought under effective control [49]. Population-based, representative prevalence estimates of STIs should be complemented by robust prevalence estimates in key populations to gain a full understanding of the burden of STIs and the impact of interventions. Without robust prevalence estimates, moving an international STI agenda forward will continue to be a challenge. Nesting STI prevalence surveys in HDSSs could provide an efficient strategy for obtaining these data.
|
10.1371/journal.pcbi.1003656 | Variability of Metabolite Levels Is Linked to Differential Metabolic Pathways in Arabidopsis's Responses to Abiotic Stresses | Constraint-based approaches have been used for integrating data in large-scale metabolic networks to obtain insights into metabolism of various organisms. Due to the underlying steady-state assumption, these approaches are usually not suited for making predictions about metabolite levels. Here, we ask whether we can make inferences about the variability of metabolite levels from a constraint-based analysis based on the integration of transcriptomics data. To this end, we analyze time-resolved transcriptomics and metabolomics data from Arabidopsis thaliana under a set of eight different light and temperature conditions. In a previous study, the gene expression data have already been integrated in a genome-scale metabolic network to predict pathways, termed modulators and sustainers, which are differentially regulated with respect to a biochemically meaningful data-driven null model. Here, we present a follow-up analysis which bridges the gap between flux- and metabolite-centric methods. One of our main findings demonstrates that under certain environmental conditions, the levels of metabolites acting as substrates in modulators or sustainers show significantly lower temporal variations with respect to the remaining measured metabolites. This observation is discussed within the context of a systems-view of plasticity and robustness of metabolite contents and pathway fluxes. Our study paves the way for investigating the existence of similar principles in other species for which both genome-scale networks and high-throughput metabolomics data of high quality are becoming increasingly available.
| Organisms are usually exposed to changing environments and balance these perturbations by altering their metabolic state. Gaining a deeper understanding of metabolic adjustment to varying external conditions is important for the development of advanced engineering strategies for microorganisms as well as for higher plants. One tool which is particularly suited for investigating these processes is genome-scale metabolic models. These large-scale representations of the underlying metabolic networks enable the integration of experimental data and application of constrain-based mathematical approaches to estimate flux rates through the chemical reactions of the network under different environmental scenarios. However, for most of these approaches the assumption of a steady-state (flux balance) is indispensable and therefore precludes the prediction of metabolite concentrations. Here, we present a data-driven observation that relates results from a flux-centric constraint-based approach that is based on transcriptomics data to metabolite levels from the same experiments. Our observations suggest that constraint-based modeling approaches in combination with high-throughput data can be used to infer regulatory principles about the plasticity and robustness of metabolic behavior from the stoichiometry of the underlying reactions alone.
| Organisms, especially plants, are exposed to almost perpetually changing environments (e.g., light intensity and quality, nutrient and water supply) to which they respond by readjusting their cellular setup to efficiently utilize available resources and to ensure viability [1]–[6]. These transitions are often systemic in that they affect almost all levels of cellular organization, starting from gene expression to protein abundances and metabolite levels [7]–[9]. Therefore, a systems-based analysis is particularly suited for understanding the responses of plants to changes in the environment. Such an approach offers the possibility to integrate data which were simultaneously collected across different cellular levels to identify dependence between processes and to aid in testing hypotheses concerning the behavior of individual components or pathways.
Constraint-based approaches provide a modeling framework which is particularly amenable for systems-based analyses, since they not only allow for the integration of high-throughput data, but also rely almost solely on the stoichiometry of the reactions included in the models. For instance, with the help of Flux Balance Analysis (FBA, for details see Material and Method section) [10], [11] condition-specific steady-state flux distributions and growth capabilities can be readily predicted [12]. Moreover, recent studies have established that integration of high-throughput data can narrow down the space of feasible flux distributions and, therefore, results in improved predictions of biomass or contributes to more physiologically realistic engineering strategies [13]–[15]. The existing constraint-based approaches, which integrate data rely mostly on transcriptomics data and assume a relationship between the expression level of a given gene and the flux boundaries of the corresponding reaction in the metabolic network [14], [16], [17].
However, one of the main drawbacks of most constraint-based approaches lies in the nature of their problem formulation, i.e., the steady-state assumption, which precludes the integration and prediction of metabolite levels (detailed in the Materials and Methods section). Therefore, these approaches usually neglect the metabolome i.e., the levels of all considered metabolites which, along with reaction fluxes, act as one of the most informative indicators of the cellular metabolic state [18]. Existing attempts to integrate metabolite levels/concentrations into constraint-based approaches are restricted to predictions of reaction directionality via thermodynamic analysis [19]–[21] or require relaxation of the steady-state assumption [22]. In the current study, we ask whether (and if so, to what extent) we can make inferences about metabolite levels from a constraint-based analysis that is based on the integration of transcriptomics data.
In order to do so, we extend a previous study in which we used microarray data from Arabidopsis thaliana [23],[24] and integrated them in a metabolic network [25] to predict flux capacities for a large set of pathways under eight different light- and temperature conditions [26]. Furthermore, to make statistical statements, we compared the flux capacity profiles to those obtained from a biochemically meaningful data-driven null model. Based on this, we defined a pathway to be differential under a given condition if it exhibits a flux capacity profile that has an average absolute greater that 2 with respect to the null model. Moreover, we introduced the concept of metabolic sustainers and modulators. Sustainers are metabolic functions that are differentially up-regulated with respect to the null model and sustain a certain functioning, whereas modulators are differentially down-regulated [23], [24] to control a certain flux and modulate affected processes. A more detailed description of this study is given in the Results section.
Here, we present observations that link predictions made from the integration of transcriptomics data to metabolomics data from the same experiment. By doing so, we bridge the gap between flux- and metabolite-centric approaches. Most importantly, our findings demonstrate that under certain conditions, metabolites acting as substrates in pathways defined as modulators or sustainers of the metabolic state show a significantly lower temporal variation in comparison to the remaining metabolites. These observations are discussed within the context of a systems-view of plasticity and robustness of metabolite content as well as reaction/pathway fluxes. Taken together, our results demonstrate the power of transcriptomic data in predicting metabolic behavior in large-scale models and suggest an underlying regulatory principle governing metabolic stability.
FBA's objective function has a large effect on the predicted flux distribution [27]. For microorganisms under ambient conditions, the maximization of growth is a widely used cellular objective [28], [29]. However, when modeling plants metabolism this assumed objective does not necessarily hold true. Plants are more complex than microorganisms. They have multiple compartments within the cell, different cell types, several tissues and organs, which make it difficult to define a single objective function for the entire plant. Defining such an objective becomes even more challenging under stress conditions which have been shown to drastically alter plant's cellular chemical composition (see [5], [30], [31] and references therein).
In consideration of the absence of a reasonable biological objective function for plants experiencing stress, in our previously presented approach [23] we did not attempt to make predictions about actual fluxes through a metabolic pathway but rather aimed at predicting flux capacities. These capacities are derived from the integration of transcriptomics data into a large-scale metabolic model and represent maximum fluxes which certain pathways can carry under certain environmental conditions. While this concept can also be applied to single reactions of a network, we relied on the investigation of the functional units, the pathways, or in a more generic terminology, the metabolic functions.
We employed a transcriptomics dataset which captures the temporal response of Arabidopsis thaliana to eight different light and temperature conditions [26] and used the data to constrain the upper and lower flux boundaries of the reactions based on the E-Flux method [16] in a recent compartmentalized genome-scale model of Arabidopsis [25]. Subsequently, we predicted the flux capacities through a set of 167 metabolic functions, from primary and secondary metabolism, for each time-point and each condition.
Furthermore, to make statistical statements about the metabolic functions under consideration we compared the resulting flux capacities to predictions from a null model as a reference state. This analysis was motivated by the need to determine behavior of a metabolic function in a particular condition irrespective of an artificially placed reference state, which may not be representative for the “naturally occurring environment” which the plant experiences in the field. The employed null model was based on the permutation of the assigned flux boundaries while keeping thermodynamic and exchange constraints unaltered. In this manner, we circumvented issues with the selection of a reference state and relied on the average behavior determined solely by the network structure and the imposed flux boundaries. We re-computed the flux capacities from the null model for 100 repetitions for each time point and condition. Based on this, a metabolic function was deemed differential if it showed an absolute greater than 2 with respect to the expectation from the null model in at least one but not all conditions under consideration. Pathways that were differentially up-regulated were termed sustainers—sustaining a certain metabolic functioning, while those that were differentially down-regulated were referred to as modulators—modulating a certain metabolic functioning. A complete list of these pathways and their classification under the eight conditions considered is given in [23].
The working hypothesis of this study was motivated by the following: The determined modulators and sustainers exhibit flux capacities significantly different from the capacities expected by the null model. Furthermore, we observed that the direction of the differential behavior is unaltered across environmental conditions (i.e., differential metabolic functions are robust, or in genetic terminology, canalized [32], [33]). Therefore, we expected the likelihood for this observation to increase if the metabolites participating in these differentially behaving functions show persistently smaller fluctuations compared to metabolites involved in other functions. To test this hypothesis, we analyzed the metabolite profiles that were collected alongside with the transcriptomics data in the same experiments [26]. A schematic representation of the overall workflow from the data mapping and integration to the statistical analysis is provided in Figure 1. For each of the eight environmental conditions and set of investigated functions, we categorized the 65 mapped metabolites (see Materials and Methods) according to the following criteria: (1) participation in (non-)differentially behaving metabolic function and (2) metabolite, substrate, or product of a pathway. In addition, we made the distinction between substrates/products and initial substrates/initial products of the pathway.
We defined a substrate of a pathway as any metabolite that acts as a substrate in a reaction involved in the pathway but not as product/intermediate of the same pathway. Furthermore, an initial substrate is a substrate in the first reaction of the pathway. In an analogous manner: a product of a given pathway is defined as any metabolite acting as a product in a reaction of the pathway but not as a substrate/intermediate of the pathway. A product of the last reaction of the pathway is defined as a final product. For the metabolic function in Figure 2, , , and are substrates, while and are initial substrates; moreover, , and are products, while and are final products.
We calculated the variability of each considered metabolite over the time course following the perturbation by calculating the coefficient of variation (CV) as described in the Materials and Methods section. In order to determine differences in the distribution of CVs over the considered categories of metabolites we employed the Wilcoxon rank-sum test (which also is applicable to non-normal distributions) at a significance level of 0.05. We considered the distribution of CVs across all metabolites, across products only, and across substrates only, in the following six comparisons of groups: modulators vs. all metabolites, sustainers vs. all metabolites, differentially behaving functions (i.e., modulators and sustainers) vs. all metabolites, modulators vs. non-modulators, sustainers vs. non-sustainers, and differentially behaving functions vs. non-differentially behaving functions. As shown in Table 1, this summed up to a total of 144 statistical tests for three categorizations of metabolites over six groupings under eight conditions. A list containing the numbers of tested metabolites for each scenario is given in the Supporting Information S1.
First, considering the group of all metabolites (Table 1 top), we found the mean CV of metabolites involved in differential metabolic functions to be smaller in comparison to the mean CV of all metabolites under one conditions, i.e., high-light ( and , ). This was also the case when considering the mean CV of metabolites in modulators and sustainers, separately ( and , respectively). Secondly, analyzing the group of products of the pathways, we did not observe any significant differences in the CVs in any of the tested groups under any of the eight conditions (Table 1 middle). In contrast, investigating the third group—the substrates (Table 1 bottom) — we found the mean CV of the substrates in modulators to be significantly smaller than the mean CV of all metabolites under two conditions, i.e., () and (). Moreover, the mean CV of the substrates in sustainers was significantly smaller than the mean CV of all metabolites under one condition, i.e., (). Altogether, the mean CV of the substrates in differentially behaving metabolic functions was significantly smaller than the mean CV of all metabolites under three conditions, namely, under (), (), and (). Furthermore, the mean CV of the substrates in modulators was significantly smaller than the mean CV of substrates in non-modulators under two conditions, i.e., () and (), while the mean CV of sustainers was significantly smaller than the mean CV of substrates in non-sustainer under (). Finally, the mean CV of the substrates in differentially behaving metabolic functions was significantly smaller than the mean CV of substrates in all non-differential functions under four conditions, namely under (), -darkness (), (), and ().
Figure 3 shows a histogram of the distribution of CVs for all measured metabolites and those that participate as substrates in the metabolic functions which were previously identified as sustainer or modulator over all eight investigated conditions. Histograms for each condition separately can be found in the Figure S2 in Supporting Information S2.
Inspecting the list of mapped metabolites, we identified 15 out of 65 to act as substrates in a differential pathway in at least one of the considered conditions, namely: alanine, pyruvate, serine, threonine, aspartate, methionine, glutamine, 2-oxoglutarate, citrulline or arginine, spermidine, glycine, glutamate, ethanolamine, valine, and -alanine. The temporal profiles of these metabolites for those conditions in which they act as substrates in modulator or sustainer are shown in Figure 4. All of these metabolites are either amino acids or essential intermediates in central carbon or nitrogen metabolism. The differential pathways they belong to fall into the larger groups of primary nitrate assimilation (glutamate, glutamine, and 2-oxoglutarate), photorespiration (glycine, serine, and ethanolamine), TCA cycle (pyruvate and 2-oxoglutarate), amino acid metabolism (alanine, arginine, threonine, aspartate, methionine, and valine) and polyamine biosynthesis (spermidine and -alanine). A discussion about the involvement of the respective differential pathways in stress responses to the eight environments investigated was already given in [23]. Apart from their involvement as substrates in modulators and sustainers, these metabolites have also been implicated in various other stress responses, e.g., anoxia [34] or hypoxia [35], oxidative stress [36], drought stress [37], or general stress responses [38], [39].
Additionally, we tested if the described patterns of robustness in the metabolite profiles can also be found in the flux capacity profiles of the differential metabolic functions they belong to. Interestingly, only for we also observe the flux capacity profiles of the differential pathways to exhibit significantly lower CVs than the non-differential pathways (). Another general observation that we made is, that for all metabolic functions for which substrate measurements were available, the CVs of these substrates were significantly lower than the CVs of the respective flux capacity profile. These two observations further underline the none-trivial interconnection between flux rates and the levels of metabolites.
Next, we investigated whether the group of substrates in differential metabolic functions shows distinct characteristics with respect to the network topology. For the analysis we neglected evidently ubiquitous cofactors, such as: , , , ADP, ATP, , , NAD(P)H, CoA, (pyro-)phosphate. This strategy has also been followed in other studies [40]. Furthermore, to arrive at a value for the connectivity of each metabolite, i.e., the number of reactions in which a metabolite is involved (as defined in [41]), we kept the compartmentalized structure of the network and considered the instances of a metabolite appearing in more than one compartment as different reactants. Based on the given stoichiometry we determined the number of reactions in which each metabolite participates. Interestingly, we find the group of measured substrates of differential pathways to be on average significantly more connected than the group of all metabolites—6.24 vs. 2.99 reactions (, Wilcoxon rank-sum test). Clearly, one has to keep in mind that our analysis is based on a generic compartmentalized network reconstruction. The connectivity values might vary in different tissues, due to the presence or absence of certain pathways. Nevertheless, we belief that the well-curated model we use in our study serves a good starting point for the analysis.
To further investigate which attributes are typical for differentially behaving pathways, we next investigated the number of (initial) substrates of the pathways. Cofactors of the considered reactions were neglected from the analysis (see Supporting Information S3). Counting the number of substrates, we found their average number in the differential pathways to be significantly smaller in comparison to all considered pathways/all non-differential pathways (2.6 and 3.3/3.5 substrates, /, respectively). In contrast to this, when considering the number of initial substrates, we found that 29.7% of differential pathways have two initial substrates, while the remaining ones have only one substrate. In the whole group of metabolic pathways and the group of non-differential pathways this value is lower (25.5% and 26.2%, respectively) although not significant.
In this study, we extended our earlier analysis of Arabidopsis's metabolic acclimation to varying light and/or temperature conditions which was based on transcriptomics data [23], [24]. Here we considered metabolomics data from the same experiment and investigated the temporal variation of the metabolite profiles. Our findings from the integrative analysis include the following: (i) for specific environmental conditions, differential metabolic functions have substrates, which on average show a lower CV than other metabolite groups tested, (ii) when considering the network topology, these substrates are on average more connected than the remaining metabolites and (iii) differential metabolic pathways have on average fewer substrates than the other metabolic functions investigated.
Closer inspection of the environmental conditions that exhibit low substrate variability leads to the following hypothesis: substrate robustness can be observed under stressful environmental conditions. Yet, we do not observe substrate robustness, or in genetic terms canalization, under conditions which are not perceived as stress by the plant (e.g., ) and moreover the canalization effect might get lost under those conditions which are too extreme or prolonged (e.g., ). The latter scenario might cause a serious disturbance of the acclimation which could potentially lead to non-resilience, i.e., non-recovery. Therefore, we believe that the observed substrate robustness is an inducible genetic mechanism, both depending on the metabolic network structure and the specific environmental condition. Determining the range of conditions that permit the observed robust behavior would be an interesting undertaking for future experimental testing.
Deriving flux values from transcriptomics data is a delicate issue. In recent years, a large set of methods have been proposed that use transcriptomics data to infer condition-specific networks, mainly applied on microorganisms (GIMME, [42], iMat [43], E-flux [16], PROM [44], MADE [45], TEAM [46]). While most approaches rely on a discretization of the expression data and employ user-defined thresholds, the here applied E-flux method does not rely on these requirements. It assumes a relationship between the amount of a certain transcript and the upper flux boundary of the respective reaction. While mechanisms, such as post-transcriptional modification and hierarchical regulation [47], [48] cannot be explicitly considered, they are implicitly accounted for by only restricting the upper flux boundary. In other words if a certain amount of transcript was measured the predicted flux can range between zero and the upper boundary; no enforcements on certain minimum flux values are made. Additionally, claims are even made with more reservation since the approach does not attempt to predict actual fluxes but flux capacities that are compliant with the data. Moreover, one needs to keep in mind that the employed metabolite data are not compartment-specific. In the analysis presented here, we assigned the same metabolite profile to each compartment-specific compound in the model. It would be interesting to investigate in future studies, when more compartment-or tissue specific metabolite data become available, if the observed patterns of substrate robustness are not only specific for certain environments, but also for particular compartments or tissues.
Finally, like any other modeling attempt, any results depend on the quality of the network as well as on the quality of the collected data. Here, we relied on the most recent and most comprehensive network reconstruction of Arabidopsis and a dataset that was collected with a single technology in a single laboratory to minimize technical artifacts.
The principle of keeping levels of metabolites involved in important pathways from exhibiting fluctuations was recently discussed in another context. Reznik at al. used the dual formulation of a classical FBA problem, which uses the maximization of biomass as a cellular objective, to compute sensitivities of the objective value to flux imbalances, i.e., deviations from the steady-state assumption [49]. The so-called shadow price of a given metabolite captures the influence of the metabolite's accumulation or depletion on the maximum value of the objective. Thereby, a negative shadow price implies that the corresponding metabolite is growth limiting. By using data from S. cerevisae under different nutrient limiting conditions the authors were able to show that the determined shadow prices negatively correlate with the growth limitation of the respective measured intracellular metabolites. Moreover, based on these findings, the authors argued that growth-limiting metabolites cannot exhibit large fluctuations. Using data from E. coli's metabolic response to carbon and nitrogen perturbations, they further demonstrated that metabolites associated with a negative shadow price indeed show lower temporal variation in comparison to metabolites with zero shadow prices in a perturbed system.
What both approaches, ours and the one briefly described above, have in common is the principle that metabolites important for a particular function exhibit less temporal variation than other metabolites. In the latter, an important metabolite is defined as a metabolite with a negative shadow price with respect to the assumed cellular objective of growth maximization. In contrast to this, our analysis is driven by integration of transcriptomics data and does not assume a particular overall cellular objective. In our approach, we consider a metabolite relevant if it acts as a substrate in a metabolic pathway which behaves differentially in comparison to a condition-specific null model for flux capacities. These relevant metabolites may thus play a role in the plant acclimation to environmental changes.
Additionally to the observed substrate robustness in differential pathways under certain abiotic stress conditions, we also showed that these substrates are on average more connected i.e., involved in more reactions. The role of these highly connected metabolites has previously been discussed in terms of evolution [40], [50]. In the latter, the authors identified among others, pyruvate, serine, aspartate, 2-oxoglutarate, and glutamate and put forward the hypothesis that the most highly connected metabolites should also be the phylogenetically oldest [40]. The connection between metabolites involved in core reaction of central carbon metabolism and their involvement in abiotic stress acclimation, together with the observation that they are on average more connected, extends this concept and puts the evolutionary structure of metabolic networks into a more dynamic context—one which also accounts for the changing environments affecting the organism.
Our third finding concerning the smaller number of substrates in differentially behaving metabolic functions has wide implications on the interplay between plasticity and robustness in metabolism. Most notably, our findings differ from claims made with respect to evolution of robustness and cellular stochasticity of gene expression. In a recent study, the author proposed that the degree to which varying cellular components combine to determine robust phenotypes may be predictive of the amount of their inherent variability. The basis for this claim is the observation that averaging over multiple independent inputs is a general way to reduce variability of molecular phenotypes [51]. This implies that the larger the number of variable inputs is, the smaller the variability of the phenotypic output will be. However, this observation does not apply to metabolic reactions which are governed by multiplicative (e.g., mass action, as the simplest) rather than averaging laws. Here we observe that fewer input variables with smaller fluctuations, do not necessarily result in smaller fluctuations of the output (i.e., the flux capacity in our case) but in robust differential behavior. Furthermore, our findings also showed that for all metabolic functions for which substrate measurements are available, the CVs of their substrates are significantly lower than the CVs of the respective flux capacity profile. This further highlights the particularities of regulation, variability, and robustness of metabolic pathways.
Robustness of certain pathway fluxes and specific metabolite concentrations have long been documented. The concept of network rigidity has initially been proposed in S. cerevisae [52]. Subsequently it has been demonstrated to be functional in plant systems too, especially in the context of central metabolism [53], [54]. Moreover, considered evidence has also accrued for certain metabolite levels to be exceptionally stable, for example the levels of alanine, pyruvate, 2-oxoglutarate, glutamine and spermidine [55], [56]. Furthermore, it has been shown that levels of metabolites such as serine coordinately control the level of expression of genes which encode multiple steps of the pathway in which they themselves take part [57].
In our view, the high stability of a pool of primary metabolites, invariant to environmental heterogeneity, fulfills two major functions. On the one hand, it efficiently sustains a set of “core” reaction rates which are deemed essential for the plant's objective function across a wide range of different stresses. On the other hand, the observed substrate stability enables a tight conditional control on a set of metabolic functions to act as modulators or sustainers in response to specific stresses only.
Finally, the fact that the robust metabolites may well be the most biologically relevant for metabolic regulation is an important point since it is at odds with the manner in which the majority of the metabolomics community assesses their data. This fact additionally highlights the potential difficulties and challenges in interpreting data from a single level of the cellular hierarchy and thus provides further grounds for integrated models such as the one we present here.
Taken together, our findings show that the integration of large-scale modeling with high-throughput data can be used to infer regulatory principles from the stoichiometry of the underlying reactions alone. Furthermore, we presented an approach that bridges the gap between flux-centric and metabolite centric view of large-scale data. Therefore, our study paves the way for investigating the existence of similar principles relating plasticity of metabolic profiles and robustness of metabolic behavior across other species for which both genome-scale networks and high-throughput (time-resolved) metabolomics data of high quality are becoming increasingly available.
The investigated data set captures the time-resolved response of Arabidopsis thaliana to changing light and/or temperature conditions [26]. The previously published data comprise time-series measurements for eight environmental conditions covering combinations of four different light intensities (ranging from high-light () to darkness) and three different temperatures (4, 21, and ). A schematic representation of the combinations of abiotic stresses is provided in the Figure S1 in Supporting Information S2. In brief, wild-type Arabidopsis thaliana Columbia-0 plants were grown in soil under short-day-conditions for 4 weeks and then transferred to long-day-conditions for another 2 weeks. Subsequently, they were exposed to the following conditions: ; ; ; ; ; ; , and . Metabolite and transcript profiles were collected from samples harvested at 22 time points ranging from 0 to 24 hours after the stress application. Further details of the experimental procedures and data processing can be found in the original publication [26]. Transcriptomics data are deposited in the array express repository (http://www.ebi.ac.uk/arrayexpress) under Arabidopsis light and temperature response ArrayExpress accession: E-MTAB-375 and they can be downloaded using the following link: http://www.mpimp-golm.mpg.de/Supplementary-Materials-for-Publications/Caldana-et-al_Filtered-Affymetrix-Gene-Expression-Data.zip. Metabolomics data are provided on the following website http://www.mpimp-golm.mpg.de/Supplementary-Materials-for-Publications/Caldana-et-al_Normalized-metabolic-data.zip.
From the total of 82 measured metabolites, 65 can be mapped onto the model. It must be noted that the metabolomics data are not compartment-specific and 61 out of 65 mapped metabolites appear in more than one compartment in the model. Due to a lack of this information, for those non-unique metabolites we assigned the same profile to each compartment-specific compound in the model. The mapping dictionary is given in the Supporting Information S4.
The mapping of the transcriptomics data from our previous study has a network coverage of 46%, i.e., 627 out of 1363 reactions can be constrained by transcriptomics data [23], [24].
FBA is a constraint-based approach for predicting steady-state fluxes in a metabolic network independent of enzyme kinetics and metabolite concentrations [10], [11]. The method solely relies on the physico-chemical constraints of the network (e.g., the reaction stoichiometry, reversibility, and maximum uptake rates) and a putative biological objective of the organism under consideration (e.g., biomass production for microorganisms under ambient conditions). A central element of the approach is the assumption of a steady-state which implies that each internal metabolite in the network is produced and consumed at the same net rate if considering the system at a small time interval, or in a mathematical representation: (1)which results in a decoupling of the flux predictions from the metabolite concentrations. Adding the above mentioned additional constraints and assumptions leads to the following linear program:(2)(3)(4)where is the stoichiometric matrix of the system under consideration in which the rows denote the metabolites and the columns represent the reactions of the model. The reaction fluxes are captured in the flux vector . The respective lower and upper boundaries of the reaction are given by and . The vector encodes the ratios at which certain precursors (e.g., amino acids, fatty acids, nucleotides, sugars) contribute to the objective function. For a detailed review of FBA and other related constraint-based optimization approaches see [11], [13].
The analysis presented here extends results recently presented in [23]. In brief: In our previous study we had simulated flux capacities through a set of a 167 metabolic functions. The simulation of metabolic functions has initially been proposed to demonstrate the quality of a metabolic reconstruction [58] and it has also been used in the original model reconstruction [25] to ensure model functionality. The set of selected pathways cover primary as well as secondary metabolism and is obtained from AraCyc/MetaCyc [59]–[61]. These databases incorporate community-wide efforts to integrate current biological knowledge and understanding of metabolic pathways. In our previous study we extended the proposed concept to simulate time-and condition-specific flux capacities by integrating transcriptomics data based on a modification of the E-flux method [16], which assumes a relationship between the amount of a certain transcript and the upper flux boundary of the respective reaction. Since the correspondence between transcript and protein abundance is crucial when using transcriptomics data to constrain flux boundaries, the approach only makes weak assumptions. Proteins are allowed to be present and active if the respective gene product is detected. In contrast, no enforcements on protein activities are made if the gene product was detected with certain abundance. Additionally, our claims are stated with even greater reservation since the approach does not attempt to predict actual fluxes but flux capacities that are compliant with the data. For details of the simulation and the list of metabolic functions refer to [23], [25].
To determine the temporal variation of the metabolite profiles we used the coefficient of variation (CV) which is defined as the ratio of the standard deviation and the mean of the observable: (5)
We applied this statistic to the mapped metabolite profiles for each condition separately. While doing so, we neglected the data for the first hour (first six time points) after the stress application during which the system experiences the strongest effect, i.e., differential regulation of pathways involved in the response. To ensure that the categorization of differential metabolic functions is robust to the removal of this time-interval we repeated the analysis of our previous study. Reassuringly, we find our results, i.e., the classification as a sustainer or modulator of a given metabolic state, to be robust to the removal of up to six time points from the beginning of the time series.
|
10.1371/journal.ppat.1002615 | An “Escape Clock” for Estimating the Turnover of SIV DNA in Resting CD4+ T Cells | Persistence of HIV DNA presents a major barrier to the complete control of HIV infection under current therapies. Most studies suggest that cells with latently integrated HIV decay very slowly under therapy. However, it is much more difficult to study the turnover and persistence of HIV DNA during active infection. We have developed an “escape clock” approach for measuring the turnover of HIV DNA in resting CD4+ T cells. This approach studies the replacement of wild-type (WT) SIV DNA present in early infection by CTL escape mutant (EM) strains during later infection. Using a strain-specific real time PCR assay, we quantified the relative amounts of WT and EM strains in plasma SIV RNA and cellular SIV DNA. Thus we can track the formation and turnover of SIV DNA in sorted resting CD4+ T cells. We studied serial plasma and PBMC samples from 20 SIV-infected Mane-A*10 positive pigtail macaques that have a signature Gag CTL escape mutation. In animals with low viral load, WT virus laid down early in infection is extremely stable, and the decay of this WT species is very slow, consistent with findings in subjects on anti-retroviral medications. However, during active, high level infection, most SIV DNA in resting cells was turning over rapidly, suggesting a large pool of short-lived DNA produced by recent infection events. Our results suggest that, in order to reduce the formation of a stable population of SIV DNA, it will be important either to intervene very early or intervene during active replication.
| New treatments for HIV have proved very successful at controlling viral replication and preventing the onset of AIDS. However, these treatments must be continued for life, because if they are stopped the virus rapidly ‘rebounds’ to its original levels. The reason for this rebound is the existence of a population of viruses that lie dormant inside cells during treatment, and reactivate as soon as treatment is stopped. This ‘latent virus’ is extremely long-lived under drug therapy conditions, and therefore presents a major barrier to viral eradication. However, very little is known about the survival and reactivation of latently infected cells during ongoing infection, because virus is being formed and destroyed all the time. We have developed a novel ‘escape clock’ approach to measure how long viral DNA lasts in monkeys. We find that, in the setting of low viral load, the lifespan of infected cells is very long, whereas during active infection there is a surprisingly high turnover of viral DNA within resting CD4 T cells. We believe this is due to high level of immune activation when there is a high level of replicating virus. This result may have important implications for the optimal timing of drug treatment.
| Treatment of HIV-1 infected individuals with highly active antiretroviral therapy (HAART) can suppress plasma viral RNA levels below the threshold of detection by standard diagnostic assays. However after cessation of even long-term HAART, virus replication is quickly re-established [1]–[4]. A barrier to viral eradication is the presence of viral DNA stably integrated into the chromosomes of resting CD4+ T cells and other long-lived cell populations, since the decay of this viral compartment is very slow [5]–[17].
Several studies suggest that persisting integrated viruses are laid down early in infection [18]–[20]. An indication for this is that HIV strains cultured from resting CD4+ T cells are genetically distinct to concurrent plasma virus [20]. However, it is generally difficult to study the precise kinetics of establishment and turnover of the latent reservoir in most human cohorts as the infecting isolate is usually not known and serial samples available during asymptomatic early infection (1–2 weeks post transmission) are difficult to acquire. Studying macaques experimentally infected with SIV overcomes these barriers. Recent studies show the utility of SIV-infected macaque models for studying long-term integrated viruses [21]–[23].
During HIV and SIV infection one typically sees immune escape at defined cytotoxic T cell lymphocyte (CTL) epitopes. CTL escape mutations (EM) are frequently generated early after the acute infection stage and typically follow predictable patterns of outgrowing wild type (WT) virus. We hypothesised that evidence for the early formation and turnover of SIV DNA may be found by comparing the dynamics of immune escape in cellular viral DNA populations to the dynamics in replicating plasma virus. If replicating WT virus (as indicated by plasma RNA) is only present during acute infection, and is completely replaced in plasma by replicating EM virus during chronic infection, the latent reservoir will be primarily WT if laid down during acute infection, but predominantly EM if laid down (or rapidly turned-over) during chronic infection. In other words, if WT viral DNA is detected in resting CD4+ T cells during chronic infection and remains at similar levels when measured later, this supports low rates of turnover of viral DNA populations during active infection.
We previously developed sensitive real-time PCR assays to quantify EM and WT viruses at a Mane-A*10-restricted SIV Gag CTL epitope (termed KP9) in replicating plasma RNA [24], essentially providing a “viral load” of both WT and EM quasispecies. For this study, we also developed PCR assays to assess WT and EM populations in cellular SIV DNA in FACS-sorted resting CD4+ T cells. After obtaining serial plasma and PBMC samples from Mane-A*10 positive SIV-infected pigtail macaques, we used the observed evolution of WT and EM replicating plasma SIV RNA viruses to model the turnover rate that resulted in the observed relative levels of WT and EM SIV DNA sequences. These analyses suggest that during periods of active high-level infection, the majority of SIV DNA in resting CD4+ T cells is turning over very rapidly. However, at lower levels of infection a substantial proportion of SIV DNA in resting CD4+ T cells is laid down early (when virus is still WT at the CTL epitope) and this WT reservoir persists at stable high levels during chronic infection.
We first studied 12 Mane-A*10+ animals in a SIV vaccine trial using an influenza virus vector, as this large study provided an extensive bank of plasma and PBMC samples (Table 1). We characterized the frequency of CTL escape mutant and wild-type variants in SIV plasma RNA (Figure 1A). The evolution of EM and WT viruses in plasma was derived by a previously described SIV Gag KP9 qRT-PCR that specifically discriminates the K165R EM virus [24]. After infection with the SIVmac251 challenge stock, peak SIV viremia with predominating WT virus was observed ∼2 weeks post infection in all animals. K165R CTL immune escape predictably occurred [24], [25] and EM virus predominated in chronic infection, being selected in preference to WT virus in the majority of animals. Several animals had plasma viremia trajectories in which there was minimal or no detectable EM virus in plasma during acute infection, and complete or near complete replacement of WT virus with EM virus in plasma during chronic infection.
To assess SIV DNA within resting CD4 T cells, we sorted cells based on their being positive for CD3 and CD4 and negative for HLA-DR and CD69/CD25 (Figure 1B), and then performed a nested PCR with the second round being specific for either WT or the K165R CTL EM virus (Figure 1C). All animals were infected with SIVmac251 that is WT at this CTL epitope. The PCR provides relative levels of EM and WT SIV DNA in resting CD4 T cells. We performed assays on FACS-sorted resting CD4 T cells obtained from PBMC samples collected over the course of infection, and compared the ratio of WT and EM virus in plasma with that in SIV DNA in resting cells (Figure 1D).
These two ratios of WT/EM are the basis of the “escape clock” that we use to estimate the SIV DNA turnover rate in resting cells. The method is outlined in the in Figure 2. Briefly, if SIV DNA turns over quickly (or has a short half-life), then we expect the fraction of WT virus in SIV DNA to closely track the ratio seen in plasma virus, since most of SIV DNA would have been recently formed from plasma virus. If, on the other hand, SIV DNA is extremely stable (or persists indefinitely), then we expect the fraction of WT in SIV DNA to reflect the accumulation of all latently infected cells over the whole previous course of infection. The archived viral DNA of each strain should then be proportional to the ‘area under the curve’ (AUC) of each virus over time. For any SIV DNA ratio in between these extremes, we could estimate the optimal half-life of SIV DNA that best fits the data using the model described in the Methods section.
In the majority of these animals with active replication (which had high viral loads) we observed that CTL escape in the plasma SIV RNA was closely followed by CTL escape in the SIV DNA from FACS-sorted resting CD4 T cells (first 8 animals in Figure 3). The estimated half-life of SIV DNA was therefore extremely short – of the order of a few days.
To confirm these results we also assessed reversion of the K165R KP9 CTL escape mutation in a Mane-A*10 negative animal infected with the SHIVmn229 challenge stock. This challenge stock had previously been passaged in a Mane-A*10+ animal and was composed largely of the K165R escape mutation [26]. In the absence of CTL pressure, we observe plasma SIV RNA rapidly reverting back to wild type, as previously reported [26]. Thus, this provides an ‘escape clock’ where EM instead of WT virus is temporarily expressed, and thus provides an excellent control for our measurements of WT∶EM ratio. Consistent with our findings with WT SIVmac251 infection, we found that the cellular SIV DNA in resting CD4 T cells also reverted back to wild type very rapidly in this animal with high levels of viral replication (Figure 4).
The rapid turnover of cellular SIV DNA in resting CD4 T cells that we observed during high level SIV infection above was surprising, given the accepted stability of the latent reservoir in subjects with very low levels of replication on HAART. To investigate the effects of plasma viral turnover on the persistence of SIV DNA, we repeated the study on a cohort of 8 Mane-A*10+ animals from a peptide immunotherapy trial. These animals had undergone ART at week 3, followed by immunotherapy and cessation of ART (week 10), leading to long-term low levels of viral replication in many animals. In these animals, escape was usually observed in the plasma following therapy interruption, leading to the rapid appearance and dominance of EM virus in chronic infection. Thus we were able to study SIV DNA dynamics in resting CD4+ T cells in chronic infection at a time of low viral loads in the absence of therapy. The results of fitting the half-life of WT DNA in animals from both trials are shown in Figure 3, in the order of increasing estimated half-life.
Analysis of the proportion of WT SIV DNA in resting cells from these animals produced a very different picture from that seen in the first cohort. In several animals, the SIV DNA in resting cells remained close to 100% WT, despite EM virus dominating the plasma for prolonged periods. When we estimated the half-life of SIV-DNA in these resting cells using the ‘escape clock’ approach, we found that in 4 animals with very low viral loads the half-life of SIV-DNA in resting cells was >20 years. In several other animals, although some turnover could be measured, the half-lives were extremely long.
We then investigated whether viral load was a correlate of the rate of SIV DNA turnover in sorted resting CD4+ T cells. We observed a significant correlation between viral load and estimated SIV DNA half-life (Figure 5), suggesting that the high levels of infection and CD4+ T cell activation may play a role in determining SIV DNA turnover.
Current HAART regimes suppress plasma HIV RNA to very low levels, but cessation of HAART results in a brisk rebound of plasma virus [1]–[4]. Cellular compartments containing viral DNA provide a stable long-term reservoir for the virus [5]–[17]. However, the dynamics of establishment and turnover of this reservoir are not well understood. In particular, the majority of studies of HIV latency have focused on HIV DNA turnover under therapy, when viral replication and CD4+ T cell activation and turnover are greatly suppressed. But is HIV DNA persistence the same during active infection?
Our approach using a qRT-PCR to track the evolution of CTL escape mutants allowed us to compare EM and WT virus in plasma RNA and cellular DNA in a cohort of Mane-A*10 positive SIV-infected pigtail macaques. By analysing SIV DNA in purified resting CD4 T cells and comparing with plasma virus, we show that WT SIV DNA can persist in some animals for many months, even where there is an absence of WT viral RNA in replicating plasma virus. Thus, the WT SIV DNA species that are laid down early during infection can persist into late infection, and turnover of this viral compartment is very slow. Importantly however, this long half-life of WT DNA was only seen in animals with low viral loads. Animals with a high viral load showed a very rapid turnover of WT SIV DNA in resting CD4+ T cells. Indeed, we observed a highly significant correlation between the average viral load in chronic infection and the estimated half-life of SIV DNA. This correlation suggests that viral load is an important factor driving SIV DNA turnover in resting CD4 T cells during active infection.
Our observation of the long half-life of SIV DNA associated with low levels of plasma virus is consistent with the previous studies of HIV DNA persistence under drug therapy, where viral loads are even lower than those observed here [14]. What is less clear from our study is why we see such a rapid SIV DNA turnover in resting cells during active infection. A number of mechanisms are possible. Firstly, it seems possible that CD4+ T cells simply don't get a chance to stay in the resting state long enough to maintain a stable integrated pool, since SIV DNA is continuously driven to productive infection because of host cell activation. The half-life that we are estimating here is then half-life spent in the “resting” pool, i.e. the time during which infected cells express CD3 and CD4 and are negative for HLA-DR, CD69 and CD25. When they activate, they are no longer sorted as resting, and are lost from the pool in the same way as if they died. One limitation to our conclusions about the latent infection is that they apply only to CD69−CD25−HLA-DR− infected CD4+ T cells in blood, which may not truly represent the latent pool, but may be a heterogeneous population containing truly latently infected cells as a small subset. Thus, although the observed average turnover of HIV DNA in these resting CD4+ T cells is sometimes extremely fast, we cannot exclude that there might be minor populations of cells harbouring much longer-lived DNA, or that indeed long-lived DNA might not be harboured at some other anatomical site. Although we also found very few effector memory or dividing cells within the sorted resting CD4 T cell population, future studies could sort even more highly refined resting CD4 T cell populations or investigate other anatomical sites to evaluate this further.
A second possibility is that the observed SIV DNA represents a mix of long-lived, integrated SIV DNA, and of short lived reverse transcription products that represent dead-ends for the virus. At low viral loads, the level of short-lived reverse transcription products is very low, as there is little virus present in plasma to produce new infections. Thus, the SIV DNA observed comes predominantly from the long-lived integrated pool, and we observe the slow SIV DNA turnover characteristic of this compartment. However, in animals with a high viral load, we may see a high level of recent infection and of short-lived reverse-transcripts. If viral loads are high enough, this pool of recent reverse transcripts is the dominant population we see, overwhelming the long-lived WT DNA, and leading to an apparent close tracking of the viral DNA in resting cells with the plasma DNA. This mechanism also predicts that the long-lived WT DNA pool always persists at the same level, but is numerically overwhelmed by the large number of copies of short-lived EM DNA when plasma viral load is high.
In our model we did not consider a possibility that the half-life of infected resting cells depends on viral strain, because we assumed that they would not express viral epitopes and would not be recognized while resting. In addition, the fraction of WT DNA in resting cells is in most cases higher than in plasma, which is not supportive of preferential killing of resting cells with WT DNA. However, it is in principle possible that WT-infected resting cells are preferentially killed during periods of fast turnover, when the WT fraction is very low and approaches that in plasma. We note though that if preferential killing of WT infected resting CD4+ T cells were driving the rapid turnover of latency, we might expect that the turnover would be correlated with CTL number.
Specifically, if this were the mechanism driving the turnover, we would expect that in animals with good CTL control (low viral loads) we should see faster turnover of HIV-DNA. However, we observed the opposite relationship. Moreover, when we analysed the correlation between the number of tetramer positive cells and HIV-DNA turnover, we found that both in early (before the appearance of EM in plasma) and in chronic infection the average number of tetramer positive cells was positively correlated with the estimated half-life of resting infected cells. This is in agreement with our interpretation that better immune control leads to less reactivation of latently infected cells.
Our analyses are in agreement with early studies in humans suggesting that latent reservoirs in resting CD4+ T cells are laid down earlier in infection and are extremely long-lived [20], [27]. There are a number of possible mechanisms by which long-lived SIV/HIV DNA may persist in cells (illustrated in Figure 6). Firstly, the individual cells bearing HIV DNA may be extremely long-lived. Secondly, these cells may turn over through homeostatic replication, with a balance of cell replication and death leading to a stable number of HIV DNA copies. Finally, it has been suggested that HIV persistence may be maintained by low levels of viral reactivation, replication, and reinfection of new cells, leading to a stable level of HIV DNA copies. This latter mechanism seems unlikely at low plasma viral loads given our results. That is, if WT DNA persistence involved reactivation, viral production into the plasma, and reinfection of new cells, then we should be able to estimate the proportion of new infections due to WT virus, by the ratio of WT∶ EM virus in the plasma. However, since in most cases we don't observe any WT virus in the plasma in chronic infection, it could at best make only a very trivial contribution to any reinfection, and could not maintain WT DNA levels at or above the AUC levels via this mechanism.
There are limitations to our data that suggest further studies. Discriminating integrated from non-integrated forms of SIV DNA was not feasible in the small numbers of sorted resting CD4 T cells without further optimising the assay. Several assays have been designed to measure the more abundant non-integrated forms of HIV/SIV such as 1 LTR and 2-LTR circular forms, and these could be used in the future to determine the level of contaminating non-integrated SIV [28]–[31]. No widely accepted method exists to measure the “true” latent reservoir and our studies of resting CD4 T cells are only an approximation of this as yet undefined cell population. PCR-based assays have the disadvantage in that much of the cellular HIV-1 DNA may be from replication-deficient virus, although for viruses to contribute to the latent reservoir they must be replication-competent [20], [32]. However, our analyses focus only on a single nucleotide change in the KP9 CTL epitope. This change (alone) is replication competent and it seems unlikely that additional lethal mutations would accumulate more commonly in either WT or the K165R EM species.
Our results provide a method for direct quantification of HIV DNA turnover during active infection, and show for the first time that SIV DNA turnover in resting CD4+ T cells is strongly correlated with viral load during chronic infection. The rapid turnover of SIV DNA in animals with high viral load suggests the resting pool of CD4+ T cells and the pool of SIV DNA may be much more dynamic than previously thought during active infection. However, the persistence of WT SIV DNA laid down in early infection in animals with low chronic viral loads indicates the importance of the earliest stages of infection in establishing the latent pool of HIV DNA. Taken together, our study highlights the importance of early viral control in preventing the establishment of persistent HIV infection.
Experiments on pigtail macaques (Macaca nemestrina) were approved by CSIRO livestock industries Animal Ethics Committees (approval number 1315) and cared for in accordance with Australian National Health and Medical Research Council guidelines.
We studied serial PBMC and plasma samples from 20 pigtail macaques involved in several SIV infection studies [33], [34]. Five macaques received no SIV vaccinations and were infected with SIVmac251 (wild type at KP9) [35]. Two macaques received influenza viruses expressing KP9 and were infected with SIVmac251 [34]. Five macaques received influenza viruses expressing KP9, KSA10 and KVA10. Eight Mane-A*10 positive pigtail macaques were enrolled in a therapeutic peptide-based vaccine trial [33]. The outline of the therapeutic study was as follows: pigtail macaques were infected intravenously with SIVmac251 at week 0 and received ART (tenofovir and emtricitibine) from weeks 3 to 10 post infection and then withdrawn. Either no treatment (controls) or OPAL treatment (overlapping 15mer Gag peptides only or peptides from all 9 SIV proteins) was given at weeks 4, 6, 8 and 10 after infection. PBMC and plasma samples were collected at regular time points on all animals. The animals and their treatment are summarized in Table 1.
To quantify virus levels of WT or EM quasispecies at the KP9 epitope we employed a discriminatory real-time PCR assay as described [24], [36]. Briefly, the assay uses a forward primer specific for either wild-type sequence or specific for the nucleotide mutation encoding the dominant K165R KP9 escape mutant. At each timepoint after infection 10 µl of plasma RNA was reverse-transcribed and then amplified by qRT-PCR using either WT or EM forward primers. A common reverse primer and FAM-labelled DNA probe were also added for quantification against the appropriate SIV Gag epitope RNA standards using an Eppendorf Realplex4 cycler. Analysis was performed using Eppendorf Realplex software. Baselines were set 2 cycles earlier than real reported fluorescence and threshold value was determined by setting threshold bar within the linear data phase. Samples amplifying after 40 cycles were regarded as negative, and corresponded to <1.5-Log10 SIV RNA copies/ml of plasma. Plasma viral cDNA was also subjected to Sanger-based sequencing as previously described [37] to confirm the EM quasispecies contained the K165R mutation detected in the qRT-PCR assay.
We studied HLA-DR-CD69−CD25− CD4+CD3+ T lymphocytes as resting CD4 T cells as these cells are commonly studied as a model for resting CD4 T cells [32], [38]–[41]. Frozen PBMC (approximately 5×106) were thawed at 37°C in RF10, centrifuged at 300 g and resuspended in 500 ul of PBS containing 2 mM EDTA. 0.5 µl live/dead (Near Infra Red –IR (APC-Cy7) viability stain/tube was added and tubes were incubated for 60 minutes on ice in the dark. Cells were washed for 5 min at 500 g, the supernatant removed and surface stained with an antibody cocktail of CD69-APC (clone L78), CD3-PE (clone SP34-2), CD4-FITC (clone L200), CD25-APC (clone BC96) and HLA-DR- PerCP Cy5.5 (clone L243) on ice in the dark for 60 minutes to avoid CD4 T cell activation. PBMC were washed in PBS containing 0.5% BSA and 2 mM EDTA, fixed in 0.1% formaldehyde and passed through filtered facs tube prior to being sorted on the FACSAria. Live resting CD4+ T cells were positive for CD3 and CD4 and negative for HLA-DR and CD69/CD25. To validate whether HLA-DR-CD69−CD25− CD4+CD3+ T lymphocytes studied were truly resting using other markers we also studied CD28 and CD29 memory markers and the cell turnover marker Ki67 in a subset of the studied animals. The activated cells were highly enriched for CD28−95+ effector memory cells. A mean of only 0.91% of the resting cells were of the effector memory phenotype (p<0.001). The activated cells were also highly enriched for Ki67 staining (p = 0.0028). A mean of only 1.64% of the resting cells were Ki67+.
DNA from sorted cells was extracted using the Qiagen mini DNA. A nested KP9-specific qRT-PCR was performed that consisted of a first round Gag specific PCR followed by a second round discriminatory KP9 qRT-PCR. The first round PCR utilized 400 nM of the Gag forward primer (5′- CAAGTAGACCAACAGCACCATCTAGCGGCAG-3′) and reverse primer (5′- CTTGTTGTGGAGCTGGTTGTGGGTGCTGCAAGTC). Amplification was performed using 2 U Expand High Fidelity polymerase, 200 nM dNTPs and 250 mM MgCl2 per reaction. Amplification consisted of 94°C for 2 minutes followed by 22 cycles of 94°C for 15 seconds, 68°C for 30 seconds and 72°C for 45 seconds, with a final extension of 72°C for 7 minutes. The second round dKP9 qRT-PCR has previously been described for the analysis of WT and EM virus in plasma RNA, was performed [24], [36]. The second round product was serially diluted 1/100 to 1/2000 to ensure that the second round qRT-PCR did not contain saturating amounts of DNA. Sanger based sequencing confirmed the ratios of WT∶EM virus observed in the real-time PCR reaction (not shown).
We start from a simple model describing WT and EM infection in resting CD4+ T cells:(1)In this model, cells infected with WT and EM (IW and IE respectively) are becoming resting (RW and RE for resting cells infected with WT or EM respectively) at a fixed rate μ and have a half-life of τR. Half-life can describe the loss of resting cells either to death or to activation. We do not have access to IW and IE from experiment, but we assume that free virus is turning over much faster than productively infected cells [42], so that plasma virus to a good approximation reflects the corresponding productively infected cell level. This proportionality holds irrespective of the cause of viral load variation – be it because of the variation in target cell numbers, immune response or drug therapy. It allows us to replace μIW and μIE in Eq.1 with fW and fE, where W and E are the plasma WT and EM viral loads, and f is a constant different from μ.
We then use this model to estimate the half-life of viral DNA in resting infected cells using the measured WT and EM viral loads and the fraction of WT DNA in resting cells bW = RW/(RW+RE). For this purpose we rewrite the system Eq.1 (with W and E replacing IW and IE) in terms of the WT fraction bW and a variable Λ representing total number of infected resting cells scaled by constant f, Λ = (RW+RE)/f:(2)The system Eq.2 has only one fitting parameter, the half-life of viral DNA τR. We find the best fit of τR for each animal by solving Eq.2 with initial conditions Λ(t = 0) = 1 and bW(t = 0) = 1 (because inoculating SIVmac251 is 100% WT) and choosing a value of δ = ln2/τR between 0 and 2 such that it minimizes the deviations from measured points of WT DNA fraction, using measured values of WT and EM viral load with exponential interpolation between time points for the variables W and E. Because the fitted values of the WT fraction bW must lie between 0 and 1, the best-fit value of δ has to minimize the expression [43]:(3)where ti are the time points when the measurements were taken, bWexpt are the measured values of the WT fraction at this point, and bWpred are the values predicted by the model. The arcsin-square root transformation of the deviations was used to normalize the error distribution.
Given the initial conditions, the solution Λ(t), bW(t) of the system Eq.2 for each value of τR is unique. Therefore it is sufficient to have one experimental WT fraction different from 1 or 0 to completely define the trajectory of WT fraction in time. We have such points for all animals. Even when a fraction looks like 1 or 0, it often deviates a little from these numbers. The largest possible value of τR that could be obtained from our model is infinity (which we report as >105 because δ = 10−5 is the lowest parameter value used that was greater than zero), and the lowest possible value is 0.35 days (if this is the best fit, we report it as <0.5 days).
The confidence intervals for each animal in Figure 5 are obtained by bootstrapping, using the errors transformed by arcsin-square root. The initial point bW(0) = 1 was not used in bootstrapping. Because the total number of measurements for each animal was small, we used the whole set of error permutations to determine the bounds of the confidence intervals.
Our model assumes that the constant f is the same for cells infected with WT and EM. However, this constant may be higher or lower in one of the strains, depending on its propensity μ to generate latently infected cells, fitness cost of mutation and type of immune response. We have therefore repeated the process by simultaneously fitting the ratio fE/fW and τR and found that this did not change the range of observed half-lives or the correlation of half-lives with chronic viral load. The details of this analysis are shown in the Text S1 in the online Supplementary material.
It should be noted that RW and RE, which we regarded as infected resting cells in Eq.1 and Eq.2, can also be interpreted as the WT or EM DNA content in resting cells (just as IW and IE can represent viral DNA in productively infected cells), In this case, the half-life of τR describes the loss of viral DNA due to cell death, degradation or resting cells becoming activated.
|
10.1371/journal.pbio.1002349 | Wuho Is a New Member in Maintaining Genome Stability through its Interaction with Flap Endonuclease 1 | Replication forks are vulnerable to wayward nuclease activities. We report here our discovery of a new member in guarding genome stability at replication forks. We previously isolated a Drosophila mutation, wuho (wh, no progeny), characterized by a severe fertility defect and affecting expression of a protein (WH) in a family of conserved proteins with multiple WD40 repeats. Knockdown of WH by siRNA in Drosophila, mouse, and human cultured cells results in DNA damage with strand breaks and apoptosis through ATM/Chk2/p53 signaling pathway. Mice with mWh knockout are early embryonic lethal and display DNA damage. We identify that the flap endonuclease 1 (FEN1) is one of the interacting proteins. Fluorescence microscopy showed the localization of WH at the site of nascent DNA synthesis along with other replication proteins, including FEN1 and PCNA. We show that WH is able to modulate FEN1’s endonucleolytic activities depending on the substrate DNA structure. The stimulatory or inhibitory effects of WH on FEN1’s flap versus gap endonuclease activities are consistent with the proposed WH’s functions in protecting the integrity of replication fork. These results suggest that wh is a new member of the guardians of genome stability because it regulates FEN1’s potential DNA cleavage threat near the site of replication.
| Accurate genome replication is essential for the transmission of genetic information, and this process is vulnerable to factors that can induce DNA damage or block replication. To avoid this and achieve genome stability, many enzymes need to coordinate their activities at the replication forks (the area where the two DNA strands separate to replicate), thereby ensuring high efficiency and fidelity. Replication occurs simultaneously in both strands, but DNA polymerases only work in one direction (5′-to-3′), and while synthesis of one strand proceeds continuously, synthesis of the other must occur via small DNA fragments that need to be joined together after the removal of their 5′ ends by the flap endonuclease 1 (FEN1). In addition, FEN1 has a gap endonuclease activity that can potentially introduce strand breaks at the sites of replication. Here we report the identification of a new member of this complex of proteins called Wuho (WH) that has a role in guarding the genome stability at the replication forks. Down-regulation of Wuho in tissue culture cells results in double-strand DNA breaks and programmed cell death. Genetic and biochemical results indicate that WH interacts with FEN1 and regulates its activities at the replication fork by promoting its flap endonuclease activity while dampening its gap endonuclease activity. These results suggest a critical role of Wuho in protecting the integrity of replication forks.
| Faithful DNA replication is important for both the proliferation of cells and transmission of genetic information. This key biological process is vulnerable to many impediments, including numerous intrinsic and extrinsic agents that can damage DNA and create blockage for the movement of enzymatic machinery at the replication forks. At this fork, there are two separate but well-coordinated DNA biosynthetic activities, one for each template strand of parental duplex. Because of the 5′-3′ unidirectional activity of DNA polymerases, the leading strand synthesis can proceed without any interruptions after the initiation, while the lagging strand synthesis is intermittently halted for multiple RNA priming events and results in generating discontinuous nascent DNA, the so-called Okazaki fragments [1,2]. The ligation of the lagging strands requires the removal of their 5′-RNA primers, and this remodeling process in eukaryotes is usually carried out by the flap endonuclease 1 (FEN1) [3–5].
The bacterial FEN1 homologue was characterized as an integral part of Escherichia coli DNA polymerase I [6,7]. In eukaryotic organisms, FEN1 was first cloned and characterized from mouse cells by Harrington and Lieber [8], and it forms a ubiquitous and conserved family of enzymes. FEN1 is a structure-specific endonuclease capable of removing DNA or RNA flaps branching out from duplex DNA [4,9]. In addition to its important role in remodeling nascent lagging strands prior to their ligation, FEN1 is involved in repairing DNA damage in the pathway of long-patch base-excision repair [10]. Interestingly, FEN1 is known to possess a gap endonuclease activity, which may be functional in genetic recombination but also poses a potential threat for replication forks [11]. The critical biological functions of FEN1 are evidenced by the observation that homozygous fen1 knockout mice are early embryonic lethal [12]. The heterozygous fen1 mutation also promotes rapid tumor progression in a cancer-prone mouse model [13]. FEN1 is known to be under both positive and negative regulations. There are a number of proteins shown to interact with FEN1, including PCNA, 9-1-1, Blm, and Wrn helicases, and they are able to stimulate the FEN1’s in vitro endonucleolytic activities [14–17]. FEN1 is under multiple post-translational modifications that can either inhibit its enzymatic activities or promote its degradation [18–20]. We describe here the discovery of a new FEN1-interacting protein, Wuho, which has an essential role in genome stability, possibly through its ability to modulate FEN1’s activities depending on the structure of the DNA substrate.
We previously identified wuho (GeneID: 31566; protein accession: NP_572307.1) through isolating a Drosophila mutant deficient in its expression and demonstrated that it has a sterile phenotype (wuho means no progeny in Chinese, abbreviated as wh) [21]. Wuho (WH) belongs to an evolutionarily conserved family of proteins (Fig 1A), known as TRM82 in yeast (GeneID: 851743; protein accession: NP_010449.1) and WDR4 in humans (GeneID: 10785; protein accession: NP_387510.1) and mice (GeneID: 57773; protein accession: NP_067297.2). It is characterized by the presence of multiple WD40 repeats and can form a disc-like structure with seven β-propeller blades [22]. WD40 proteins are known for their function in mediating the formation of macromolecular complexes important for multiple cellular processes [23]. WH’s homologue in yeast, TRM82, is the non-catalytic subunit for heterodimeric m7G46 tRNA methyltransferase with the catalytic subunit known as TRM8 [24]. However, it is unclear that the tRNA methylase activity of TRM8/TRM82 complex has any essential cellular functions since mutations in either gene do not affect yeast viability [24]. On the other hand, our work here demonstrates that wh has an essential function in mice and that it has a conserved critical role in maintaining genome stability in metazoans, likely through WH’s function of regulating FEN1’s enzymatic activities.
In Drosophila, wh has an important function in the growth and differentiation of germline cells [21]. To probe the cellular functions of wh in mammalian cells, we utilized small interference-RNA (siRNA) to knockdown WH expression in mouse JB6 and human HFW cells. Compared with control cells and cells treated with non-targeting control RNA, siRNA specific for wh can reduce WH expression in both JB6 and HFW cells (Fig 1B). Depletion of mouse WH (mWH) by simWH and human WH (hWH) by sihWH results in greatly reduced viability of JB6 and HFW, respectively (Fig 1C). The loss of viability is likely due to a failure in maintaining genome stability. The depletion of WH induces DNA strand breaks as shown by γ-H2AX signals detected by western blot (Fig 1D) and also by comet assay in a neutral buffer aiming to detect double strand breaks (Fig 1E). We probed the biological functions of wh in Drosophila S2 cells and showed that WH knockdown results in loss of viability, DNA damage, and apoptosis (S1 Fig). These experiments thus demonstrate that wh has an evolutionarily conserved function, in the metazoan cultured cells, to maintain genome stability.
To demonstrate the relevance of DNA damage induced by the depletion of WH to the loss of viability, we examined whether these cells were affected by apoptosis. By agarose gel electrophoresis of the isolated genomic DNA, we showed that depletion of hWH results in degradation of nuclear DNA in HFW cells and the generation of the characteristic fragmentation pattern of DNA ladders (Fig 2A).
We investigated the signaling pathway connecting DNA strand breaks to apoptosis after hWH knockdown. An initial signaling event due to double-strand breaks is expected to involve the activation of the ATM/Chk2 pathway [25,26]. This is confirmed by our observation that ATM activation through Ser1981 phosphorylation and Chk2 phosphorylation at Thr68 occurred following hWH depletion, while the ATR/Chk1 pathway was not affected (Fig 2B). A key regulatory molecule, p53, coordinates DNA damage signaling and subsequent cell cycle arrest and apoptosis [27]. hWH knockdown leads to the phosphorylation of Ser15 in p53, which is known to play a critical role in its pro-apoptotic activity [28,29].
Activation of p53 can lead to apoptosis through both transcription-dependent and–independent pathways to trigger the release of the Bcl2 family of proteins from mitochondria [30–32]. The subsequent release of cytochrome c can lead to the activation of Caspase-9 and the downstream Caspase-3 and cleavage of PARP (reviewed by [33,34]). The activation of the mitochondria-based Caspase-9 pathway was observed following hWH depletion, while the alternative Caspase-8 pathway was not activated (Fig 2C).
The notion that hWH has a critical function in preserving genome stability and cell survival is thus supported by the above data that depletion of WH leads to DNA strand breaks, ATM/Chk2/p53 activation and Caspase-9 initiated cell death (Fig 2D). Such a function of hWH is not restricted to human cells only; we have used mouse JB6 cells to demonstrate that the knockdown of mouse WH (mWH) results in the same pathway of DNA damage signaling and cell death (S2 Fig), which suggests a conserved function of WH in mammalian cells.
The results presented so far have showed that WH depletion resulted in DNA damage and cell death, but the order of these two events was not addressed here. It is possible that WH knockdown can promote apoptosis, and that DNA fragmentation associated with programmed cell death generates the damage signals observed here. To investigate the direct consequence of the removal of WH, we examined both the time course of appearances of key molecules and the effects of specific Caspase inhibitors on these molecules. Following the treatment of mWH siRNA (simWH), γ-H2AX appears before the activated Caspases and the cleavage product of PARP (Fig 3A), suggesting that DNA damage precedes the cell death program. This notion is further reinforced by results from treatment with Caspase-specific inhibitors. If apoptosis induces DNA damage signaling, Caspase inhibition should stall the appearance of γ-H2AX. With the addition of inhibitors for Capsase-8, -9, or -3 (Z-IETD-FMK, Z-LEHD-FMK, and Z-DEVD-FMK, respectively), the level of γ-H2AX was unaffected (Fig 3B). It is interesting to note that inhibition of Caspase-9 or -3 extinguishes the apoptosis response (no cleavage of PARP), while the addition of the Caspase-8 inhibitor has no effect, further confirming that apoptosis goes through the Caspases 9 and 3 pathway (see Fig 2D). These results suggest that WH plays an important role in genome stability and that its depletion leads to DNA breaks and cell death.
Since the tumor suppressor protein p53 has a key role as a gatekeeper in maintaining genome stability by regulating the choice between cell death and cell cycle arrest upon genome damage (reviewed by [35]), we were interested in testing whether p53 plays a central role in directing the hWH-knockdown cells to apoptosis. We used a pair of isogenic cell lines with different p53 status, HCT116 p53+/+ and HCT116 p53-/- [36], to test their response following hWH knockdown. Treatment with sihWH results in a similar reduction of hWH by 80% in both cell lines and induces DNA breaks as evidenced by appearance of γ-H2AX (Fig 4A). But only the cells with wild-type p53 status showed reduced cell viability (Fig 4A), and this was linked to the appearance of apoptotic markers including cleaved Caspase-3 and PARP (Fig 4B). To further confirm that their difference in entering apoptotic program is primarily determined by p53 status, we reversed p53 status by using siRNA knockdown in HCT116 p53+/+, and introduced a DNA vector for ectopic p53 expression in HCT116 p53-/-. After the manipulation to alter p53 status, we observed that the appearance of apoptotic markers and reduction in viability were reversed consequentially (Fig 4C and 4D). These results plus those showing p53 activation following hWH depletion suggest that hWH affects genome stability and that p53 has a key role in determining the cellular response to the genotoxic consequence resulting from hWH knockdown.
Given that siRNA knockdown experiments suggest that wh has a critical function in genome maintenance and cell growth in mammalian cells, we sought to test its function in an organismic context. We constructed a knockout vector to delete the second and third exons in the mWh gene (Fig 5A). We examined 110 progeny from a cross of heterozygous mWh knockouts and found no pups that were homozygous mWh nulls (Table 1). But the pups of the heterozygous mutant and those of the homozygous wild type followed an expected 2:1 ratio. Upon dissecting the ovaries of pregnant females, we could identify mWh null by genotyping with PCR (Fig 5B), and these embryos can be detected up to day 10.5 (Table 1). Western blot experiments using embryo samples showed that mWH is depleted in the null and the heterozygous has about half of the wild-type amount (Fig 5C). Interestingly, we could detect both DNA damage and apoptotic signals, γ-H2AX and cleaved PARP, in mWH null embryos (Fig 5C). This suggests the lethality of mouse embryos follows a similar pathway to that of cultured cell lines with WH-knockdown. In addition, some of the null embryos at day 9.5 showed a resorbed morphology. There are apparent morphological differences in null embryos at day 10.5 and they showed varying aberrations, possibly due to different degrees of resorption. E10.5 nulls can be severely resorbed or have abnormality of brain development and internal bleeding (Fig 5D). The genetic experiments with Drosophila and mice, as well as the siRNA knockdowns with tissue culture cells, demonstrated the critical role of wh in cell growth and development, possibly through involvement in maintaining genome stability.
To probe the molecular mechanism by which hWH participates in maintaining genome stability, we sought to determine its interacting partners by co-immunoprecipitation (coIP) and analysis with mass spectrometry. From a human 293 cell line with inducible expression of hWH tagged with V5 and hexahistidine epitopes, we isolated the nuclear extract and incubated it with agarose beads conjugated with V5 monoclonal antibody. The bead-bound proteins were eluted, resolved by polyacrylamide gel electrophoresis, and analyzed by Mass Spectrometry (Fig 6A). Three of the identified proteins may provide information regarding hWH’s functions: METTL1 (Methyl Transferase-like Protein 1, human homolog of the yeast tRNA methyltrasferase subunit TRM8), FEN1, and PCNA. The identity of these proteins was also confirmed by western blot of the bead-bound proteins (Fig 6B).
The WH homolog in yeast is TRM82, the non-catalytic subunit of tRNA methyltransferase, that forms a heterodimer with the catalytic subunit TRM8 [24]. Human homologs of TRM8 and TRM82 can form a complex that has tRNA methyltransferase activity [24,37]. However, neither TRM8 nor TRM82 is essential for yeast growth [24,38]. Results presented in the next section also suggest that tRNA methyl transferase activity is not relevant to genome instability when hWH is depleted.
FEN1 and PCNA can form a complex and function at the replication fork to remove the RNA primers and remodel Okazaki fragments for facilitating the religation of the lagging strands [5]. Since earlier research had no indications that either one can interact with WH, we sought to establish their biochemical interactions. Whereas hWH partners were identified here through coIP with an ectopically expressed and tagged hWH, we carried out coIP experiments with endogenous hWH and FEN1 from HCT116 p53+/+ cells to verify their interactions (Fig 6C). These results demonstrate that hWH/FEN1 interaction does not depend on the presence of an ectopically expressed protein with an epitope tag. To examine if the binding is through direct protein/protein interaction, we purified recombinant proteins of hWH tagged with GST and FEN1 with hexahistidine and used GSH beads and Ni(II)-NTA resin, respectively, to pull down their partners (Fig 6D). These results demonstrate that hWH and FEN1 can have direct protein/protein interaction without the need of a mediator partner. As an alternative approach to directly demonstrate the interaction between FEN1 and hWH, we monitored the fluorescence anisotropy change of a fluorophore-tagged hFEN1 due to its association with hWH (Fig 6E). The increase in anisotropy depends on the addition of hWH and is saturable at higher hWH concentrations, confirming that a specific linkage between FEN1 and hWH results in a change in rotational diffusion time and a higher anisotropy for the tagged protein. The anisotropy data can be used to determine a dissociation constant of 130 nM.
Similar pull-down experiments with purified recombinant hWH and PCNA, however, failed to demonstrate their interactions (S3A and S3B Fig). But FEN1 is known to interact with PCNA [39], and the interaction of PCNA with FEN1 and many other replication proteins is through the PIP (PCNA-Interacting Protein) motif [40]. Since most of the PCNA partners possess the PIP motif [41], and hWH does not have one, it is not surprising that we could not observe a direct interaction between hWH and PCNA. It is possible that hWH associates with FEN1/PCNA complex through its binding to FEN1, and functions at the replication forks in facilitating the maturation of Okazaki fragments. Indeed, we can demonstrate that the pull-down of PCNA by hWH only occurs in the presence of FEN1 (S3B Fig).
To further probe hWH’s function in genome stability, we investigated its intracellular localization, especially with respect to the sites where DNA synthesis was ongoing. Both FEN1 and PCNA are localized at the site of newly synthesized DNA [42]. We therefore stained HCT116 p53+/+ cells for hWH, FEN1, PCNA, and EdU (5-ethynyl-2′-deoxyuridine), a thymidine analog, marking nascent nucleotide incorporation during DNA replication (Fig 7A). There were over 95% overlaps of the pixels from the images of FEN1, PCNA, and EdU when compared with hWH. The results showed that these three proteins were co-localized with DNA replication sites, suggesting that hWH is involved in DNA replication.
Besides confocal fluorescence microscopy, we also utilized a super-resolution imaging technique dSTORM, direct stochastic optical reconstruction microscopy [43], to verify the co-localization pattern between hWH and EdU. The dSTORM system is based on total internal reflection fluorescence (TIRF) microscopy. Therefore, we can obtain both TIRF and dSTORM (super-resolution) images from the same sample, reinforcing the validity of dSTORM result (Fig 7B, upper and lower panels, respectively). The TIRF and dSTORM images demonstrated co-localization of hWH and EdU. The higher magnification inserts of these images further support the co-localization pattern even at the super-resolution level (Fig 7B, lower panels). The localization of Wuho at replication loci is also shown recently with iPOND using the method of chemical cross-linking and proteomic analysis (listed as WDR4 in Table S6 in [44])
The notion that hWH has a function to protect DNA integrity at the replication fork predicts that upon hWH knockdown, DNA damage should occur at the site of replication. We tested this hypothesis by monitoring the localizations of replication and damage sites marked by EdU and γ-H2AX, respectively, during the time course following hWH knockdown (Fig 8). We observed that γ-H2AX staining colocalizes with EdU (the overlap of EdU to γ-H2AX is 92.9%, and the reciprocal of γ-H2AX/EdU is 5.9%), consistent with the idea that hWH’s genome stability function is at replication sites (Fig 8A). The γ-H2AX staining begins to appear at 48-h after hWH siRNA treatment, and it persists to 72-h (Fig 8B).
To further verify the role of hWH in protecting replication fork integrity, rather than having a direct role in DNA repair, we treated HCT116 p53+/+ cells with X-ray-irradiation and monitored hWH’s expression levels and nuclear distribution. We observed appearance of the DNA damage marker γ-H2AX in a time and dose dependent manner but with hWH’s expression levels remaining unchanged (S4A Fig). We also examined hWH’s intracellular localization to investigate whether hWH colocalizes with DNA damage loci. Immunostaining results show that hWH’s localization is independent of irradiation-induced γ-H2AX signals (S4B Fig). These results thus show that Wuho localizes at and protects replication forks, but it is not directly linked to DNA repair process.
The association of hWH with FEN1/PCNA, its localization at the site of DNA synthesis, and its knockdown resulting in DNA strand breaks suggest a plausible mode of action for hWH’s role in genome stability. The nuclease activity of FEN1 while necessary for removing RNA primers and DNA damage, also presents a possible threat to genome integrity especially near the replication forks. If hWH can modulate the nuclease activities of FEN1, it may thereby mitigate a potential hazard posed by FEN1. This would also have the implication that the DNA damage following the reduction of hWH expression is mediated through FEN1. We tested this hypothesis by showing that while knocking down FEN1 only minimally affects cell viability, double knockdown of hWH and FEN1 improves the viability in comparison with hWH knockdown alone (Fig 9A). Moreover, we tested three separate sihFEN1 knockdown siRNAs, and they all were able to ameliorate the viability loss due to hWH knockdown. The rescuing ability is correlated with the knockdown efficiency with the one less capable of reducing FEN1 expression (sihFEN1-3) having lower viability restoration (lanes 6 and 7 versus 8, Fig 9A). It is also interesting that double knockdown of hWH and FEN1 mitigates the DNA strand breaks or the levels of γ-H2AX relative to hWH knockdown alone (Fig 9A). This phenomenon of synthetic rescuing can be replicated in both human (Fig 9A) and mouse cells (S5 Fig). Interestingly, the rescuing effect by FEN1 knockdown can be reversed by ectopic expression of FEN1 (lanes 6 versus 8, Fig 9B). With FEN1 overexpression under the conditions of both ectopic and endogenous expression, knockdown of hWH becomes even more cytotoxic (lane 7). These data are consistent with the notion that hWH is able to modulate FEN1 nuclease activities and that a reduction in hWH’s expression presents a vulnerability for the unchecked nucleolytic action of FEN1.
Both our work and earlier work [24] suggest hWH has another interacting partner, METTL1, the catalytic subunit of tRNA methyltransferase. We therefore found it important to address whether METTL1 has a role in hWH’s genome stability. In contrast to hWH, knocking down METTL1 has no effects on DNA strand breaks or cell viability (S6A Fig). Furthermore, double knockdown of hWH and METTL1 does not improve either genome instability or cell viability relative to single hWH knockdown. That METTL1 is not relevant to genome stability is supported by its intracellular localization. In contrast to hWH, the nuclear localization of METTL1 is not coincident with loci of nascent DNA synthesis, as evidenced by confocal microscopy (S6B Fig). hWH thus has two intracellular functions, one associated with METTL1 as a partner for tRNA modification and the other involved in controlling FEN1 activities at the replication fork.
We employed biochemical approaches to directly test the hypothesis that hWH modulates FEN1’s nuclease activities. FEN1 can remodel Okazaki fragments through its 5′ endonuclease activities on either single or double flap substrates [5]. Interestingly, FEN1 possesses a low level of gap endonuclease activity at replication forks, a potentially hazardous activity for an enzyme located at the site of DNA synthesis [11]. FEN1’s gap endonuclease activity also can target double strand DNA with single strand region. Besides endonuclease activity, FEN1 has 5′ exonuclease ability. To test hWH’s effect on FEN1’s various nucleolytic functions, we generated five model substrates using fluorophore-tagged oligonucleotides and determined FEN1’s nuclease activities with and without hWH (Fig 10A). While hWH does not affect FEN1’s exonulease activity in nicked substrate, it exerts distinct effects on FEN1’s activities on the other DNA substrates. FEN1 displays the highest activity toward double flap substrate over other substrates, and for FEN1’s gap endonuclease activity on the Y-shape substrate, it is more active toward the lagging strand than the leading strand. These results are consistent with what were reported earlier [11,45,46]. Intriguingly, hWH stimulates FEN1’s flap endonuclease activity in single and double flap substrates but inhibits FEN1’s gap endonuclease in Y-shape and other gap substrates (Fig 10A). Either promotion or repression of these activities exhibits a dose dependence on hWH. These assays reported above were analyzed with DNA gel electrophoresis under denaturing conditions. We have also carried out assays using gap and Y-shape substrates but with samples processed for analysis using gel electrophoresis under non-denaturing conditions and obtained simialr results (S7 Fig). While the effects of hWH are most apparent at the highest concentration tested, 1 mM, representing a molar ratio in a range of 20- to 1,000-fold over FEN1, we could clearly observe statistically significant changes at much lower hWH concentrations as well. We could detect the stimulation of the flap endonuclease activity using single flap substrate with hWH at a 20-fold ratio over FEN1, and in a marked contrast, inhibition of gap endonuclease activity with Y-fork substrate at 2-fold molar ratio. These diametric effects of hWH on FEN1 depending the substrate structures are consistent with the biological functions of hWH proposed here.
While the biochemical basis for how hWH can regulate FEN1’s activities remains to elucidated, it is likely that the direct association of hWH with FEN1 as shown here affects its nucleolytic activities. But it remains possible that hWH can directly binds DNA with differential affinities. To test this possibility, we monitored binding of hWH with either double flap or Y-shape substrate with electrophoretic mobility shift assay. FEN1 can bind both DNA as reported in a previous study [45]. On the other hand, hWH does not show significant affinity with either (S8 Fig). Our results here suggest that hWH can interact with FEN1 and regulate its activities in a structure-specific manner, thereby providing a guardian role for FEN1’s biologically relevant functions at the replication fork (Fig 10B).
In Drosophila, wh mutations affect the growth and development of germline cells and result in male sterility and a greatly reduced fertility in female flies [21]. Because of maternal storage of WH protein in eggs, zygotic mutations of wh affect embryogenesis to a lesser extent, and some of the wh nulls can survive to adulthood. In an interesting contrast, we showed here that mouse homozygous nulls are embryonic lethal at days 9.5–10.5, demonstrating that the critical function of wh in cellular growth and development is evolutionarily conserved.
To further elucidate the mechanistic basis of wh’s essential functions, we developed a system for using siRNA to knockdown WH in tissue culture cells. WH knockdown in Drosophila, mouse, and human cells resulted in reduced viability and apoptosis. These cellular defects are due to double strand breaks, and through the signaling pathway of ATM/Chk2/p53 lead to Caspase-9-activated cell death. While programmed cell death will ultimately result in DNA fragmentation, our data showed that DNA cleavage due to WH knockdown is upstream to apoptosis, not caused by it. The time-course experiments following WH knockdown showed that DNA damage signals preceded Caspase activation. Using Caspase-specific peptide inhibitors, we showed that they mitigate Caspase cleavage/activation and PARP breakdown but not affecting DNA damage induced by WH-knockdown. While inhibitors for Caspases-9 and -3 blocked PARP cleavage, Caspase-8 inactivation had no effect on PARP processing. This result further supports the proposed pathway that apoptosis is mediated through Caspases-9 and -3, by the mitochondria-mediated intrinsic signaling mechanism, and not through the extrinsic signaling that would follow Caspase-8 activation. The apoptotic response emanating from genome instability is usually mediated through a p53-controlled and mitochondria-mediated pathway with ensuing activation of Capases-9 and -3 [33,34]. Therefore, WH-knockdown can lead to DNA strand breaks and result in programmed cell death through p53 activation.
To gain additional insight into the molecular mechanism by which WH can protect genome stability during the growth and development of an organism, we endeavored to examine its interacting partners by immunoprecipitation and Mass spectrometry. We identified a known interacting protein, METTL1, the catalytic subunit of tRNA methyl transferase, which heterodimerizes with hWH (also known as WDR4). However, it is unclear that WH’s function in genome stability is related to the activity of tRNA methyalse, since the yeast mutants defective in either subunits (mutations in TRM8 and TRM82) are viable and have no dramatic phenotypes [24,38]. Our attention was directed to the identification of FEN1 and PCNA in the immunoprecipitates since both proteins are known to be important in genetic stability [4,41]. Subsequent biochemical experiments with purified proteins demonstrated that FEN1/hWH are capable of associating with each other. There is no direct affinity between hWH and PCNA. The association of PCNA is likely through FEN1 since PCNA and FEN1 are known interacting partners [39] and such interactions have important functions in the maturation of precursors of nascent lagging strand near replication forks [3]. Interestingly, the presence of hWH at the site of DNA replication along with FEN1/PCNA is also supported by the co-localization results using immunofluorescence microscopy.
FEN1 is a structure-specific nuclease that can remove extrahelical 5′ DNA or RNA flaps; its function is essential for replication and repair [5]. While the presence of FEN1 at or near the replication forks is critical for the maturation of the newly synthesized Okazaki fragments, it also poses a potential threat for replication forks since FEN1 can cleave at the fork junctions where trihelical segments intersect, and generate double strand breaks [11]. Our biochemical result that hWH can differentially modulate FEN1’s activities depending on the structure of DNA substrate, suggests a plausible function of hWH in protecting genome stability. hWH is able to suppress FEN1’s endonucleolytic cleavage of fork and gap DNA substrates relative to its activities toward flap substrates, thereby minimizing the wayward activity of FEN1 while still maintaining its functionally important ability to process the nascent lagging strands. In our activity assays, we observed hWH modulating FEN1’s multiple nucleolytic activities with hWH concentrations up to 1 mM. In the intracellular environment, hWH/FEN1 may cooperate with other proteins as components in a macromolecular complex at the replication fork. Therefore there may be more effective interactions between hWH and FEN1 to regulate FEN1’s activities in the replication machine. Alternatively, hWH may possess post-translational modifications as yet to be characterized to enable hWH to interact more efficiently with FEN1. In contrast to other interacting proteins that can either promote or inhibit FEN1’s activities, hWH has diametric effects on the gap versus flap endonuclease activities, which are consistent with the proposed functions of hWH in protecting fork integrity. This proposed function of hWH in genome stability is further supported by the RNA knockdown experiments of hWH and FEN1. While the addition of siRNA for FEN1 results in negligible effects on cell viability and genome integrity, double knockdown of hWH and FEN1 by siRNA can partially mitigate loss of cell viability and DNA damage caused by hWH knockdown. This result thus suggests that FEN1 is at least partially responsible for DNA damage caused by reducing the intracellular WH levels.
WH is a member of an evolutionarily conserved family of proteins with multiple WD40 repeats, which has its homologs in yeast known as TRM82 and in mammals as WDR4 [21,24,47]. Proteins with WD40 repeats usually have a disk-like β-propeller structure with multiple blades composed of anti-parallel β-sheets [48]. WD40-repeat proteins are known mediators in the assembling of protein complexes important for various cellular functions [49]. Besides FEN1, WH’s homologs are known to interact with the catalytic subunit of tRNA methyl transferase, TRM8 in yeast and METTL1 in mammals [24,37]. The crystal structure of TRM8/TRM82 complex has been solved, revealing that TRM82 has a β-propeller structure of seven blades with the edges of two of the blades contacting TRM8 [22]. It is uncertain whether WH’s interaction with tRNA methylase has an important role in its genome guardian function. However since yeast mutants of trm8 or trm82 do not have any serious phenotypes and, as we have shown here that siRNA knockdown of METTL1 does not affect cell viability and genome integrity, tRNA methylation through WH’s function is likely not a major contributor to genome stability.
FEN1, on the other hand, is known to be an essential protein for DNA replication and repair, and has specific functions in removing RNA primer on nascent lagging strands near replication forks and in long patch base excision repair [5]. It therefore comes as no surprise that there are a number of proteins discovered to date that interact with FEN1, including PCNA, 9-1-1, Werner syndrome protein (WRN), and Bloom’s syndrome protein (BLM), and that these partner proteins can stimulate FEN1’s flap endonuclease activity about 5–11 fold [14,17,39,50,51]. Besides protein/protein interactions, FEN1 is also subject to multiple cell cycle-specific post-translational modifications [20]. Our results presented here demonstrated WH as a new interacting partner with FEN1 and that their association can modulate the structure-specific endonuclease activities of FEN1. Such regulatory functions of FEN1 may have a critical role in the growth and development of a multicellular organism and in the maintenance of its genome stability.
All the animals utilized in this study were maintained in a specific pathogen free environment under the guidelines of Academia Sinica Institutional Animal Care and Use Committee.
All mammalian cell lines including JB6 mouse epidermal cells, HFW human fibroblast cells, and HCT116 p53+/+ and HCT116 p53-/- cells were maintained in Dulbecco’s Modified Eagle Medium (DMEM) supplemented with 10% heat-inactivated fetal bovine serum (FBS), 100 unit/ml penicillin, and 100 μg/ml streptomycin at 37°C in a humidified incubator with 5% CO2. Drosophila melanogaster Schneider 2 (S2) cells were maintained in Schneider’s medium supplemented with 10% heat-inactivated FBS, 100 units/ml penicillin, and 100 μg/ml streptomycin at 28°C. All reagents used in cell culture were from Invitrogen (Grand Island, NY, United States).
We synthesized peptides LKKKRQRSPFPGSPEQTK from protein sequences of mouse Wuho and DGHAKKMRPGEATLSC from human Wuho, and used them to immunize rabbits. The antibodies were purified by affinity chromatography with peptide antigens before being used for western blot or immunofluorescence.
Small interference RNA (siRNA), both gene-specific ON-TARGET and controls Non-targeting pool siRNA, were purchased from Dharmacon (Chicago, IL, US), and used for knockdown experiments following manufacturer’s protocol. Single siRNAs against human or mouse FEN1 were purchase from Ambion (Grand Island, NY, US). The sequences for all the siRNA used in this work are listed in S1 Table.
Cellular viability was determined by MTT assay using a water-soluble reagent WST-8 (Dojindo, Kumamoto, Japan).
DNA laddering in apoptotic cells was used as a cell death assay. DNA fragments were purified by Suicide Track Kit (Millipore, Billerica, MA, US), and analyzed by electrophoresis in 2% agarose gels.
DNA double strand breaks were monitored by comet assay (single cell gel electrophoresis in neutral conditions) as described before [52]. Migration of fragmented DNA escaping from the nucleus in each cell was measured with the COMET Assay III program (Perspective Instruments, Suffolk, United Kingdom) and expressed by the parameter of the tail moment (a product of the tail length and intensity).
Knockout of the mWh gene was carried out by the Cre-loxP system with a targeting construct to delete exons 2 and 3. The backbone of the targeting vector, which harbors a mWh genomic region including exon1, 2, and 3, was subcloned from a BAC clone RPCI23.C (Invitrogen) by BAC Subclone Kit (K003, GENE BRIDGES). The insertions of one loxP site between exon1 and 2 and one PGK-neo cassette (positive selection marker) flanked by two loxP sites behind exon3 were achieved by the recombineering method using Quick and Easy Conditional Knockout Kit (loxP/Cre) (K005, GENE BRIDGES). A PGK-DT (diphtheria toxin, negative selection marker) cassette was inserted into the vector at the EcoRV and SalI sites. The resulting construct was linearized by SnaBI prior to electroporation into C57BL/6J ES cells. After screening by neomycin treatment, targeted ES cell clones were microinjected into C57BL/6-C2J blastocysts for generating chimera mice. The male chimera mice were then crossed with wild-type C57BL/6-C2J females. The black pups with targeted allele were crossed with EIIa-Cre mice [53] to generate mWh heterozygous mice (+/-) carrying one mWh null allele, through deleting exon2 and 3 between the first and the third loxP sites. Homozygous deletion mice (-/-) were generated by the intercross of +/- mice.
After marking littermates by ear notching, genomic DNA was isolated from ear tissue for genotyping. Genotyping was performed by PCR with two primers (ACCACGAGCCTAGAGGATCAGTGGC and TTGTCTGTCTGTGGGAGGGCCTGAG), which can identify wild-type and null alleles by amplifying 3.6 kb and 0.5 kb fragments, respectively.
A site-specific recombination system was used to generate a stable cell line of 293T with an integrated gene for hwh under a tetracycline-inducible promoter. A plasmid vector containing pcDNA5/FRT/TO (Invitrogen) was inserted with hwh cDNA flanked with V5 and hexahistidine epitope tags, and the resulting construct was used to transfect Flp-In-293 cells (Invitrogen) to allow for integration of hwh cDNA into the genomic FRT site. Clonal selection of stable recombinants was carried out with 300 μg/ml hygromycin and 10 μg/ml blasticidin in the media. Cells were induced with 1 μg/ml tetracycline for 24 h, and harvested cells were resuspended in a hypotonic buffer of 5 mM KPO4 (pH 7.8), 2 mM MgCl2, 1 mM EDTA, and protease inhibitor cocktail (Roche, Basel, Switzerland). Cells were lysed by douncing and nuclei were collected by centrifugation at 3 kg for 5 min. The nuclear pellet was suspended with nuclear extraction (NE) buffer containing 20 mM Hepes (pH 7.9), 400 mM NaCl, 10% glycerol, 1 mM EDTA, 0.2% NP-40, and protease inhibitor cocktail. The nuclear extract was incubated with 50 μl mouse anti-V5 agarose beads (Sigma, St. Louis, MO, US) with rotation at 4°C overnight. In parallel, nuclear extract of control Flp-In-293 cells transfected with vector DNA was processed under identical conditions. The beads were washed 5 times with NE buffer prior to boiling in SDS-PAGE sample buffer without any thiol reagents. The sample with eluted protein was replenished with 5% β-mercaptoethanol and boiled again before electrophoresis in a 4%–12% SDS-polyacrylamide gel. Following Coomassie Blue staining, protein bands were sliced and identified by Mass Spectrometry.
FEN1 was purified using a published protocol [54] with a construct kindly provided by Dr. Robert Bambara. For purification of hWH, we made a vector with hWH cDNA inserted into pET23b (Novagen, Madison, WI, US) carrying a hexahistidine tag at its C-terminus, and transformed it into
BL21(DE3)pLysS. Expression of hWH was induced by 1 mM IPTG at 37°C for 4 h. The cell pellet was lysed with a buffer containing 20 mM Tris pH 8.0, 300 mM NaCl, 5 mM β-mercaptoethanol, 10% glycerol, 20 mM imidazole, and protease inhibitor cocktail. The soluble fraction was applied to the HisTrap FF Crude column (GE, Waukesha, WI, US), and hWH was eluted with 200 mM imidazole. The pooled factions were applied to a Hitrap Q FF column (GE), and eluted by a NaCl gradient from 0.05–1 M with hWH peaked at 0.3 M NaCl. The pooled fractions were dialyzed in a buffer containing 20 mM Tris pH 8.0, 0.3 M NaCl, 5 mM β-mercaptoethanol, and 50% glycerol, and stored at -30°C.
For co-immunoprecipitation (co-IP) with cell lysates, HCT116 p53+/+ cells were lysed in RIPA buffer plus protease inhibitor cocktail. Cell lysate (1 mg/ml in protein) was pre-cleared by incubating with 50 μl of PureProteome Protein A Magnetic Beads (MILLIPORE) at 4°C for 1 h. The cleared lysates were mixed with magnetic beads conjugated with an antibody against target protein, either rabbit anti-hWH antibody or mouse anti-FEN1 antibody (abcam, Cambridge, UK), and incubated with 0.1 U/ml micrococcal nuclease at 4°C overnight. The partner proteins bound to beads were detected by western blots.
The direct pull-down assay was carried out with hWH, FEN1, and PCNA with a C-terminal tag of GST, hexahistidine and FLAG, respectively. To detect FEN1 binding with hWH by fluorescence anisotropy, we constructed an expression vector for a fusion protein of hFEN1 with a C-terminal tag of tetracysteine (FEN1-CCPGCC) [55]. The purified protein was incubated with 1mM TCEP for 30 min at 25°C prior to labeling with ReAsH reagent for 2 h at 25°C. The anisotropy of the fluorescence from FEN1-ReAsH was measured after adding different amounts of hWH; the binding data were used to determine the dissociation constant [56].
Prior to fixation with 4% paraformaldehyde, HCT116 p53+/+ cells were fed with a culture medium containing the thymidine analog, 5-ethynyl-2′-deoxyuridine (EdU) (Click-iT EdU Imaging Kit, Invitrogen), in 10 μM for 15 min. The fluorophore (Alexa488 azide) was linked to EdU by click reaction catalyzed by cuprous ions [57] marking the nascent DNA. Fixed cells were probed with rabbit anti-hWH, mouse anti-FEN1 (abcam), and goat anti-PCNA (Santa Cruz Biotechnology Inc., Santa Cruz, CA, US) antibodies. The fluorophore-conjugated secondary antibodies were conducted by incubation with Alexa555 anti-goat, Alexa594 anti-mouse, and Alexa647 anti-rabbit antibodies (Invitrogen). Images were collected with a TCS-SP5-AOBS-MP microscope (Leica, Wetzlar, Germany). The extent of image-overlaps was determined with Metamorph (Molecular Devices, Sunnyvale, CA, US).
We also used dSTORM (direct stochastic optical reconstruction microscopy) to map the EdU and hWH localization with super-resolution imaging [43]. Cells were pulse-labeled with EdU, and the fixed cells were labeled with rabbit anti-hWH antibody, followed by Cy5 anti-rabbit secondary antibody (Invitrogen). The samples were immersed in 100 mM mercaptoethylamine, PBS, before image collection by dSTORM.
PAGE-purified DNA oligonucleotides with or without fluorophore of FAM or Cy5 were obtained from PURIGO (Taipei, Taiwan). Their sequences are listed in S2 Table. Oligonucleotides were adjusted to 100 μM with 10 mM Tris-HCl pH 8.5. The fluorophore-labeled oligonucleotide was annealed to unlabeled oligonucleotides (S3 Table) in 10 mM Tris-HCl pH 8.5, 100 mM KCl, and 5 mM MgCl2, with temperatures slowly ramping from 95°C to 25°C at a rate of 0.01°C/sec. Annealed products were separated by polyacrylamide gel electrophoresis in Tris-borate-EDTA buffer (TBE-PAGE). DNA substrates were eluted by immersing gel slices in 0.5 M ammonium acetate at 4°C with rotation overnight. The sample was applied to an Illustra NAP-5 column (GE) and eluted with 10 mM Tris-HCl pH 8.5 with 100 mM NaCl. The concentrations of fluorophore-labeled DNA substrates were measured by TBE-PAGE with fluorophore-labeled DNA oligonucleotides as standards. Design of oligonucleotides is based on a previous reference [58], with modifications.
Fluorophore-labeled DNA substrates with concentration of 50 nM were incubated with FEN1 at 37°C for certain time (1 min for single-flap and double-flap, 5 min for nick, and 30 min for Y-shape and gap) in 20 μl reaction buffer containing 50 mM Tris (pH 8), 50 mM NaCl, and 5 mM MgCl2. Reactions were terminated by the addition of a 2x stop mix containing 95% formamide, 10 mM EDTA (pH 8), and 0.2% Orange G. Reaction products were denatured at 95°C for 5 min and separated by denaturing TBE-PAGE with 7M urea polyacrylamide gel and fluorescence images were captured and analyzed by ImageQuant LAS 4000 (GE). The assays on FEN1’s gap endonuclease activity in Y-shape and gap substrates were also examined by native TBE-PAGE.
Fluorophore-labeled double flap substrate and Y-shape substrate of 50 nM was incubated with either FEN1 or hWH (concentrations ranging among 25, 50, 100, 250, 500, and 1,000 nM) at 37°C for 10 min in 20 μl binding buffer containing 50 mM Tris (pH 8) and 50 mM NaCl. Incubation products were separated by TBE-PAGE and images were processed by ImageQuant LAS 4000 (GE).
|
10.1371/journal.pntd.0005899 | Cost-effectiveness of a national enterovirus 71 vaccination program in China | Enterovirus 71 (EV71) has caused great morbidity, mortality, and use of health service in children younger than five years in China. Vaccines against EV71 have been proved effective and safe by recent phase 3 trials and are now available in China. The purpose of this study was to evaluate the health impact and cost-effectiveness of a national EV71 vaccination program in China.
Using Microsoft Excel, a decision model was built to calculate the net clinical and economic outcomes of EV71 vaccination compared with no EV71 vaccination in a birth cohort of 1,000,000 Chinese children followed for five years. Model parameters came from published epidemiology, clinical and cost data.
In the base-case, vaccination would annually avert 37,872 cases of hand, foot and mouth disease (HFMD), 2,629 herpangina cases, 72,900 outpatient visits, 6,363 admissions to hospital, 29 deaths, and 945 disability adjusted life years. The break-even price of the vaccine was $5.2/dose. When the price was less than $8.3 or $14.6/dose, the vaccination program would be highly cost-effective or cost-effective, respectively (incremental cost-effectiveness ratio less than or between one to three times China GDP per capita, respectively). In one-way sensitivity analyses, the HFMD incidence was the only influential parameter at the price of $5/dose.
Within the price range of current routine vaccines paid by the government, a national EV71 vaccination program would be cost-saving or highly cost-effective to prevent EV71 related morbidity, mortality, and use of health service among children younger than five years in China. Policy makers should consider including EV71 vaccination as part of China’s routine childhood immunization schedule.
| Enterovirus 71 (EV71) has caused great morbidity, mortality, and use of health service in children younger than five years in China. Recently, effective and safe vaccines against EV71 have been approved. Whether EV71 vaccination should be included as part of China’s routine childhood immunization schedule is unknown. In this study, we built a decision model to evaluate the health impact and cost-effectiveness of a national EV71 vaccination program in China. We find that vaccination would annually avert 567,500 cases of hand, foot and mouth disease (HFMD), 40,000 herpangina cases, 1,093,500 outpatient visits, 95,500 admissions to hospital, 435 deaths, and 14,000 disability adjusted life years based on the current Chinese birth cohort size. The break-even price of the vaccine was $5.2/dose. Within the price range of current routine vaccines paid by the government, a national EV71 vaccination program would be cost-saving or highly cost-effective. Policy makers should consider including EV71 vaccination as part of China’s routine childhood immunization schedule.
| Enterovirus 71 (EV71) is one of the major agents that cause outbreaks of hand, foot, and mouth disease (HFMD) and herpangina worldwide[1]. Since the 1990s, the epidemic has mainly affected the Asia-Pacific region and EV71 has become a major public health issue across this region[2,3,4,5].
HFMD is characterized with fever and cutaneous lesions on hands, feet and buttocks, along with oral lesions. Although most cases are mild and self-limiting with an average duration of 7 days, approximately 1% can rapidly develop severe and even life-threatening complications such as encephalitis, aseptic meningitis, pulmonary oedema/hemorrhage and heart failure[1].
During the period from 2008 to 2012, China reported more than 7 million children with HFMD, of which around 45% were associated with EV71[6]. During the period from May 2008 to December 2014, China reported death of 2,225 children due to HFMD, with a case-fatality rate of 0.03% and 93% of them were associated with EV71 [6,7].
Current treatment is only to relieve symptoms. No specific drug to treat EV71 infection is available [1]. With limited impact of personal and environmental hygiene, vaccination is considered as the most effective and promising strategy to prevent HFMD and herpangina caused by EV71 [8].
Since 2013, three phase 3 randomized clinical trials (RCTs) to evaluate efficacy of inactivated EV71 vaccines in infants and young children have been completed in China [9,10,11]. The vaccines showed high efficacy and satisfactory safety to provide protection against EV71-associated diseases and are now available in China.
In 2010, before the key clinical trials were initiated, an cost-effectiveness analysis estimated economic value of a future vaccine against EV71[12]. Here, to assist policy makers in evaluating the implication of a national EV71 vaccination program in China, we reassessed the public health impact and cost-effectiveness of EV71 vaccination, using new evidence on the vaccine safety and efficacy as well as updated clinical and economic data on EV71 associated infections.
Using Microsoft Excel, a decision tree model was built to calculate the net clinical and economic outcomes of EV71 vaccination compared with no EV71 vaccination (Fig 1). This model adopted Markov chain and hypothesized a 2012 birth cohort of 1,000,000 Chinese children. As most affected cases are younger than five years and the rates of severe illness and mortality decrease substantially in older children and adults[6], the model’s time horizon was five years after birth. The time step was one year. If children experienced symptomatic infection of EV71, they died or suffered from one of the following diseases: herpangina, mild HFMD, and severe HFMD[6]. Patients with HFMD were categorized as severe if they had any neurological complications (encephalitis, aseptic meningitis, or flaccid paralysis), or cardiopulmonary complications (pulmonary hemorrhage, pulmonary oedema, or myocarditis), or both; otherwise, they were classified as mild cases[6]. According to experience in China, almost all cases with HFMD make outpatient visits before deciding to receive home care or to be hospitalized for further treatment; a small number of cases with mild HFMD, almost all cases with severe HFMD and almost all death cases are hospitalized; herpangina alone is not an indication for hospitalization. Life years lost after the 5 years were taken into accounted. Accordingly, the model simulated events over a 5 year horizon but accounted for outcomes over the total lifetime of the simulated individuals. The primary result was presented as costs per disability adjusted life year (DALY) averted.
The overall incidence of HFMD was 1.2 per 1,000 person-years from 2008 to 2012, varied among provinces ranging from 0.2 in Tibet to 3.1 in Hainan according to the Chinese Center for Disease Control and Prevention (China CDC)[6,13]. EV71 accounted for 45% of mild, 80% of severe, and 93% of fatal cases and these proportions did not vary significantly with age among children aged 5 and under[6]. Thus, we calculated the annual incidences of mild, severe, and fatal EV71-associated HFMD by age from the corresponding overall incidences of HFMD by age and the proportions associated with EV71. We used their average values of four years (2009–2012) in our base-case analysis (Table 1).
Herpangina has not been included in surveillance system in mainland China. The reporting of cases of herpangina and EV71 are aggregated together in Taiwan; specific data on the epidemiology of herpangina are not available. Fortunately, studies supplied information to calculate the ratio of patients with EV71-associated herpangina to that of EV71-associated HFMD. The ratio ranged from 0.044 to 0.11, with weighted mean of 0.069, using study sample size as the weight[11,14,15] (Table 1). The incidence of EV71-associated herpangina was calculated using this ratio and the incidence of EV71-associated HFMD from the China CDC[6].
Several studies reported the spectrum of complications of severe EV71-associated HFMD; nevertheless, they were single-center in design, had small-sized sample or short duration of enrollment, or the cases were selected[14,15,16,17]. Chen et al summarized hospitalized cases of EV71-associated HFMD in Taiwan from 1998 to 2005[18]. These cases were reported to surveillance systems at the Taiwan CDC by 538 hospitals of various levels. Based on them, the proportion of each complication was calculated for this analysis (Table 1).
The data from the largest pediatric infectious disease center in Shanghai between 2007 and 2010 showed a hospitalization rate of 14% for all 28,058 patients diagnosed as HFMD and 54% of the inpatients were positive for EV71[19]. The hospitalization rate of EV71-associated HFMD was calculated as: 0.14×(0.54/0.45) = 0.168, in which 0.45 represented the proportion of EV71 in all HFMD cases according to the China CDC[6] (Table 1). Another survey from Guangdong reported a similar hospitalization rate [20].
The frequencies of outpatient visits for each symptomatic case were not available specifically for EV71-associated HFMD and herpangina. The frequency for overall HFMD patients was used in this analysis (Table 1).
Recently, three multicenter, randomized, double-blind, placebo-controlled phase 3 trials evaluated the efficacy and safety of inactivated EV71 vaccines in healthy infants and young children in China[9,10,11]. The 1-year efficacies ranged from 90% to 97.4% against EV71-associated HFMD. We performed a meta-analysis using a random-effect model. The results showed an overall efficacy of 95%, with 95% confidence interval of 90%-98%.
One of the trials reported efficacy against EV71-associated herpangina[11]. However, due to sparse events, no significant result was reached. In the absence of other data, this analysis assumed that the vaccine efficacy against EV71-associated herpangina was the same as that against HFMD.
Extended follow-up of one trial showed that the antibody titers were maintained at a high level through two years post-vaccination [9,21]. There are no long-term results for the other two trials. However, one of them reported consistent titers from month 6 to month 12 post-vaccination [11] and the other one reported slightly waned titers at day 180 after vaccination [10]. Therefore, we assumed that the titers do not wane significantly in our model, of which the time horizon is just five years.
The rate of adverse events within 28 days after vaccination was 56% on average (range 47%-71%)(Table 1)[9,10,11]. Most of the adverse events of EV71 vaccines were mild. Serious adverse events, which were considered to be associated or most likely associated with vaccination, happened only in 0.04% of the participants (range 0–0.1%)(Table 1)[9,10,11].
The schedule of vaccination against EV71 was two doses, 4 weeks apart[9,10,11], given at 3 and 4 months of age [22]. As this schedule is the same as that for the first two doses of diphtheria, tetanus and pertussis (DTP) vaccine, DTP coverage was used to estimate EV71 vaccine coverage[22]. Due to the lack of data on the coverage of the second DTP dose, data on the third DTP dose was used. Data are limited to estimate the efficacy of a single dose of EV71 vaccine. It was assumed to be 50% in this analysis[22].
We estimated DALYs using 2010 life expectancy data of China[23]. DALYs are the sum of years of life lost (YLLs) and years of life lost due to disability (YLDs)[24].
The durations of herpangina and mild HFMD are both 7 days on average[25]. According to Xu et al, the mean duration of hospitalization of severe HFMD (including critical cases) was 16 days[26]. There is no data on the duration of disability after discharge from hospital. Therefore, the duration was underestimated. In the base-case analysis, we assumed no disability following discharge. In sensitivity analysis, we explored how its uncertainty influenced the cost-effectiveness results.
Disability weights (DW) for each condition were taken from the World Health Organization’s estimates and a previous cost-effectiveness analysis (Table 1)[12,27]. DWs for conditions with combined complications were not available. For simplicity, the highest DW was used if the patients suffered more than one complication.
This cost-effectiveness analysis was conducted from a societal perspective. The costs for EV71-associated HFMD and herpangina incorporated direct medical costs and non-medical costs for physician visits, medications, lab tests, and transportation, and indirect costs for work loss (S1 and S2 Tables). To today, seven studies have gathered these cost data from outpatient visits and hospitalizations in various regions of China[28,29,30,31,32,33,34]. The reported costs were weighted by the reported number of cases in each study to estimate average costs for each treatment setting (outpatient or hospitalization) (Table 1).
In China, vaccines are either supplied by commercial market or Expanded Program on Immunization (EPI). The latter is paid by the government. This analysis is to give an implication whether EV71 vaccines should be included in EPI in China. As the prices of vaccines in EPI are no more than $4.59 per dose, the analysis showed more concern for the case of $5.0 per dose (close to $4.59). As far as we know, recently EV71 vaccines have become commercially available in China and the price is around $30-$40 per dose, varied among regions. Right now the vaccines are paid by parents and the coverage is relatively low according to experiences from other commercial vaccines in China. The vaccine price may change in the future. Therefore, we performed the analysis at a range of prices for vaccines. Our analysis used the range from $2.5 to $40 ($2.5, $5, $10, $20, $30, $40) per dose because this range covers almost all prices of China made vaccines.
The price of vaccine administration was estimated at 3 Chinese Yuan (CNY) per injection according to subsidy policies to health facilities for vaccine administration (range 2–4 CNY, to cover the costs of nurse labor, syringe and transportation and storage of vaccine)[35]. The costs of vaccine-associated adverse events were considered in this analysis and they were obtained from published studies (Table 1)[12,36]. All costs were converted to 2012 US Dollars (1 US Dollar = 6.30 CNY) using the medical care component of the Consumer Price Index[37].
Incremental cost-effectiveness ratio (ICER) was calculated using the following formula:
ICER=(Costno vaccination−Costvaccination)/(Effectno vaccination−Effectvaccination)
The numerator was the difference in total costs with or without vaccination. The denominator was DALYs that vaccination averted. There is no official guidance on discounting in China. All costs and DALYs were discounted to 2012 amounts at a rate of 3% annually (range 0–10%) according to Weinstein et al[38]. The cost-effectiveness thresholds were based on the WHO standard (highly cost-effective, ICER < GDP per capita; cost-effective, GDP per capita < ICER < 3×GDP per capita; and not cost-effective, ICER > 3×GDP per capita)[39]. The GDP per capita for China in 2012 was approximately $6,300[37].
To assess the robustness of the model and to identify influential model inputs for which additional data are warranted, one-way sensitivity analyses were performed at each level of vaccine price. There are substantial heterogeneities of disease incidence and costs among different regions in China. Thus, a two-way sensitivity analysis was performed to evaluate their influence on the base-case results for the case of $5 per dose. The ranges of model inputs for sensitivity analysis were all listed in Table 1. All data used in this study are available through references.
Table 2 shows the clinical events in a birth cohort of 1,000,000 Chinese infants followed for five years with or without EV71 vaccination. EV71 vaccination would be expected to annually avert 37,872 cases of EV71-associated HFMD, 2,629 cases of EV71-associated herpangina, 72,900 outpatient visits, 6,363 admissions to hospital, 29 deaths, and 945 DALYs among children younger than five years.
The economic burden of EV71-associated HFMD and herpangina incorporating direct and indirect costs is approximately 13 million dollars per 1,000,000 Chinese infants followed for five years. Table 3 shows the costs per DALY averted by the EV71 vaccination program at various prices per dose. According to WHO cost-effectiveness criteria, the vaccination program would be cost-saving at $2.5 and $5.0 per dose, cost-effective at $10, and not cost-effective at $20, $30 and $40. The break-even price of the vaccine is $5.2 per dose. When the price is less than $8.3 or $14.6 per dose, the vaccination program would be highly cost-effective or cost-effective, respectively.
Fig 2 shows the impact of HFMD incidence on the cost-effectiveness of EV71 vaccination at various prices per dose. As the incidence of HFMD falls below 0.3, 0.5, 0.9, 1.6, 2.4, and 3.2 per 1,000 person-years, the vaccination program would be not cost-effective at the prices per dose of $2.5, $5, $10, $20, $30, and $40, respectively.
Table 4 shows how other parameters influence the ICER comparing EV71 vaccination with no vaccination. At prices per dose less than $5, EV71 vaccination is still cost-saving or highly cost-effective when the parameters are varied across their ranges. At $10, the discount rate is the only influential parameter. When the cost and DALYs are both discounted at a rate more than 6%, EV71 vaccination would be no longer cost-effective. At $20, EV71 vaccination would be cost-effective only when the cost and DALYs are both discounted at a rate less than 2%. At the prices more than $30, EV71 vaccination would not be cost-effective when any parameter in Table 4 is varied across its range. A series of tornado diagrams show the rank of parameters’ influence on ICER at the prices per dose of $10, $20, $30, and $40, respectively (S1 Fig).
Fig 3 shows how HFMD incidence and disease costs influence the ICER when the vaccine price is $5 per dose. If the disease costs increase by 50%, the vaccination program would be not cost-effective in regions where the incidence of HFMD is below 0.4 per 1,000 person-years. If the disease costs decrease by 50%, the incidence making vaccination not cost-effective is below 0.6 per 1,000 person-years.
The results of this study suggest that a national vaccination program against EV71 would result in substantial decline in morbidity, mortality, use of health service, and DALYs in China. Based on an actual Chinese birth cohort size of around 15 million a year[40], EV71 vaccination would be expected to annually avert 567,500 cases of EV71-associated HFMD, 40,000 cases of EV71-associated herpangina, 1,093,500 outpatient visits, 95,500 admissions to hospital, 435 deaths, and 14,000 DALYs among children younger than five years. As EV71-associated early death contributes to the great majority of DALYs (Table 2) and current vaccines are highly effective, health benefit of the program mainly comes from its role in avoiding EV71-associated death.
The prices of vaccines in China’s routine childhood immunization program paid by the government are no more than $4.59 per dose[22]. Therefore, if the vaccine is priced in accordance with previous prices of under $4.59, then this vaccine would be cost-saving. This result is only sensitive to the HFMD incidence among all clinical and cost parameters. If the HFMD incidence is below than 0.5 per 1,000 person-years, that is, the incidence of EV71-associated HFMD is below 0.23 per 1,000 person-years, a national EV71 vaccination program would be unaffordable at the price per dose of $5. If the program is determined by local government, Western China provinces (Tibet, Xinjiang, Sichuan, Qinghai, and Gansu) with the lowest incidences should carefully balance the program’s economic cost and health benefit[13]. Besides the HFMD incidence, the discount rate, costs of EV71-associated HFMD, and hospitalization rate due to EV71-associated HFMD are the most influential parameters on the ICER comparing vaccination with no vaccination. However, at the price of $5 per dose and with the baseline incidence, they do not affect the cost-effectiveness.
This study has several methodological strengths. First, three large RCTs of high quality supplied solid clinical evidence on the efficacy and safety of EV71 vaccines in Chinese infants and children[9,10,11]. Second, nationwide epidemiology data on EV71-associated diseases were available and clinical outcomes were based on large populations[6,18]. Besides, the incidences of EV71-associated HFMD were age-specific so that we could assess the disease burden more precisely. Third, the disease costs incorporating direct and indirect costs came from seven surveys across regions varied in economic development levels in China[28,29,30,31,32,33,34]. These cost data benefited our study.
Due to lack of data, our study has several limitations. First, we did not consider long-term sequelae in EV71-infected children with severe complications. Although most recovered, two studies showed that some children with EV71 brainstem encephalitis (especially stage III) had residual cognitive and motor deficits at follow-ups after their hospitalization[41,42]. Second, we did not consider EV71-associated diseases other than HFMD and herpangina, mainly including upper respiratory tract infection and diarrhea. Third, according to the experience of our center and other pediatric centers in China, some children with severe complications died after discharge and they are not reported to the disease surveillance system. Fourth, this study adopted a static cohort model instead of a dynamic infection transmission model because the vaccine’s impact on the overall force of infection is not clear. Thus, the current analysis underestimated the disease burden of EV71 infection as well as vaccine effect. If the above four factors are considered, vaccination program would be more cost-effective.
Before the emergence of clinical evidence on EV71 vaccines, a previous study forecasted the economic value of a future vaccine against EV71[12]. Although the study also prefers routine vaccination in China (cost-effective when vaccine cost is $25 and efficacy ≥70% or cost is $10 and efficacy ≥50%), several key clinical and cost parameters are different between the study and ours: (1) based on 2009 data, the incidence of EV71 infection was lower than ours which was based on data of four years; (2) a percentage of more than 26% of severe cases in all EV71 infected cases was quite high compared to our data from the China CDC; (3) the efficacy of vaccine ranged from 50%-90%, which was quite lower than ours (90%-98%); (4) disease costs came from an American population and was converted to China hospital costs using an indirect method. These costs were much lower than those from recent seven studies directly surveying economic burden of the disease[28,29,30,31,32,33,34].
The EV71 vaccines in the three phase 3 trials were all developed on the basis of subgenotype strain C4. Fortunately, other subgenotype strains of EV71 have not been reported in mainland China [8,43]. However, in other Asian regions, B4, B5, C2, and C5 have been reported[8,43,44]. Studies showed that C4 vaccine could elicit cross-neutralizing response with other subgenotype strains[45,46,47]. Nevertheless, the degree of cross-protective immunity and the potential escape evolution for EV71 are unknown. More importantly, this cross-protective immunity has not been tested in clinical trials. The incidence of EV71-associated diseases, disease costs, and medical resources in other Asian regions are all different from China. Further studies should evaluate the public health impact and economic value of these vaccines in other Asian regions.
This study focused on EV71 vaccination versus no EV71 vaccination. Future analyses should evaluating the cost-effectiveness of vaccination against Coxsackievirus A 16 (CA16) or other enteroviruses causing HFMD. The predominance of CA16 is comparable with EV71 nationwide and CA16 often becomes the main epidemic strain in some regions of China[6]. Strategies of EV71 or CA16 vaccination alone and combined vaccination may need evaluation. Chen et al have shown that the co-administration of inactivated EV71 vaccine with a commercial pentavalent vaccine did not affect the antibody response of each vaccine[48]. When considering the incorporation of EV71 vaccine into the EPI, future analyses should also evaluate strategies of co-administration with other EPI vaccines.
A national EV71 vaccination program would prevent a substantial portion of EV71 related morbidity, mortality, outpatient visits, and admissions to hospitals among children younger than five years in China. Within the price range of current routine vaccines paid by the government, the program is cost-saving or highly cost-effective. Policy makers should consider including EV71 vaccination as part of China’s routine childhood immunization schedule.
|
10.1371/journal.ppat.1004577 | Extreme Divergence of Wolbachia Tropism for the Stem-Cell-Niche in the Drosophila Testis | Microbial tropism, the infection of specific cells and tissues by a microorganism, is a fundamental aspect of host-microbe interactions. The intracellular bacteria Wolbachia have a peculiar tropism for the stem cell niches in the Drosophila ovary, the microenvironments that support the cells producing the eggs. The molecular underpinnings of Wolbachia stem cell niche tropism are unknown. We have previously shown that the patterns of tropism in the ovary show a high degree of conservation across the Wolbachia lineage, with closely related Wolbachia strains usually displaying the same pattern of stem cell niche tropism. It has also been shown that tropism to these structures in the ovary facilitates both vertical and horizontal transmission, providing a strong selective pressure towards evolutionary conservation of tropism. Here we show great disparity in the evolutionary conservation and underlying mechanisms of stem cell niche tropism between male and female gonads. In contrast to females, niche tropism in the male testis is not pervasive, present in only 45% of niches analyzed. The patterns of niche tropism in the testis are not evolutionarily maintained across the Wolbachia lineage, unlike what was shown in the females. Furthermore, hub tropism does not correlate with cytoplasmic incompatibility, a Wolbachia-driven phenotype imprinted during spermatogenesis. Towards identifying the molecular mechanism of hub tropism, we performed hybrid analyses of Wolbachia strains in non-native hosts. These results indicate that both Wolbachia and host derived factors play a role in the targeting of the stem cell niche in the testis. Surprisingly, even closely related Wolbachia strains in Drosophila melanogaster, derived from a single ancestor only 8,000 years ago, have significantly different tropisms to the hub, highlighting that stem cell niche tropism is rapidly diverging in males. These findings provide a powerful system to investigate the mechanisms and evolution of microbial tissue tropism.
| Microbes evolve to infect structures favoring their transmission in host populations. A large fraction of insects are infected with Wolbachia bacteria. Usually Wolbachia are transmitted the same way we inherit our mitochondria, via the eggs from the mother. In fruit flies, to favor maternal transmission, Wolbachia infect the microenvironment containing the egg producing stem cells, called the “stem cell niche”. Targeting of the stem cell niche is evolutionary conserved in female fruit flies, observed in all Wolbachia strains analyzed to date. Remarkably, in males, we find many Wolbachia strains not infecting the stem cell niche present in the testis. We report a surprising diversity in stem cell niche infection in males, contrasting with extreme conservation in females. We further show that even closely related Wolbachia strains in D. melanogaster display rapidly evolving patterns of stem cell niche targeting in males. Understanding the molecular mechanisms driving these differences will identify sex specific features of stem cell niche biology. Because Wolbachia promote insect resistance against human diseases transmitted by mosquitos, Wolbachia are becoming a valuable tool in the control of several diseases, including Dengue and malaria. Knowledge emerging from this research will also provide novel tools towards Wolbachia based strategies of disease control.
| The evolutionary interests of males and females are frequently divergent. Sexual conflict arises when phenotypes that enhance the reproductive success of one sex reduces the fitness of the other sex [1]. A well-characterized example in Drosophila is sperm competition between males. Sperm competition results in rapid evolution of sperm proteins which up-regulate females' egg-laying rate and reduces her desire to re-mate with another male. However, these proteins also shorten the female's lifespan reducing her fitness [reviewed by 2].
Vertically transmitted reproductive parasites, such as Wolbachia, can also cause sexually divergent phenotypes in males and females. Wolbachia are obligate intracellular bacteria present in a large fraction of insects, as well as spiders, mites, crustaceans, and filarial worms. They are primarily vertically transmitted from mother to offspring in a manner analogous to mitochondrial inheritance, although there is extensive evidence of horizontal transmission in nature [3], [4]. For intracellular bacteria, vertical transmission often favors infected female hosts, which is also the case for Wolbachia [5]. There are several Wolbachia-induced phenotypes favoring the infected female, including parthenogenesis, feminization, male killing, and cytoplasmic incompatibility [6]. Each of these phenotypes ultimately results in the spread of more infected female hosts. In such cases, maternally transmitted bacteria can act as selfish genetic elements driving sexual conflict [5].
For successful vertical transmission, Wolbachia need to be present in the eggs laid by infected females. It has been shown in Drosophila that Wolbachia display a strong tropism for the germline, in particular, the oocyte, to ensure a high percentage of vertical transmission [7]–[10]. Although vertical transmission is prevalent, Wolbachia also can spread horizontally across individuals and species [3], [11], [12]. Colonization of the germline is a prerequisite for the infection to become successfully established into a population. We have previously shown that upon recent infection, Wolbachia colonize the stem cell niches in the Drosophila ovary, favoring vertical transmission after horizontal transfer [13]. Furthermore, stem cell niche tropism in the ovary is a highly evolutionarily conserved phenotype across the Drosophila genus, present in 100% of ovaries analyzed [14]. Wolbachia also infect the putative stem cell niches in the ovaries of other species, such as the bedbug and leafhopper [15], [16] indicating that the selective pressure for Wolbachia targeting of ovarian stem cell niches to favor transmission extends beyond the Drosophila genus.
Wolbachia have also been shown to display tropism to the stem cell niche present in the testis in Drosophila mauritiana [17]. However, the conservation of this phenotype across the Drosophila genus is unknown. Here we show that the evolutionary conservation of stem cell niche tropism present in females is not maintained in the male lineage. In fact, Wolbachia niche tropism in the testis, compared to the female results, represents a pronounced sexual dimorphism in the evolutionary history of Wolbachia stem cell niche tropism. Furthermore, we′ve identified that both Wolbachia and host factors modulate hub tropism in this system. Finally, we show that very closely related Wolbachia strains infecting the same host differ significantly in the densities at which they colonize the hub, indicating that hub tropism is a rapidly diverging phenotype in males.
In the testis, the germline stem cells (GSCs) and cyst stem cells (CySCs) reside at the “hub”, a structure at the apical tip of the testis (Fig. 1A). The hub is a group of 10 to 16 somatically derived cells forming the microenvironment supporting the stem cells, referred to as the niche [18]. It has been shown that the GSCs receive maintenance signals from both the hub and the CySCs, hence both are considered to be part of the stem cell niche for the GSCs. However for the context of this study, niche tropism in the testis refers to Wolbachia infection of the hub only. To investigate whether Wolbachia niche tropism is as pervasive in the hub, as previously shown in the ovary [14], we surveyed various Drosophila species infected with different strains of Wolbachia (Fig. 1; S1 Dataset; see S1 Table for the sources for the stocks used in this analysis).
Using confocal imaging and immunohistochemistry, we analyzed the density of Wolbachia infection in the hub cells as compared to the density of Wolbachia in the surrounding tissue (see Materials and Methods). We found that Wolbachia target the hub at varying frequencies and densities across the Drosophila genus (Fig. 1, S2 Table, S1 Dataset). 3 out of 9 species showed very little to no Wolbachia infection in the hub (Fig. 1 H–J, quantification in K), indicating that hub tropism is not pervasive across the Drosophila genus. 6 out of 9 species analyzed, however, did have Wolbachia tropism to the hub, ranging from 17% of niches infected to 95% of niches infected (Fig. 1 B–G, K, see also Materials and Methods). The 6 Drosophila species- Wolbachia strain pairs with hub tropism fall into two groups with significantly different frequencies and densities of tropism: 3 had very high frequencies and densities of hub infection: D. ananassae wAna, D. melanogaster wMel, and D. mauritiana wMau; and 3 had moderate frequencies of Wolbachia tropism to the hub: D. yakuba, wYak, D. tropicalis wWil, and D. simulans wRi. In the ovary, tropism to the somatic stem cell niche is found at high frequencies in every individual of all Drosophila species analyzed [14]. In contrast, tropism for the hub is found in only a fraction of the species analyzed.
Similar to the results for hub tropism, the frequency of tropism to the germline stem cell niche (GSCN) in the ovary was shown to be variable across the Drosophila genus (Fig. 2A and [14]). We reasoned that Wolbachia tropism to the hub in the testis could simply be a byproduct of GSCN targeting in the ovaries. Interestingly, however, the presence of hub tropism does not correlate with the presence GSCN tropism (S3 Table, Correlation Test, p = 0.773). Although tropism in males and females is correlated in some strains (5 out of 9, e.g. wMau displays high frequencies of both hub tropism and GSCN tropism and wSh does not have tropism to either the hub or the GSCN), there are also others that do not (4 out of 9). The Wolbachia strain displaying one of the highest frequencies of GSCN tropism in the ovary (wNo, 99% [14]), displays no tropism to the hub (0%, Fig. 1 I and K). Conversely, a Wolbachia strain displaying a high frequency of tropism to the hub (wMel, 71%, Fig. 1 C and K) does not target the GSCN in the ovary (1%, [14]). These data reveal that Wolbachia stem cell niche tropism does not correlate with GSCN tropism in the female.
Previously, we have shown that the pattern of GSCN tropism is evolutionarily conserved across the Wolbachia lineage ([14] and Fig. 2). To assess whether hub tropism was also conserved across the Wolbachia lineage, we aligned the frequencies of hub tropism on the Wolbachia phylogenetic tree (Fig. 2). We quantified the correlation of hub tropism pattern with the Wolbachia phylogeny using a computer simulated model of randomized character distributions to compare with the distribution of niche tropism pattern on each of the phylogenies, as previously described [14]. We found that it is highly probable that the distribution of hub tropism is completely independent of the Wolbachia phylogeny (S2 Fig.). Similarly, when we compared hub tropism to the Drosophila phylogeny, we found no clear correlation between the two (S3 Fig.). Quantification of the relationship revealed that frequency of hub tropism bears no correlation with the Drosophila phylogeny (S4 Fig.).
An important Wolbachia related phenotype that also bears no correlation with host or microbial phylogenies is cytoplasmic incompatibility (CI). CI is a reproductive phenotype resulting in reduced embryo hatching when a Wolbachia infected male mates with an uninfected female. We examined the possibility of a correlation between tropism to the hub and CI by comparing our tropism data to previously published reports on the levels of CI across the Drosophila genus (S4 Table) [19]–[23]. This analysis shows that some species with high levels of CI have different levels of tropism (i.e. wSh and wRi have 0% and 17% hub tropism, respectively). Conversely, some species with low levels of CI also have a wide range of hub tropism phenotypes (i.e. wTei and wMau have 2.3% and 71% hub tropism frequencies, respectively). Although hub tropism is highly divergent even amongst closely related strains of Wolbachia, similar to CI, there does not seem to be a correlation between these two phenotypes (S4 Table, Correlation test, p = 0.267).
We next aimed to elucidate if host or bacterial factors influence the highly dynamic nature of the hub tropism phenotype. To investigate this question, Wolbachia strains backcrossed into a different host were used to assess Wolbachia strain versus host background influence on hub tropism, as previously described [14]. D. mauritiana wMau, which displays hub tropism (Fig. 1D and Fig. 3) and D. sechellia wSh, which does not display hub tropism (Fig. 1J and Fig. 3) and their hybrid offspring were utilized in this study (See Material and Methods).
Wolbachia strain wSh, infecting its native host, D. sechellia, and its non-native host, D. mauritiana, displays no hub localization, regardless of host genetic background (Fig. 3, S5 Table). This result suggests that Wolbachia wSh is incapable of hub tropism in either species. However it does not rule out the possibility that the hosts share a mechanism for excluding wSh from the hub. Therefore, a lack of tropism in both hosts cannot provide insight into whether the host or microbe is providing factors contributing to hub tropism.
The analysis of wMau hub tropism allows further probing into this question. Wolbachia strain wMau infecting its native host, D. mauritiana, and its non-native host, D. sechellia, displays tropism for the hub, suggesting that the Wolbachia strain is driving this phenotype. However, the frequency of targeting in the hybrid host is 3-fold lower than in the native host (Fig. 3C, green bars). Statistical analysis of frequency data indicates that both host genetic background and Wolbachia strain can significantly affect the frequency of hub tropism (Fisher's exact test, p = 8.309×10−5 and p = 2.267×10−10, respectively). These results are in contrast to previous data in the ovaries where only the Wolbachia strain drives tropism. wMau can efficiently target the GSCN in the ovary of both its native and hybrid host, greater than 80% of niches infected, regardless of the host genetic background [14]. The wMau frequency data in the male support the hypothesis that the Wolbachia strain is directing hub tropism. However, because the frequency of targeting is not as robust in the hybrid host compared to its native host, a role for the host is also implicated.
In relation to Wolbachia density in the hub, the data indicate that the Wolbachia encoded factors play a major role in both native and hybrid hosts. The overall density at which wMau infect the hub is conserved (Fig. 3 B and C, native host solid green bar, hybrid host hatched green bar, S4 Table). Similarly, wSh hub titers, compared to the surrounding tissue, are less than 1 in both native and hybrid hosts (Fig. 3 B and C, native host solid red bar and hybrid host hatched red bar, S4 Table). Linear regression analysis of density data indicates that the Wolbachia strain, rather than the host genetic background, modulates Wolbachia density in the hub (P = 0.045 and P = 0.56, respectively). With respect to both frequency and density, the overall data reveal that factors encoded by both the host species and the Wolbachia strain influence hub tropism in the Drosophila testis.
To further investigate the role of Wolbachia on hub tropism, we then analyzed different Wolbachia strains in the same host species. We took advantage of D. simulans, which is a host to many different Wolbachia strains. We investigated two strains of D. simulans flies differentially infected with wRi and wNo and their backcrossed offspring. Flies were backcrossed to account for any genomic divergence between host strains, as previously described [14]. D. simulans flies infected with Wolbachia wRi display hub tropism in about 33% and 43% of hubs analyzed for the parental and backcrossed hosts, respectively (Fig. 4, S6 Table). D. simulans wNo displays hub tropism infrequently (2% and 15% of hubs highly infected for the parental and backcrossed hosts, respectively, Fig. 4, S6 Table). Although the frequencies of hub tropism for each Wolbachia strain increase in the backcrossed hosts, the general trend remains, where wRi targets the hub at a higher frequency than wNo. To quantify the relative contributions of host and bacterial factors towards hub tropism, logistical regression was performed. Wolbachia factors have a significant effect on hub tropism as compared to no significance of the host genetic background in the D. simulans hybrid flies (p = 0.0000552 and p = 0.927 respectively). These results indicate that when host factors are kept constant, Wolbachia strain factors are sufficient to significantly modulate the frequency of hub tropism.
In the previous analyses of hybrid crosses, hub tropism of distantly related Wolbachia strains were compared, first with different host species (Fig. 3), then within the same host species (Fig. 4). These results indicate that although the fly host can play a role in hub tropism, Wolbachia can significantly affect tropism on its own. In both cases, we were comparing Wolbachia strains from the A and B supergroups. We next investigated if the observed diversity of niche tropism is still present between more closely related Wolbachia strains. To address this question, we analyzed hub tropism of several Wolbachia strain variants infecting Drosophila melanogaster that diverged from a single ancestor within the last 8,000 years [24], [25].
Hub tropism of wMel-like (wMel, wMel2, and wMel3) and wMelCS-like (wMelCS, wMelCS2, and wMelPop) Wolbachia strains were analyzed. These Wolbachia strains were introgressed into the same D. melanogaster genetic background with the same microbiota [25]. The data reveal that the three wMel-like Wolbachia strains have significantly different tropism phenotypes from the wMelCS-like strains (Fig. 5, S7 Table). The wMel-like strains target the hub at similar frequencies, between 25% and 50%, and at similar densities, about 1.5-fold higher than the surrounding tissue. The wMelCS-like strains target the hub at significantly higher frequencies (P<0.05) and densities (P<0.001) than the wMel-like strains. Within the wMelCS-like group, wMelPop targets the hub at a significantly higher frequency (100%) than wMelCS2 (77%; P = 0.005), but not wMelCS (90%). However, wMelPop targets at a significantly higher density than both wMelCS and wMelCS2 (P<0.0001; S1 Movie). Interestingly, wMelPop densities increase to the point where the hub cells burst open in approximately 20% of hubs (S5 Fig. and S2 Movie). The finding that the wMel-like and wMelCS-like Wolbachia variants, all derived from a single ancestor only 8,000 years ago, have significantly different frequencies and densities of targeting indicates that hub tropism is a rapidly diverging phenotype.
A fundamental aspect of Wolbachia-host interactions is the type of tissue preferentially infected by the bacteria. We have previously shown that Wolbachia tropism to the stem cell niches in the female Drosophila ovaries is important for vertical transmission, and that this tropism is ubiquitous across the Drosophila genus. Furthermore, closely related Wolbachia strains tend to display the same patterns of tropism in the ovary, indicating the importance of maintaining this phenotype for vertical transmission [14].
If the major role of niche tropism is related to Wolbachia transmission, evolutionary theory predicts that there should be reduced selective pressure to maintain niche tropism in males, since Wolbachia is not transmitted through the sperm. Patterns of Wolbachia niche tropism in the filarial nematode (B. malayi, D. immitis, L. sigmondontis, M. unguiculatus, and O. dewittei japonica) support this concept, where Wolbachia colonization of the distal tip cell (the nematode equivalent of the stem cell niche) and subsequent germline invasion occurs only in females [26]. In agreement, the results shown here indicate a reduced level of conservation of hub tropism phenotype, contrasting with previous observation in females [14]. The stem cell niches in the ovary and testis are well characterized and have several signaling pathways in common [27]. The robust sexual dimorphism in the evolutionary conservation of niche tropism, indicates that Wolbachia could be recognizing novel sex specific differences in these cells [28].
Wolbachia-induced host phenotypes related to stem cell biology and testis physiology have been previously described [17], [23]. We investigated whether hub tropism correlates with those known Wolbachia-related reproductive phenotypes. Because GSCN tropism in the ovary was shown to not be ubiquitous across the Drosophila genus, we reasoned that hub tropism could simply be a byproduct of GSCN tropism in the female. However, the frequencies of GSCN and hub tropism only correlate in 5 out of the 10 species and are not statistically significant.
On the cellular level, another phenotype we have previously shown was a Wolbachia-dependent increase in the rate of germline stem cell division (GSCD) in the ovaries of D. mauritiana. Although a similar trend exists in the D. mauritiana testis, the up-regulation of GSCD was not shown to be significant, showing a lack of conservation of a phenotype derived in the females to boost their spread [17].
A third important Wolbachia mediated phenotype, cytoplasmic incompatibility (CI), is a consequence of Wolbachia modification of sperm during spermatogenesis, causing embryonic lethality of uninfected eggs fertilized by sperm from infected males [reviewed by 29]. Although the precise mechanism is not well understood, the sperm from infected males is modified (mod+) and an infected egg with the appropriate rescue factor (resc+) is required for embryo viability [30], [31]. Several lines of evidence suggest that the modification of the sperm occurs at the chromatin level [32]–[34]. Extensive analyses of Wolbachia population dynamics and localization during spermatogenesis have demonstrated that CI is a non-cell autonomous effect caused by a diffusible Wolbachia factor during spermatogenesis [35]. Interestingly, local factors secreted by the hub can act on the germline stem cell. Since niche factors are extrinsic to the stem cell, they can affect the testis germline stem cell and consequently their sperm-forming progeny in a non-cell autonomous fashion. Niche factors have also been shown to cooperate with chromatin remodeling complexes towards control of germline stem cell maintenance and differentiation [36]. Therefore, we attempted to correlate our tropism data with published data regarding CI levels of several Wolbachia strains across the Drosophila genus. However, we found no correlation between Wolbachia hub tropism and CI, suggesting that Wolbachia's presence in the hub is not required for the CI effect. This suggests that either Wolbachia factors modify the sperm later in spermatogenesis or if Wolbachia-derived factors are affecting early spermatogenesis events towards CI, it is independent of Wolbachia infection of the niche.
Literature shows that both the host species and Wolbachia strains have rapidly evolving aspects that could contribute to the dynamic evolutionary changes in Wolbachia hub targeting shown here. Regarding the host, several testis specific genes, male seminal fluid proteins, and spermatogenesis genes have been shown to be rapidly evolving [37]. Furthermore, proteins related to GSC biology are also undergoing recurrent positive selection [38]. From the perspective of the bacteria, Wolbachia genomic analyses suggest that these bacteria have one of the most highly recombining intracellular bacterial genomes, with many genomic differences between closely related strains [39]–[42].
We investigated the relative contribution of both host and bacterial factors towards hub tropism phenotype. Unlike in the ovary where host derived factors did not play a role [14], in the testis, host factors could not be ruled out. When comparing distantly related Wolbachia strains and host species (D. mauritiana and D. sechellia hybrid lines), the data indicate that both host and Wolbachia derived factors contribute to the differences in hub tropism. One possibility is that there is selective pressure on the host driving rapid evolution of the hub intracellular environment to counteract negative effects of Wolbachia colonization of the testis niche. Although there is no evidence in the literature for positive selection of hub proteins, genes in the neighboring germline stem cell have been shown to be undergoing positive selection [38], [43]. Independent of differential host factors, we were able to confirm Wolbachia's role in hub tropism. By comparing distantly related Wolbachia strains in the same host species (D. simulans lines), we were able to confirm that Wolbachia derived factors significantly modulate hub tropism.
To assess how quickly this modulation of hub tropism can evolve, we investigated if very closely related Wolbachia strains that have recently diverged could display diverse hub tropism phenotypes. Several variants of the wMel strain of Wolbachia naturally infecting D. melanogaster exist [44], [45]. Due to strict maternal transmission, congruent Wolbachia and mitochondrial lineages made it possible to trace these lineages back to a single common D. melanogaster ancestor existing around 8,000 years ago [24], [25]. We investigated hub tropism of wMel-like (wMel, wMel2, and wMel3) and wMelCS-like (wMelCS, wMelCS2 and wMelPop) Wolbachia strains which have been shown to induce differential protection against viruses [25]. The wMel-like and wMelCS-like subgroups can be separated into three statistically distinct groups based on their density of hub infection (1: wMel, wMel2, and wMel3; 2: wMelCS and wMelCS2; 3: wMelPop), indicating that they have evolved distinct cellular tropisms. These data demonstrate that hub tropism is a rapidly diverging phenotype.
The fast paced changes in the hub tropism phenotype during the evolution of these different Wolbachia strains raises the questions of what mechanisms are driving these rapid changes and is adaptive evolution occurring. If Wolbachia tropism for the hub is causing an unfavorable phenotype in the host, a molecular arms race will result where both the host and microbe will rapidly evolve [46], [47]. We did not find any correlation of hub tropism with CI, germline stem cell division, or with other obvious testis related phenotypes. It is possible that hub tropism may have a phenotypic effect on the host, but at the moment these are unknown and we have no evidence supporting adaptive evolution in response to a host-microbe arms race driving rapid changes in hub tropism in wMel strains.
Another possibility is that genetic drift is driving the extreme divergence in hub tropism that we report here. At every generation, from embryonic development through the mature egg, Wolbachia undergoes several bottlenecks: only the Wolbachia present in the germplasm of the embryo will colonize the primordial germ cells [8], [10]. Within the germline, only the Wolbachia present in the oocyte is transmitted to the progeny [7], [9], [10]. This effectively reduces the genetic effective population sizes and increases the rate of fixation of mutations by drift. There are several studies highlighting the role of genetic drift driving high rates of genome sequence evolution in vertically transmitted endosymbionts [reviewed by 48]. The data presented here suggest that mutations that are neutral regarding niche targeting in the female may affect niche tropism in the male. If these mutations do not affect Wolbachia overall fitness in the females and do not interfere with transmission, they can be fixed by drift and result in significant niche tropism evolution in males.
At the moment it is difficult to identify the specific molecular underpinnings resulting in the differences in niche tropism phenotypes between these strains. A possible molecular player involved in hub tropism could be encoded by the gene region known as ‘octomom’. This region was found to be amplified several times in wMelPop, and contains genes predicted to be involved in DNA replication. It has been proposed to be responsible for the wMelPop over-replication phenotype [25], although there are conflicting reports [49]. This could explain the highest titers present in wMelPop-infected hubs. However, there are other unknown factors contributing to the range of hub tropism phenotypes observed in the other wMelCS-like and wMel-like strains, since they have only once copy of the octomom region. The wMel variants are defined by several polymorphic genetic markers [25], [44], [45], [49]. There are 108 single nucleotide polymorphisms (SNPs), a tandem duplication, and seven insertion-deletion polymorphisms between the wMel and wMelCS-like (wMelPop) strains [25]. Further characterization of niche tropism of different strains in the same host genetic background, together with additional sequencing of diverse strains, will allow the correlation of Wolbachia genomic features with patterns of niche tropism. Future identification of Wolbachia proteins modulating the different levels of hub tropism will provide insights into the evolutionary mechanism driving this rapid divergence in males and the robust sexual dimorphism of stem cell niche targeting.
Here we presented tropism differences in Wolbachia strains well characterized at the genomic level in a Drosophila species with a large repertoire of transgenic and genetic tools. These findings provide the foundation to dissect the molecular mechanisms involved in Wolbachia hub tropism. Furthermore, the differences in stem cell niche tropism between males and females may reveal sex specific differences in the biology of stem cell niche being recognized by Wolbachia. Identification of the Wolbachia factors involved in tissue tropism is fundamental in understanding how bacteria spread and infect their hosts in nature and will provide additional tools towards vector and disease control.
Fly stocks used in this analysis and their sources are listed in S1 Table. Drosophila species naturally infected with Wolbachia comprising the melanogaster subgroup were selected, along with two additional species outside the melanogaster subgroup: D. tropicalis and D. ananassae, belonging to the willistoni and ananassae subgroups, respectively. Introgression crosses for hybrid analysis experiments were performed as previously described [14]. D. melanogaster flies infected with the several wMel Wolbachia variants were introduced into the same genetic background as described elsewhere [25].
Flies were raised at room temperature and fed a typical molasses, yeast, cornmeal, agar food, with the exception of D. sechellia flies which were supplemented with reconstituted Noni Fruit (Hawaiian Health Ohana, LLC) [50].
For consistency and proper comparison to previous analysis of niche tropism in the female, males in this study were aged to seven days at room temperature (with the exception of the D. simulans hybrids for Fig. 4, which were dissected upon eclosion, see Toomey et al, 2013 for details). At least 20 flies were dissected for each sample, and total N's of hubs analyzed are listed in the Supplemental tables for each experiment. Testis were fixed using a 4% paraformaldehyde solution and subjected to immunostaining as previously described [13]. The mouse anti-hsp60 (Sigma, 1∶100) antibody was used to visualize Wolbachia. Hub markers were either rat anti-α-catenin (DSHB, DCAT1, 1∶40) or rat anti-DE-Cadherin (DSHB, DCAD2, 1∶20). Nuclei were counterstained with Hoechst (1 µg/ml, Molecular Probes).
Images of the hub were acquired using a FV1000 confocal microscope. Wolbachia signal intensity in the hub and surrounding area were measured in Z-stacks of images using MatLab software for image quantification. Manual masks were drawn around the hub structure as well as the surrounding soma and germline using only the hub marker and DNA. Wolbachia density was measured within each mask and Wolbachia infection of the hub was considered tropism if the density relative to the surrounding soma and germline was at least 1.5-fold increased. A 1.5-fold threshold for tropism was previously determined to best represent what visually appears to be a higher density of Wolbachia in the niche versus the surrounding tissue [14]. Raw data showing density ratios is provided in S1 Dataset.
We utilized a computer simulation model of randomized character distributions to compare with the distribution of niche tropism pattern on each of the phylogenies to quantify the correlation of niche tropism pattern to the Wolbachia and Drosophila phylogenies (S1 and S3 Figs.) [51]. We used tree length as a measurement for goodness of fit for the distribution of a character, such as the tropism pattern, as aligned with the phylogeny. Tree length is defined as the total number of steps required to map a data set onto a phylogenetic tree.
To determine the three significant groups for tropism in Fig. 1, a two-sample test for proportions was used on frequency data (Fig. 1K) and T-tests were used for density data (Fig. 1L). A Bonferroni correction was applied to account for multiple comparisons.
To determine the significance of host genetic background versus Wolbachia strain (Fig. 4) on the frequency of hub tropism a logistical regression was performed on frequency data as previously described (Fig. 4B) [14]. When “zero” frequencies are present, logistic regression analysis was replaced by a Fisher Exact Test (Fig. 3B). For density data, a linear regression was performed (Fig. 3C).
To determine if the frequencies of targeting between Wolbachia strains were significantly different (Fig. 5B), a two-sample test for proportions was used. If there were more than two strains being compared a Chi-square test was performed. To determine if the differences in densities were significant, pair-wise t-tests were performed (Fig. 5C).
|
10.1371/journal.pgen.1002128 | A Functional Variant in MicroRNA-146a Promoter Modulates Its Expression and Confers Disease Risk for Systemic Lupus Erythematosus | Systemic lupus erythematosus (SLE) is a complex autoimmune disease with a strong genetic predisposition, characterized by an upregulated type I interferon pathway. MicroRNAs are important regulators of immune homeostasis, and aberrant microRNA expression has been demonstrated in patients with autoimmune diseases. We recently identified miR-146a as a negative regulator of the interferon pathway and linked the abnormal activation of this pathway to the underexpression of miR-146a in SLE patients. To explore why the expression of miR-146a is reduced in SLE patients, we conducted short parallel sequencing of potentially regulatory regions of miR-146a and identified a novel genetic variant (rs57095329) in the promoter region exhibiting evidence for association with SLE that was replicated independently in 7,182 Asians (Pmeta = 2.74×10−8, odds ratio = 1.29 [1.18–1.40]). The risk-associated G allele was linked to reduced expression of miR-146a in the peripheral blood leukocytes of the controls. Combined functional assays showed that the risk-associated G allele reduced the protein-binding affinity and activity of the promoter compared with those of the promoter containing the protective A allele. Transcription factor Ets-1, encoded by the lupus-susceptibility gene ETS1, identified in recent genome-wide association studies, binds near this variant. The manipulation of Ets-1 levels strongly affected miR-146a promoter activity in vitro; and the knockdown of Ets-1, mimicking its reduced expression in SLE, directly impaired the induction of miR-146a. We also observed additive effects of the risk alleles of miR-146a and ETS1. Our data identified and confirmed an association between a functional promoter variant of miR-146a and SLE. This risk allele had decreased binding to transcription factor Ets-1, contributing to reduced levels of miR-146a in SLE patients.
| Genome-wide association studies have identified quite a number of susceptibility loci associated with complex diseases such as systemic lupus erythematosus (SLE). However, for most of them, the intrinsic link between genetic variation and disease mechanism is not fully understood. SLE is characterized by a significantly upregulated type I interferon (IFN) pathway, and we have previously reported that underexpression of a microRNA, miR-146a, contributes to alterations in the type I IFN pathway in lupus patients. Here we identified a novel genetic variant in the promoter region of miR-146a that is directly related to reduced expression of miR-146a and is associated with SLE susceptibility. The risk allele of this variant confers weaker binding affinity for Ets-1, which is a transcription factor encoded by a lupus susceptibility gene found in recent GWAS. These findings suggest that reduced expression of Ets-1 and its reduced binding affinity to the miR-146a promoter both may contribute to low levels of this microRNA in SLE patients, which may contribute to the upregulated type I IFN pathway in these patients. To our knowledge, this is also the first piece of evidence showing association between a genetic variant in a promoter region of a miRNA gene and a human disease.
| Systemic lupus erythematosus (SLE) is a chronic autoimmune disease with a complex etiology and diverse clinical manifestations [1]. The role of genetic factors in the SLE risk has long been established, and demonstrated in familial aggregations, twin studies, and sibling recurrence rates [2]. Recently, high-throughput technologies have facilitated genome-wide association studies (GWASs) across different populations. This approach, accompanied by large-scale replications, has not only confirmed the association of many established susceptibility genes, but has also presented convincing evidence of novel genetic loci involved in SLE [3]–[8]. As members of the Asian Lupus Genetics Consortium, we have also performed a GWAS in Asian populations and have identified variants in ETS1 and WDFY4 that are associated with SLE [9]. A combination of GWAS data from different ethnic groups will clearly provide new insights into the genetics of SLE and further our understanding of the pathogenesis of lupus [10], [11].
To use genomic tools to study the mechanisms of SLE, we and others have independently identified a gene expression signature for lupus patients using microarray profiling [12]–[14], which highlights the pathogenic role of the abnormal activation of the type I interferon (IFN) pathway in human lupus [15]–[17]. Intriguingly, recent investigations suggest a genetic contribution to the variability observed among individuals in the production and signaling of IFN [17], and advances in the genetics of SLE highlight the strong association between the risk of developing lupus and gene variants connected to the production and effects of type I IFN [11], [18].
We recently used a microRNA (miRNA) profiling assay to examine the involvement of miRNAs in SLE, because miRNAs are novel gene expression regulators [19] and important players in shaping the immune response [20]–[22]. This profiling identified a reduction in miR-146a expression in lupus patients, and we showed that the underexpression of miR-146a contributes to lupus pathogenesis by deregulating the activation of the IFN pathway [23]. However, why miR-146a levels are reduced in patients with SLE remains unresolved. miR-146a is encoded at 5q33.3. Interestingly, recent data from GWASs in both European and Asian populations have indicated that this region is a novel susceptibility locus for SLE [3], [7], [9], suggesting a plausible role for a genetic variant around miR-146a in modulating its expression and thus the disease risk.
Several studies have demonstrated unambiguously that genetic variants in miRNA precursors (pre-miRNA) can affect miRNA expression levels by interfering with the miRNA maturation process and are thus associated with disease susceptibility [24]–[26]. We postulate that genetic variants in both the miRNA promoter and the precursor region may alter mature miRNA production. Given the critical regulatory role of miR-146a in the type I IFN pathway and the abovementioned genetic association between this pathway and SLE susceptibility, polymorphisms in the miR-146a gene could also potentially confer a disease risk. To assess whether genetic variants modulate miR-146a expression and thus contribute to the risk of developing SLE, we sequenced the promoter and key regulatory regions of the miR-146a precursor to identify potential functional variants that might be associated with SLE susceptibility. Our subsequent replication and functional studies provide evidence that single-nucleotide polymorphism (SNP) rs57095329 in the miR-146a promoter, which affects its mature level, can confer SLE susceptibility.
miR-146a is located at 5q33.3. The transcription start site (TSS) of its primary transcript (pri-miR-146a) has been identified [27]. To characterize the essential regulatory region for subsequent genetic analysis, we first cloned miR-146a upstream fragments with variable 5′ ends into the pGL3-basic reporter plasmid to analyze its promoter activity. We found that the inclusion of a fragment from nucleotide (nt) −1,091 to nt −611, which contains a known NF-κB-binding site characterized in THP-1 cells [27], was consistently robust to promote luciferase activity in HeLa cells (Figure S1). The inclusion of the more distal region (nt −1,998 to nt −1,091) enhanced neither the basal nor phorbol myristate acetate and ionomycin (hereafter referred to as “PMA+Iono”) -induced activity of the promoter (Figure S1B). Therefore, to look for new genetic variants and to characterize their potential association with SLE, we designed four pairs of primers with which to sequence the upstream region that spans the 1,105-bp promoter (nt −1,091 to nt +14) and the consecutive first exon of pri-miR-146a (Figure S1A), in 360 individuals (180 SLE patients and 180 controls), which served as the discovery panel. We also sequenced the 452-bp region centered on miR-146a precursor or exon 2 of pri-miR-146a, because it potentially affects mature miR-146a production [26], in the same discovery panel. A total of 12 variants were identified, with nine already reported in the dbSNP database Build 130 (Table S1). Five variants had a minor allele frequency (MAF) of >1% (rs17057381, rs73318382, rs57095329, rs6864584, and rs2910164; Figure S1A). Therefore, we extended our sequencing analysis to examine these five SNPs in up to 816 patients and 1,080 controls, who were all Chinese Han individuals living in Shanghai. In this expanded study panel, only rs73318382 and rs57095329 showed an association with SLE (Table S2). These two SNPs are separated by 304 bp and are in strong linkage disequilibrium (LD; r2 = 0.81; Figure S2). When a Bonferroni correction was applied, the association of rs57095329 with SLE remained highly significant (P = 4×10−4). Given that rs57095329 is identified through our candidate region sequencing approach and not included in the HapMap database, it is not surprising that this SNP has not been included in commercial SNP arrays. Because published GWASs in SLE of both Asian and European populations detected association signals at rs2431697 and rs2431099 (Figure S3), 15 kb and 8 kb upstream from miR-146a TSS, respectively, we extended our genotyping of rs2431697 and rs2431099 using 1,896 Shanghai samples. Both SNPs showed significant association with the disease (Table S2), while rs57095329 produced the best association signal among the three SNPs in the same dataset (Table S2). Therefore, we focused on rs57095329 in the subsequent experiments.
We replicated the association between rs57095329 and SLE using a TaqMan genotyping assay in another two panels from Hong Kong, China, and Bangkok, Thailand. We also added 1,536 patients from the central China area to our mainland China cohort, and the newly added patients showed an allele frequency for rs57095329 very similar to that in the discovery panel (MAFs of 20.53% and 20.77%, respectively). This replication provided consistent evidence for the association, revealed by an allelic association analysis (Table 1). When all the samples were included (3,968 patients and 3,214 controls in total) to conduct a meta-analysis, there was strong evidence that the minor G allele of rs57095329 conferred a risk of SLE (Pmeta = 2.74×10−8, odds ratio [OR] = 1.29, 95% confidence interval [CI] = 1.18–1.40; Table 1). There was no significant difference among the ORs for the three independent cohorts (P = 0.33), when the Breslow–Day test installed in PLINK was used [28], although the SNP showed significant allele frequency differences in respective controls. Recessive mode of action seemed to be supported in the Chinese mainland cohort and the cohort from Thailand (OR = 2.47 and 2.11 for the two cohorts, respectively), compared with the allelic OR of 1.35–1.36 for the two cohorts. However, this was not supported by the result for the Hong Kong cohort, where the same OR was observed for both the recessive mode and the allelic test (OR = 1.18), reflecting certain variations among the different cohorts.
We also examined whether the genetic variant is specifically associated with disease risk in patients with lupus nephritis. Although only the discovery panel in the Chinese mainland cohort showed a significant association in a patient-only analysis, a similar trend was also observed in the Hong Kong and Bangkok cohorts, with a marginal P value of 0.093 and an OR of 1.105 when patients with nephritis were compared with patients without it (Table S3).
We explored the association between rs57095329 and miR-146a expression. Mature miR-146a levels were determined with a TaqMan microRNA assay in 86 healthy controls with known genotypes and available RNA samples. Compared with individuals with the AA genotype, individuals with heterozygous AG genotype for rs57095329 had lower levels of miR-146a (P = 0.0438; Figure 1), while individuals with GG genotype had the lowest miR-146a levels (P = 0.0197; Figure 1). This association indicates that rs57095329, located in the miR-146a promoter, may function by regulating the transcription activity and expression levels of miR-146a.
To explore molecular mechanisms of the association between rs57095329 and miR-146a expression, we examined whether the variant is functionally significant by altering the miR-146a promoter activity. We generated reporter gene constructs containing either rs57095329 allele and transfected different cell lines with the reporter plasmids, so that the effect of each allele on the miR-146a promoter activity could be evaluated in the context of the full-length promoter. First, the construct carrying the A allele had higher basal activity in Jurkat T cells than the construct carrying the risk-associated G allele, when a luciferase assay was performed 24 hours after electroporation (Figure 2). This finding is consistent with our previous observation of an association between reduced miR-146a expression and SLE disease. Moreover, when the cells were activated by PMA+Iono or anti-CD3 plus anti-CD28 antibodies after transfection, the induced activity of the promoter with the A allele remained higher (Figure 2). Similarly, an approximately 50% reduction in the activity of the promoter with the risk-associated G allele was observed in HeLa cells under both rested and PMA+Iono-activated conditions (Figure S4A). This difference in promoter activity was also consistently found in steady-state Raji B cells and 293T cells (Figure S4B and S4C). Considering these data together, our reporter gene assay showed that the disease-associated G allele reduced the promoter activity of miR-146a.
To examine whether allelic difference in promoter activity may be attributable to their different binding capacities for nuclear factors, two probes corresponding to the 24-bp miR-146a promoter region, centered on rs57095329, were synthesized and biotin-labeled for an electrophoretic mobility shift assay (EMSA), and unlabeled oligonucleotides were used as the “competitors”. Nuclear extracts were then prepared from resting and anti-CD3+anti-CD28-activated Jurkat cells. As shown in Figure 3, probe A formed much more DNA–protein complexes with the nuclear extracts from resting Jurkat cells than did probe G (lane 2 versus lane 7), indicating that the promoter carrying the A allele of rs57095329 binds more robustly to nuclear proteins. Once the cells were activated, both probes were able to bind more nuclear proteins. Similarly, in this case, probe A exhibited much stronger binding than probe G (Figure 3: lane 4 versus lane 9). All the DNA–protein complexes were reduced or abolished by the addition of excessive corresponding “competitor” oligonucleotides, demonstrating the binding specificity. Similar results were observed with PMA+Iono-stimulated Jurkat and HeLa cells (Figure S5).
The findings described above illustrated that rs57095329 alleles conferred differential binding affinity of nuclear extracts to the miR-146a promoter. To identify which proteins bind at or near this SNP to regulate the expression levels of miR-146a, we performed a bioinformatics search. The Genomatix online tool suggested that the multipotent transcription factor Ets-1 binds to the rs57095329 region (Figure S6). Interestingly, it has been shown that mutation of this predicted Ets-1-binding site resulted in a great reduction in the activity of an miR-146a promoter–reporter gene [29]. Here, we performed the following assays and further confirmed the pivotal role of Ets-1 in regulating miR-146a expression: the transient expression of Ets-1 greatly enhanced the reporter gene activity from the full-length miR-146a promoter in Jurkat cells, compared with that of another transcription factor, PBX1 (Figure 4A); knockdown of Ets-1 by small interfering RNA (siRNA) in Jurkat cells directly impaired the induction of pri-miR-146a upon T-cell activation (Figure 4B); and the overexpression of Ets-1 dramatically enhanced, whereas the knockdown of Ets-1 consistently reduced, the miR-146a promoter–reporter gene activity in HeLa cells (Figure S7A and S7B).
We assessed whether the allelic difference of rs57095329 in regulatory activity is attributable to different binding affinity for Ets-1. We cotransfected different amounts of Ets-1 with the reporter gene construct containing either the A or G allele of rs57095329 into HeLa cells. Increasing the protein levels of Ets-1 greatly enhanced the promoter activity of both constructs. However, the activity ratios of the two constructs (G/A) gradually decreased (Figure S8), suggesting that the inferior ability of the G-allele-containing sequence to bind Ets-1 could be compensated by increasing the levels of Ets-1. To examine the binding affinity more directly, we performed a promoter pulldown assay using streptavidin-conjugated agarose beads. When incubated with nuclear extracts from anti-CD3+anti-CD28-activated Jurkat cells, the biotin-labeled A probe bound more Ets-1 protein than the G probe, as demonstrated by western blotting analysis of total agarose beads precipitated proteins with an anti-Ets-1 antibody (Figure 4C). Taken together, these results demonstrate that rs57095329 alters Ets-1 binding, and the risk-associated G allele is less competent than the A allele in the regulation of miR-146a expression.
It is intriguing that our previous GWAS and that of others identified an association between a functional variant of ETS1 and SLE [8], [9]. Therefore, we investigated whether there is an interaction between the risk variants of the two genes, rs1128334 in ETS1 and rs57095329 in miR-146a. No epistatic effect was detected between the two variants (P = 0.46), when analyzed with a conditional logistic regression test, with the interaction between the two variants treated as a covariate using PLINK [28]. However, we observed additive effects of the risk alleles of miR-146a and ETS1, suggesting that individuals carrying two or more of these alleles are at greater risk than those carrying only one allele (Figure 5). This additive effect between ETS1 and miR-146a SNPs is also supported by a consistent increase in OR values in analysis of the enlarged samples containing 4,302 individuals with known genotypes for both SNPs (Table S4).
miRNAs have been shown to play an essential role in immune homeostasis, and aberrations in the miRNA-mediated regulation of immune-cell development and function has been linked to autoimmune diseases [30]. In an miRNA profiling study, we recently identified a significant reduction in miR-146a expression in lupus patients [23]. Here, we extended this study to determine why the expression of miR-146a is reduced in SLE patients. Prompted by the genetic association between the type I IFN pathway and the risk of SLE and by evidence that polymorphic variants in miRNA precursors can modulate miRNA biogenesis and disease risk, we sequenced key regions of pri-miR-146a and identified an SLE-associated SNP, rs57095329, within the promoter of miR-146a, which functionally affects miR-146a expression levels and thus contributes to the risk of SLE. The association of this variant with SLE was consistent in three independent cohorts from mainland China, Hong Kong, and Bangkok (Thailand). Individuals carrying the risk-associated G allele tended to express lower levels of miR-146a. To the best of our knowledge, this is the first report of an association between a functional genetic variant in an miRNA promoter and a human disease. It will be interesting to investigate the association between rs57095329 and SLE in other ethnic groups. Another functional variant located in the miR-146a precursor, rs2910164, has been associated with cancer development [26], [31], [32], but showed no significant association with SLE in our initial sequencing experiments.
Among the multiple immunological aberrations present in lupus patients, the type I IFN system is thought to play a crucial role in its pathogenesis [15]–[17]. Intriguingly, a number of genes involved in IFN signaling have already been associated with various autoimmune diseases, including SLE [33]. Functional variants in genes encoding key components of the IFN pathway, such as TYK2, IRF5, and STAT4, have been identified and characterized, and their association with SLE has been extensively replicated [34]–[39]. Our recent work characterized the role of miR-146a as a negative regulator of the type I IFN pathway by targeting key signaling proteins [23]. Here, the delineation of an SLE-susceptible variant of the miR-146a promoter further supports the notion that polymorphic variants linked to IFN pathway molecules contribute to the pathogenesis of lupus. miR-146a is embedded in a non-coding RNA with a previously unknown function, so our findings highlight the importance of exploring genetic variants in such regions, which have been more or less ignored in previous genetic studies.
Our findings underline the regulatory role of Ets-1 in miR-146a expression, and attribute the allelic difference of rs57095329 to different affinity for Ets-1. Rs57095329 is not located at the core sequence of the Ets-1-binding site (Figure S6), so it only causes an affinity difference, whereas Ets-1 recognition is still well preserved. Nevertheless, the risk-associated G allele of rs57095329 does affect the strongly conserved A residue near the Ets-1 core motif (Figure S6), highlighting the relevance of this SNP. Besides, this is a germ-line regulatory polymorphism and thus potentially functions in each cell type, as reflected in our consistent observation of the reduced activity of a reporter gene carrying the risk-associated G allele in various cell lines (Figure 2 and Figure S4). The attenuation of the promoter activity by the risk-associated G allele of rs57095329 thus accounts, at least partly, for the underexpression of miR-146a in lupus patients. Intriguingly, ETS1 has been characterized as a susceptibility gene for SLE in GWAS results from others and our group [8], [9]. The reduced expression of ETS1 was shown to be associated with the risk-associated allele of rs1128334 compared with the protective allele, identified in an allelic expression assay [9]. The finding that Ets-1 knockdown led to an inability to induce miR-146a expression in vitro (Figure 4B) was consistent with reduced miR-146a expression in patients with SLE who have reduced Ets-1 levels. It seems that both rs1128334 in ETS1 and rs57095329 in miR-146a may reduce the expression of miR-146a, through the reduced availability of Ets-1 and a reduced binding affinity for Ets-1, respectively. However, we did not detect interaction between the two variants. This may not be surprising because both of these variants only have a quantitative effect on their respective functions. Therefore, an additive effect is observed between the two variants (Figure 5 and Table S4) rather than a strong interaction, which would be the case if both of them totally abolished a function. We propose a working model for the genetic link between ETS1 and miR-146a to illustrate the genetic contribution to the reduced expression of miR-146a in SLE patients (Figure 6). We fully appreciate that Ets-1 can modulate a large collection of genes for their expression, and are not trying to limit the contribution of its 3′UTR SNP to SLE solely to affecting miR-146a expression. Yet our functional study, the genotype-expression data, and the additive effect of the two SNPs together provide an interesting connection between these two SLE susceptibility genes.
Recent SLE GWASs identified disease association of two SNPs (rs2431697 and rs2431099) that are upstream of miR-146a gene region [3], [7], [9] and our genotyping confirmed their association on our Chinese samples (Table S2). We therefore performed the following analysis to clarify the genetic signals of association across this region: 1) Imputation of rs57095329 into our Asian GWAS dataset, using the individuals of Asian ancestry from the 1000 genome project as the reference panel, suggested that this SNP represented an independent signal (conditional P value of 0.90); 2) rs2431099 and rs2431697 was in intermediate LD with each other while they were not in LD with rs57095329 (Figure S2); 3) Conditional analysis indicated that the association of rs57095329 with SLE was independent of those detected at rs2431697 and rs2431099, while the association of rs2431099 with SLE could be attributed to rs2431697 (Table S5); 4) Haplotype analysis showed that rs57095329 and rs2431697 were two independent SLE-associated loci, while rs57095329 had a stronger association in Chinese (Table S6); and 5) There was no correlation between miR-146a expression levels and the genotypes of rs2431697 or rs2431099 (Figure S9). Therefore we have newly identified a relevant SNP (rs57095329) by direct sequencing that the genotyping arrays in GWASs missed due to incomplete coverage, and these SNPs may confer a disease risk through different and independent mechanisms.
In conclusion, our findings add an miRNA gene, miR-146a, to the list of SLE-susceptible genes. A genetic variant of the miR-146a promoter, rs57095329, is functionally significant in modulating the expression of miR-146a by altering its binding affinity for Ets-1.
This study was conducted according to the principles expressed in the Declaration of Helsinki. Informed consent was obtained from all subjects. The Shanghai study was approved by the Institutional Review Board of Renji Hospital. The studies of the Hong Kong, Anhui, and Thai samples were approved by the Institutional Review Board of the University of Hong Kong and Hospital Authority, Hong Kong West Cluster, New Territory West Cluster, and Hong Kong East Cluster; the Research Ethics Committee of Anhui Medical University; and the Ethics Committee of the Faculty of Medicine, Chulalongkorn University, respectively.
We recruited 816 SLE patients and 1,080 sex- and age-matched controls, all of whom were from the Chinese Han population in Shanghai, China. Other Chinese mainland samples consisted of 1,536 SLE patients living in central China, collected by collaborators in Anhui province. For the independent replications, samples collected by collaborators in Hong Kong (case vs control: 1,152 vs 1,152) and Bangkok, Thailand (464 vs 982, respectively) were included. All SLE patients fulfilled the American College of Rheumatology (ACR) classification criteria for SLE, and 1,254 patients met the ACR criteria for lupus nephritis.
Consecutive overlapping amplicons corresponding to the miR-146a promoter region were amplified from genomic DNA extracted from peripheral blood leukocytes. The products were purified and directly sequenced on a 3730 automated sequencer (Applied Biosystems). The 452-bp DNA region around the miR-146a precursor was also amplified and sequenced. The primer pairs used are shown in Table S7.
In the replication stage, SNP rs57095329 was genotyped with the specified TaqMan SNP genotyping probes (Applied Biosystems). The assay was run on a 7900HT sequence detection system (Applied Biosystems) and the data were analyzed with the affiliated SDS software, version 2.3. The genotypes of rs57095329 were found to be in Hardy–Weinberg equilibrium (P>0.01) in the controls of all three cohorts. The average call rate for all samples was 92%.
Total RNA was extracted from peripheral blood leukocytes or cultured cells using TRIzol (Invitrogen), followed by reverse transcription using a reverse transcriptase kit obtained from Takara. To determine the quantity of pri-miR-146a, the cDNA was amplified by real-time PCR with SYBR Green RT–PCR kit (Takara), and the expression of RPL13A was used as the internal control. The primers used are shown in Table S7. To determine the quantity of mature miR-146a, the specific TaqMan MicroRNA Assay kit (Applied Biosystems) was used, and the expression levels were normalized to snRNA U6. The assays were performed on a 7900HT real-time instrument (Applied Biosystems). Relative expression levels were calculated using the 2−ΔΔCt method.
To create the miR-146a promoter–luciferase reporter constructs, three fragments of variable lengths, corresponding to the upstream region of the TSS of pri-miR-146a, were amplified and cloned into the pGL3-basic luciferase vector (Promega). To compare the activities of miR-146a promoters containing the different rs57095329 alleles, the full-length 1105-bp fragment was amplified from individual homozygous templates. The ETS1 overexpression vector was a kind gift from Dr Gang Pei, and the PBX1 overexpression plasmid was created by replacing the inserted ETS1 sequence with the PBX1 coding sequence. The primers used are shown in Table S7. All constructs were verified by sequencing.
Jurkat and Raji cells were grown in RPMI 1640 medium supplemented with 10% fetal bovine serum. These two cell lines were electroporated with 2 µg of the indicated luciferase reporter vector and 0.2 µg of a modified pRL-TK vector, using a nucleofector device (Amaxa). Alternatively, the reporter gene vectors were electroporated in combination with 1.5 µg of an ETS1- or PBX1-expressing vector. For the knockdown of ETS1, 3 µg of ETS1 siRNA or negative control oligonucleotides (all from GenePharma, Shanghai) were transfected. HeLa and 293T cells were grown in Dulbecco's modified Eagle's medium supplemented with 10% fetal bovine serum. These two cell lines were transfected using Lipofectamine 2000 (Invitrogen), with the ETS1- or PBX1-expressing vector or ETS1 siRNA alone, or in combination with 50 ng of the indicated luciferase reporter vector and 5 ng of a modified pRL-TK plasmid. Where indicated, an irrelevant “carrier” vector was added to ensure that equal total amounts of plasmid DNA were transfected among the groups. For cell activation, Jurkat and HeLa cells were stimulated with PMA (100 ng/mL; Sigma) and ionomycin (1 µM; Sigma) for the indicated times. Alternatively, Jurkat cells were activated with plate-bound anti-CD3 antibody (coating solution: 5 µg/mL; eBioscience) and soluble anti-CD28 antibody (2 µg/mL; eBioscience).
Cells were cultured for 24 hours or 48 hours after transfection with the reporter gene vectors together with the ETS1 expression vector or siRNA, respectively. The cells were then maintained resting or activated for 6 hours and lysed. Their luciferase activity was measured on a luminometer (LB960; Berthold) using the Dual-Luciferase Reporter Assay System (Promega). The ratio of firefly luciferase to Renilla luciferase was calculated for each well.
Jurkat and HeLa cells (1×107) were activated or left to rest for 2 hours, and then their nuclear proteins were extracted with a Nuclear Extract Kit (Active Motif), according to the manufacturer's protocol. The protein concentrations were determined with the DC Protein Assay Kit (Bio-Rad). Double-stranded allelic probes were synthesized and labeled with biotin by Takara (the sequence is shown in Figure S6). EMSA was carried out with a gel-shift kit purchased from Active Motif. The competition assay was performed by adding cognate unlabeled oligonucleotides. After incubation, the protein–DNA complexes were separated on a nondenaturing 6% polyacrylamide gel and then transferred to a nitrocellulose membrane (Millipore). The signals were detected using a luminoimage analyzer.
The pulldown assay was performed following a protocol described elsewhere [40], with slight modification. Biotin-labeled allelic probes were incubated with equal amounts of nuclear extract from activated Jurkat cells for 2 hours at room temperature, in the presence of streptavidin–agarose beads (GE Healthcare) and protein inhibitors. The precipitated protein–DNA complex was dissociated from the agarose beads by suspending the pellet in Laemmli sample buffer (Bio-Rad) and heating it. The supernatants were then subjected to SDS–PAGE. The proteins were transferred onto a PVDF membrane (Bio-Rad), blotted with an anti-Ets-1 antibody, and detected with ECL solution (Pierce). To evaluate the Ets-1 protein levels after the transfection of the overexpression vectors or siRNAs, the Jurkat and HeLa cells were lysed in RIPA buffer (Thermo Scientific), and the supernatants were similarly used for immunoblotting. Anti-ETS1, anti-GAPDH, and horseradish-peroxidase-conjugated secondary antibodies were all obtained from Santa Cruz Biotechnology.
For single SNP analysis, PLINK was used for the basic allelic test and other tests in the patients and the controls [28]. LD patterns were analyzed and displayed with HaploView [41]. Review manager was used to perform meta-analysis. IMPUTE version 2 was used to perform imputation. Other data were analyzed with GraphPad Prism 4 software, version 4.03. The nonparametric Mann–Whitney test was used to compare miR-146a expression between the genotype groups, and an unpaired t test was used to compare reporter gene activities. Two-tailed P values<0.05 were considered to be statistically significant.
|
10.1371/journal.pcbi.1007229 | Predictions of time to HIV viral rebound following ART suspension that incorporate personal biomarkers | Antiretroviral therapy (ART) effectively controls HIV infection, suppressing HIV viral loads. Suspension of therapy is followed by rebound of viral loads to high, pre-therapy levels. However, there is significant heterogeneity in speed of rebound, with some rebounds occurring within days, weeks, or sometimes years. We present a stochastic mathematical model to gain insight into these post-treatment dynamics, specifically characterizing the dynamics of short term viral rebounds (≤ 60 days). Li et al. (2016) report that the size of the expressed HIV reservoir, i.e., cell-associated HIV RNA levels, and drug regimen correlate with the time between ART suspension and viral rebound to detectable levels. We incorporate this information and viral rebound times to parametrize our model. We then investigate insights offered by our model into the underlying dynamics of the latent reservoir. In particular, we refine previous estimates of viral recrudescence after ART interruption by accounting for heterogeneity in infection rebound dynamics, and determine a recrudescence rate of once every 2-4 days. Our parametrized model can be used to aid in design of clinical trials to study viral dynamics following analytic treatment interruption. We show how to derive informative personalized testing frequencies from our model and offer a proof-of-concept example. Our results represent first steps towards a model that can make predictions on a person living with HIV (PLWH)’s rebound time distribution based on biomarkers, and help identify PLWH with long viral rebound delays.
| Antiretroviral therapy (ART) effectively controls HIV infection, holding HIV viral loads to levels undetectable by commercial assays. Therapy interruption is followed by rebound of viral loads to high, pre-therapy levels, but there is significant heterogeneity in the timing of the rebound to those high levels. Some rebounds occur within days, weeks, or even, rarely, years. Here we develop a mathematical model to characterize rebounds occurring within two months of treatment interruption. Li et al. (2016) report biological markers that correlate with the time between ART interruption and viral rebound. We incorporate this information to parametrize our model so that our model can make predictions on time to rebound tailored to the individual undergoing ATI. Our parametrized model can aid in design of clinical trials to study infection dynamics following treatment interruption. We also use our model to gain insight into the underlying within-host viral dynamics. For example, we refine previous estimates of viral recrudescence after ART interruption and determine a recrudescence rate of once every 2-4 days. Our results represent first steps towards a model that can make predictions on an person living with HIV’s rebound time based on personal biomarkers, and help identify patients with long viral rebound delays.
| Antiretroviral therapy (ART) for HIV infection can very effectively control the infection and hold the amount of circulating virus below the level detectable by clinical assays, improving both the quality and length of life. ART suspension generally is followed by HIV rebound to high viral loads [1], and consequently the standard of care for people living with HIV (PLWH) is to maintain life-long ART. However, there is significant heterogeneity in rebound times. In a pooled analysis of participants from six AIDS Clinical Trials Group (ACTG) analytic treatment interruption (ATI) studies to identify predictors of viral rebound, Li et al. reported widely varying times to viral rebound, with a significant number of participants maintaining viral suppression to undetectable levels for up to 2 or more months in the absence of ART [2]. In a follow-up study, Li and his team identified a cohort of post-treatment controllers (PTCs) from these ATI studies, who maintained viral loads ≤ 400 HIV RNA copies/mL for ≥24 weeks [3, 4]. Previous reports of these rare PTCs include the VISCONTI cohort, 14 PLWH who initiated ART within three months of their estimated date of infection who were able to control HIV infection for a prolonged period after stopping ART [5]. Results from the VISCONTI study and others suggest that PTCs may control HIV by a mechanism distinct from that of spontaneous HIV controllers [6, 7]. However, the factors that mediate delayed timing of HIV rebound are not well understood.
Since ART comes with a number of drawbacks including side-effects and cost, the search for biological indicators (biomarkers) of lasting ART-free HIV remission has become a priority in HIV cure research [8, 9]. Studies have already begun to bear fruit, with recent studies revealing a variety of immunological biomarkers for delayed rebound and infection control [2, 3, 10–13]. While such studies are informative, they offer limited insight into the mechanisms underlying viral rebound or post-treatment control. Mechanistic modeling inference has an established history of advancing our understanding of HIV [14, 15]. In this study we combine data on markers associated with rebound identified by Li et al. [2] with mechanistic mathematical models to gain deeper insight into mechanisms of viral rebound.
Modeling within-host HIV infection and treatment is a well-established field [14, 16–24]. By fitting models to clinical data, many parameters describing HIV dynamics such as the viral clearance rate, the infected cell death rate, and the viral burst size have been estimated [25–27]. Existing models have mainly focused on the kinetics of early infection and the effects of treatment. A few papers have focused on HIV control and the time to viral rebound after treatment cessation. These include those of Hill et al. [28, 29], Pinkevych et al. [30, 31] and Fennessey et al. [32], in which the authors all assume viral rebound is the outcome of latent cell activation. Pinkevych et al. used data from treatment interruption trials to provide the first estimates of latent cell activation rates that lead to observable viremia [30, 31]; using a related approach Fennessey et al. investigated SIV viral rebound in macaques infected with barcoded virus, to generate more detailed insights into viral rebound [32]. However, this modeling does not account for individual-level heterogeneity in viral rebound dynamics [33]. Hill et al. used continuous time branching processes, which are well-suited for small populations, to model within-host viral rebound dynamics. The primary results in Hill et al. [28, 29] are estimates of viral rebound time distributions, used in combination with careful and thoughtful consideration of within-host parameters to evaluate the needed efficacy of therapeutic agents that one day may be able to reduce the latent reservoir. In their model, Hill et al. assumed that latently infected cells may die or activate and that newly activated cells can die or generate infected cell offspring that are infected, as a proxy for tracking virus that in turn infects new cells [28, 29]. Thus, Hill et al. assumed that average viral growth immediately following viral recrudescence is, on average, exponential, which may not be the case.
In this study, we take up the hypothesis that latent cell activation causes viral rebound in the short term (<60 days) [28, 29], but that significantly delayed rebounds are associated with additional mechanisms of infection control, such as anti-HIV immune responses [23]. For example, there is evidence that T-cell exhaustion markers are predictive of shorter time to viral rebound [13] and that levels of HIV-specific T cell responses is associated with viral load after ATI [11]. We focus on short-term delays. We fit a simple stochastic model of viral rebound extended from our previous studies [21, 24] to the viral rebound data from the ACTG ATI studies [2]. In contrast to Pinkevych et al. [30] we address uncertainty in latent reservoir rebound dynamics, by modeling the time between a “successful” latent cell activation and detectable viremia stochastically and given by one of a variety of probability density functions. We integrate into our model biomarkers with observed impact on time to viral rebound, e.g., an individual’s expressed HIV reservoir, i.e., levels of cell-associated HIV RNA (HIV CA-RNA or CA-RNA), and ART regimen pre-analytic treatment interruption (ATI) [2, 34]. We discuss biological insight offered by parameter estimates from data, in particular on the average rate of latent cell activations that cause viral rebound [30, 31, 33]. Our model output is a cumulative probability density function for the probability of an individual’s viral rebound at time t. The output can be used for ATI clinical trial design; in particular, one can derive from our modeling a viral load testing schedule for participants to meet study objectives.
Our aim is to construct a model that predicts the viral rebound time, i.e., the time between suspension of therapy and detectable viremia. We model viral dynamics following cessation of therapy using the central assumption that activation of latently infected cells drives viral rebound. To estimate model parameters, we use data collected from ACTG ATI studies.
Written informed consent was provided by all study participants for use of stored samples in HIV-related research. This study was approved by the Pennsylvania State University Institutional Review Board, the Los Alamos National Laboratory Institutional Review Board, and the Partners Institutional Review Board.
The description of the data we employ and associated collection methodologies are fully explained in [2]. Briefly, participants in six ACTG ATI studies (ACTG 371 [35], A5024 [36], A5068 [37], A5170 [38], A5187 [39], and A5197 [40]) were included if they were on suppressive ART, received no immunologic interventions (e.g., therapeutic vaccination, interleukin-2), and had HIV-1 RNA less than 50 copies/ml at the time of ATI (N = 235 participants). We restricted the data we analyzed to the participants who showed viral rebound ≤ 60 days after ART cessation, in part to accommodate model simplifications, such as our assumption that the latent reservoir size remained constant from ATI to the time of viral rebound (see Model, below). Of the N = 210 participants who rebounded within 60 days of ATI, N = 84 met the following additional criteria for further study: 1) had peripheral blood mononuclear cells (PBMCs) and plasma available for HIV reservoir quantification while on ART prior to the ATI and 2) had cell-associated HIV-1 RNA (HIV CA-RNA) above the level of detection at ATI. Cell-associated DNA was also measured but did not have a significant association with time to viral rebound [2] and is neglected in our analysis. Finally, Li et al. noted that viral rebound delays were greater in study participants whose pre-ATI ART regimen contained non-nucleoside reverse transcriptase inhibitors (NNRTIs) [2]. We therefore distinguish between regimens containing a NNRTI (50/84 study participants) and those that do not (34/84 study participants).
Early after treatment interruption, most studies reported weekly viral load measurements, with the exception of A5170. In total, in the subset of study participants we study here (N = 84/235), 41 participants had approximately weekly or more frequent viral load measurements, while the remaining participants had viral loads measured more frequently than monthly, with a median of 7 days (range 1-35 days; interquartile range (IQR) 6-24 days). The timing of viral rebound was defined as sustained viral loads of at least 200 HIV RNA copies/mL.
Viral load data is shown in Fig 1a. In this present study we model the time of viral rebound, which occurs at some point between a study participant’s last undetectable and first detectable viral load measurement (threshold of detection 200 HIV RNA copies/mL). We will therefore use those time points to estimate model parameters for each ATI study participant. The times of last undetectable measurement and first detectable measurement are shown in Fig 1b as line segments spanning the detection window per study participant, with color indicating whether the ART regimen included (red) or excluded (blue) NNRTIs. Although the median time between viral load tests up to the time of detectable viremia, across the 84 study participants, is 7 days, the median time window between the last undetectable measurement and the first detectable viral load measurement, shown in Fig 1b, is 20 days (range 4-35 days; IQR 7-27 days).
We assume that activation of latently infected cells drives viral rebound [29, 30] and model the ensuing dynamics as illustrated in Fig 2. Not all latently infected cell activations cause viral rebound. We assume that activation is followed by rounds of viral replication, which may cause viral populations to grow to detectable levels, thereby causing viral rebound, but may also die out. We define q as the probability of extinction, i.e., the probability that the rounds of viral replication following latent cell activation die out, that is, that the activation of a latently infected cell does not cause viral rebound. We further define a “successful latent cell activation” as one that does cause viral rebound. In the time preceding a successful latent cell activation, we envision viral dynamics similar to those modeled in [24], with potentially many latent cell activations followed by a few rounds of viral replication, with the lineages ultimately going extinct. We also assume that any activations pre-ATI and resultant lineages go extinct as drug is still restricting viral spread (see Incorporating rebound indicators and and Discussion).
Finally, we assume that there is a delay between successful latent cell activation and detectable infection, where “detectable infection” corresponds to study participant viral loads exceeding 200 HIV RNA copies/mL. There is some debate as to the dynamics following latent cell activation. For example, in vitro observations suggests that an activated latently infected cell may produce significantly less virus than a productively infected cell [41]. Further, following the application of latency reversing agents, latently infected cells dynamics may not conform to the observed dynamics of productively infected cells [42], and latently infected cells may divide before they get fully activated and produce virus [28, 41]. We therefore avoid the common assumption that an activated latently infected cell is the same as a productively infected cell [21, 24, 28]. The need of target cells to infect in the proximity of an activated latently infected cell may also pose a challenge in initiating detectable infection in vivo. Finally, heterogeneity in viral growth rates (once there is enough virus for exponential growth) among individuals, due to difference in the infecting virus and host restriction factors, is also a factor in the early dynamics [30, 31, 33]. Because the dynamics of latent cell activation and infection spread before a detectable level of viremia is attained are unknown, we absorb these dynamics into a delay-time distribution, D(t) (Fig 2) reflecting these various sources of heterogeneity. We will test differing assumptions on this delay, for example taking a fixed or distributed delay (e.g. a lognormal distribution).
We restrict ourselves for now to short-term viral rebound (≤ 60 days). Over that short time period, we can assume that the latent reservoir size is approximately constant with value L0 (<3% estimated reduction, assuming that while the virus is undetectable the latent reservoir continues to decay with a half-life of 44 months [43, 44]). We assume that latently infected cells are activated at an average rate a, so activated cell influx occurs at the constant rate aL0. In general, we anticipate variability in the activation rate a; for example, most latently infected cells are memory cells [45], and activation may depend on encounters with cognate antigen, whose rates may be expected to vary according to the rarity of the associated pathogen. However, per the law of large numbers, given the large latent reservoir size in most individuals [46], the time to detection will approximately depend on the average activation rate. Mathematically we employ a multi-type branching process framework to derive an expression or viral rebound at time t. We use this probability to ultimately derive likelihood functions and fit the model to data.
To sum up, in our model we assume that heterogeneity in observations of viral rebound across individuals result from four components. Two depend on the individual study participant and derive from observed correlates of time to viral rebound: (i) The replication-competent reservoir size L0, which we will assume is reflected in the HIV CA-RNA level [2, 34], and (ii) the probability q that the activation of a latently infected cell does not cause viral rebound, which may be affected by the pre-ATI ART regimen [2]. The remaining two components arise from stochastic within-host dynamics: (iii) the rate of latent cell activations that are “successful”, which we model as a Poisson rate, and will result in an exponential distribution in time to first successful activation, and (iv) the stochastic delay between successful activation and detection, which we model using different stylized distributions.
We begin by estimating model parameters. We then go on to discuss the consequences of our estimates, in particular the impact of NNRTI-containing drug regimens on viral rebound times, and our estimates of successful latently infected cell activation times. Finally, we discuss how our model predictions can be used to inform clinical trial design.
We use the Davidon-Fletcher-Powell optimization algorithm to estimate model parameters as(1 − q0), k, and parameters associated with the delay distribution, for our viral rebound model that neglects NNRTI status, Eq (1), and accounts for NNRTI status, Eq (2). A summary of these parameter estimates are provided in Tables 2 and 3, respectively, with complete details provided in S1 and S2 Tables, respectively. We use the Akaike Information Criterion (AIC) to compare how well the models explain the data.
We show in Fig 7a predictions on the mean viral rebound time as a histogram across ATI study participants, depending on HIV CA-RNA and sorted by NNRTI status. Our model predicts that the mean time to viral rebound is delayed in individuals including NNRTIs in their pre-ATI ART regimen. This is not a surprise, as we were motivated to include the effects of NNRTIs by the observations of statistically significant delay in [2]. However, our model predictions offers additional nuance: the variation in time to rebound is also larger. Fig 7b shows a histogram of model-predicted standard deviations in time to viral rebound across the study population, again sorted by NNRTI status. The increased variability may be explained by different NNRTI drugs and individualized differences in rates of drug metabolism. We recover similar results using parameter estimates derived from other delay distribution assumptions (Table 1; not shown). The wider variation suggests that rebound times in PLWH taking NNRTIs are less predictable, and in the context of clinical trials, the inclusion of individuals on an NNRTI-based regimen may alter sample size and power calculations.
We can use our parameter estimates to predict a distribution in average time to successful activation across the 84 individuals. For clarity we stress that “average” is at the individual level, since the latent reservoir is a heterogeneous population of cells. The reservoir is primarily composed of latently infected memory cells, of which there are several types, e.g., central memory, transitional memory and effector memory [44]. Investigations of latent reservoir decay after initiation of therapy show multiple decay phases [57–59], which suggest a heterogeneous population of latently infected cells with different half-lives. Further, a memory cell is only activated when it encounters its cognate antigen, e.g. a bacterial or viral peptide that it recognizes. Finally, recent evidence suggests that the reservoir is in part made up of clonal populations [60]. Therefore the activation time will vary across latently infected memory cells, depending on the rate at which an individual’s immune system is challenged with different antigens. However, if the number of cells is large, as we expect it to be in PLWH, rebound times are well described by the average.
Our model predicts that the average frequency of successful activations in an individual is given by as(1 − q)log10 (CA-RNA). Fig 8 shows a histogram for the predicted time to successful activations, 1/[as(1 − q)log10 (CA-RNA)], across the ATI study population, assuming a Weibull-distributed (Fig 8a) or fixed (Fig 8b) detection delay.
Table 4 gives our model-predicted frequency of successful reactivation from latency, depending on the delay distribution assumption, with model-predicted means and 5th, 50th (median), and 95th percentiles. Previous modeling by Pinkevych et al. [30] estimated that the average frequency of successful reactivation from latency is about once every 6 days, and a range of 5-8 days. The result of Pinkevych et al. [30] addressed neither potential heterogeneity in the unclear latent cell kinetics post-activation [41, 42, 61], nor heterogeneity in viral growth rates as a source of rebound delays [31, 33]. In our modeling, when we neglect heterogeneity by taking a fixed delay between successful activation and infection detection, we predict an average of about 5 days (90% confidence interval 3-8 days), on par with the results of Pinkevych et al. However, when we account for that heterogeneity via alternate, i.e., non-fixed delay densities D(t), we recover a shorter average frequency of successful reactivation from latency, with successful activations occurring on average every 2-4 days (see Table 4).
Our result also adds nuance to previous estimates. While we estimate that the mean frequency of successful activation from latency is once every 2 (Weibull distributed detection delay) to 5 (fixed detection delay) days, we also report the population range of estimated frequencies within the 5th and 95th quantiles at most once every 3-8 days, approximately, neglecting heterogeneity and assuming a fixed detection delay, and at least 1-3 days, approximately, assuming a Weibull-distributed delay. For the purposes of further calculation, we can also use these data to estimate a population-level distribution for an individual’s average frequency of successful reactivation from which we can sample, see S2 Fig. We find that the data shown in Fig 8 is best described by a lognormal distribution, tact ∼ Lognormal(μ, σ2) with μ = 0.74 (standard error 0.03) and σ = 0.29 (standard error 0.02) with a Weibull-distributed delay, which gives a mean average frequency of successful reactivation of once every 2.2 days with 90% confidence interval (1.3,3.4) days (S2a Fig), and μ = 1.61 (standard error 0.03) and σ = 0.30 (standard error 0.02), neglecting heterogeneity and assuming a fixed delay, which gives a mean of once every 5.2 with 90% confidence interval (3.1,8.1) days (S2b Fig).
We envision the primary use of our parameter estimation to be ATI clinical trial design. Our model predicts a probability density function for a study participant’s time to viral rebound following ATI, depending on that participant’s pre-ATI log10(HIV CA-RNA) with level and ART regimen. We can use the predictions to plan testing intervals to capture rebound times to within study-objective specificity. We use tools from survival analysis, treating 1-PVR(t), 1-(cumulative probability of viral rebound function), as the survival function, S(t) = 1 − PVR(t).
We have developed a simple stochastic model to predict the time to viral rebound for people living with HIV (PLWH) who undergo analytic treatment interruption (ATI). Our model predictions take the form of probability distributions in time, which can be interpreted as survival functions. Our model integrates PLWH-specific data to individualize predictions based on (1) pre-ATI ART regimen and (2) pre-ATI HIV CA-RNA, both shown to be associated with times to viral rebound [2, 3, 34]. Thus it distinguishes itself from previous studies, which have emphasized population average predictions [28–30].
In our modeling we focused on short-term viral rebound following ATI, which we restrict to ≤ 60 days. We used our model, parametrized with ATI study participant data [2], to provide a population distribution of “successful” latent cell activation rates, i.e., the rate of latent cell activations that induce viral rebound. We recover an average frequency of activations leading to viral rebound of approximately 2-4 days, depending on our assumption on the delay between activation and the increase of viremia to detectable levels. The most appropriate delay distribution cannot be resolved with existing data. Our estimate of the successful activation rate is shorter than that of Pinkevych et al. [30] who estimate an average of about 6 days between successful activations, and a range of 5-8 days [30]. Pinkevych et al. modeled viral rebound by assuming that the number of study participants controlling HIV at time t after ART cessation is exponentially decaying in time, and attribute that decay rate to latent cell activation. With data from treatment interruption trials, they provided the first estimates of latent cell activation rates that lead to observable viremia [30]. While our parameter estimation approach differs, the primary reason we obtain a shorter frequency of latent cell activation is that Pinkevych et al. did not explicitly address heterogeneity in the events leading to viral rebound. This heterogeneity comes from many sources including the kinetics of events that occur within a latent cell post-activation [41, 42, 61], including bursty transcription from the HIV-1 promoter that can lead to toggling between latent and pre-productive infection [64–66] and heterogeneity in subsequent viral growth rates [31, 33]. We account for heterogeneity from all sources, via the delay distribution D(τ), and thus we refine their previous estimate. However, our activation dynamics and delay distribution ignore any host factors, such as HLA allelles, that may generate inter-individual variability in time to detectable viremia. These and other host factors may become increasingly important as we strive to improve personalized predictions of viral rebound distributions.
We hypothesize that viral rebounds occurring after many months, or even years [2, 67, 68]—or not at all [5], i.e., post-treatment control—are associated with host mechanisms, such as immune responses [2, 5, 13, 23, 69]. Markers of T-cell exhaustion are associated with times to viral rebound [13]. Li et al. also noted that study participants treated early (within 6 months of exposure to HIV) showed later post-ATI viral rebound than those treated during the acute phase of infection [2]. The recent observation of rebound following ATI delayed by 7.4 months, in an individual treated within an estimated 10 days of exposure to HIV [70], is consistent with previous observations that early ART treatment is associated with delayed viral rebound timing and increased chances of post-treatment control [2, 3]. Delayed ART initiation appears to decrease the chances of sustained post-treatment viral control, potentially due to the expanding diversity of the HIV reservoir [71], immune exhaustion [72], or increasing CTL escape mutations that will diminish the effectiveness of cell-mediated immune responses [73].
As a consequence of our hypothesis, predictions of late viral rebounds may require more sophisticated models. In this present study, modeling viral rebound as a consequence of viral replication engendered by latent cell activation, we excluded data from ATI study participants whose viremia returned only after many months or years [2]. The current model acts as a necessary foundation upon which immunologic data can be incorporated when they are available to model post-treatment control. However, late viral rebounds form only a minority of dynamics following ATI, just 25 or ∼10% of 235 ATI study participants in [2], and thus our modeling, which predicts the probability of viral rebound at time t for PLWH following ATI in the short term, describes post-ATI dynamics in the majority of individuals. We can therefore reasonably use our modeling to aid in ATI clinical trial design, in particular determining post-ATI testing frequency, according to study objectives. If using our model for study design and implementation, we would advise reevaluation of the few individuals who achieve 60 days with no rebound with a more sophisticated approach including testing for immunological markers of HIV control [2, 5, 13, 23, 69].
From our modeling we can also identify gaps in data that may be invaluable in improving modeling insights into viral rebound and control. In creating our individualized predictions, we focused on some of the first biomarkers identified to be associated with delays in viral rebound [2]; work ongoing in identifying further covariates of control [2, 3, 10–13] will aid in further refining models of HIV rebound or control [28–30, 32, 74]. When commenting on clinical trial design we must acknowledge the practical limitations, since these studies rely on PLWH who take time out of their own lives to regularly get tested. Our data shows a median of 12 clinic visits per patient, with many making upwards of 30 clinic visits [2]. These volunteers make their contributions with the knowledge that the scientific advancements gained may not benefit them. Therefore calling for frequent testing—which from a modeling perspective would be ideal—is problematic. However we note, due to the lack of data in the first week following ATI, that we made simplifying assumptions, such as neglecting ART decay kinetics in study participants whose pre-ATI ART regimen excluded NNRTIs. Therefore we call for more regular data collection in the 1-2 weeks if possible, which may be particularly illuminating in characterizing rapid viral rebound and thus improving parameter estimation, without over-burdening clinical trialists or generous volunteers.
In our modeling we neglected the inherent heterogeneity of the latent reservoir and associated latent cell activation rates. Latently infected cells are in majority memory cells [45], each of which may be specific for a pathogen or set of pathogens. There is evidence that the reservoir is composed of clonal populations [53, 60, 75], so there may be genetically homogeneous subsets of cells, but even cells in clones may exhibit differing activation dynamics. However, in modeling viral rebound for large latent reservoir sizes, we neglect this heterogeneity in favor of the mean activation rate. Heterogeneity in activation rate becomes important as latent reservoir sizes gets small, and we move towards elimination, but that is not the focus of this present study.
The latent cell activation rate, a, is part of the recrudescence rate, as(1 − q0), which we estimate from the data and which we assume to be the same for all individuals. Note that we can only estimate the parameter combination as(1 − q0), and not a itself. Future modeling efforts may aim to rectify this possibly by including additional data, such as direct estimates of the pre-ATI viral reservoir size and viral growth rates and viral set points post-ATI for each study participant.
In accounting for the delay in viral rebound observed in Li et al. (2016) associated with inclusion of NNRTIs in the pre-ATI ART regimen [2], we assumed that the probability that latent cell activation induces viral rebound decays expontentially to q0 at rate k. Our estimate for k, which should account for the rate of drug decay post-ATI, was at least order of magnitude lower than the mean NNRTI concentration decay rates in plasma [55]. One explanation is that most T-cells and latently infected cells reside in lymphatic tissues [76, 77]. The pharmacokinetics and pharmacodynamics in the lymphatic tissues are not clear, although there is evidence that drug penetration is lower [78, 79]; commensurately, drug clearance may also be slower. Since our model crudely treats the whole body as homogeneous, the expression for decaying effectiveness of the drug must average the dynamics in different tissues, potentially explaining the inconsistencies.
It is also interesting to note that the NNRTI efavirenz’s clearance is dependent on CYP450 2B6 gene polymorphism and there are certain polymorphisms that increase plasma half life such that drug levels above the 95% inhibitory concentration maybe present for > 21 days after treatment interruption [48]. However, we acknowledge that our estimates of the rate at which drug loses effectiveness will need to be further refined and validated. In future studies we will attempt to refine q(t) and more carefully address its variability, potentially, by considering specific drug pharmacokinetic/pharmacodynamics in tissues, when available, and data on viral dynamics post-rebound to inform the reproductive ratio in absence of ART. It is also possible that ART decay may occur in a biphasic manner with our estimate of k reflecting the terminal elimination phase. Additional pharmacokinetic studies are needed to explore this possibility.
We also simplified our model by taking a constant reservoir size, neglecting factors contributing to long-term reservoir decay such as latent cell death and proliferation [45]. Preceding viral rebound, we anticipate that the latent reservoir would continue to decay at on-therapy rates [43, 80] resulting from these dynamics. But the reservoir is typically large and its decay is slow, with a 44 month half-life on average [43, 80], so in our 60-day rebound period, the average reservoir size would decrease by less than 3%. Thus our constant reservoir size assumption is reasonable. We can extend our simple model to include latent cell proliferation and death, derived in the S1 Text. Intriguingly, the resulting expression for probability of viral rebound gives the natural activation rate a as, in principle, an identifiable parameter, in contrast to our simple model, for which only the successful latent cell activation rate, (1 − q)aL0, is identifiable. Unfortunately, to disentangle a from other parameters in the extended model, we require more refined data, as discussed based on the mathematics in the S1 Text. Such data may be difficult to obtain; as it is, the data from Li et al. [2] which we employ is the most extensive and well-curated ATI study data currently available.
In integrating study participant data into our viral rebound model, we made the assumption that log10(cell-associated HIV RNA) is proportional to the size of the replication-competent portion of the latent reservoir, for which we currently have no direct methods to measure [46]. We were motivated by Li et al., who showed a negative correlation between HIV CA-RNA and time to detectable viremia [2]. Li et al. also noted a negative correlation between pre-ATI viral load, measured using single copy assays, and time to detectable viremia [2]. Since on-therapy viremia may be associated with rounds of replication resulting from latent cell activations [24], pre-ATI viral load may be a better measure of the replication-competent portion of the latent reservoir. It also may not: if ART efficacy is near 100%, there may be only very limited rounds of replication. In that case, on-therapy viremia would represent primarily virus released from activated latently infected cells, and would not be a good measure of replication competence, although the higher the fraction of cells releasing HIV, the higher the probability that the population will include cells releasing replication competent virus [53]. Regardless, we did not incorporate this observation into the current model because of the paucity of data (41/235 ATI study participants with viral load measured above the level of detection), but we acknowledge that the pre-ATI viral load may be a better predictor of replication competent latent reservoir size.
Our simple model predictions suggest that any given treated HIV+ individual who will undergo what we term ‘short-term rebound’ is capable of reproducing almost the full amount of variability in rebound times seen across the 84 study participants under consideration (c.f. Figs 5b, 5c and 7), with some finite probability, which varies across study participants. For example, the model predicts that the probability of rebound within two weeks of ATI across study participants ranges from 11% to 90%, depending on the study participant, and the probability of rebound after 6 weeks ranges from 10−5% to 15% (Fig 5b and 5c). Predicted mean times to viral rebound vary from days, without NNRTIs in the pre-ATI ART regimen, to weeks with NNRTIs (Fig 7). One interpretation is that the observed variability in rebound times is driven most strongly by stochasticity in the delay between successful activation and viral detection, and in the time to reactivation, with an individual’s level of HIV CA-RNA influencing the time to viral detection (Fig 3b). However, we suspect that the similarity in predicted viral rebound time probability density functions in the absence of NNRTIs, in particular Fig 5a and 5b, also importantly reflects the uncertainty in the data with respect to actual rebound times; recall that the median time between the last undetectable viral load measurement and first detectable viral load measurement is 20 days (mean 18 days). While parameter estimation with more frequent observations would improve predictions and resolve this question, such data may only be obtained with difficulty, as again the data we employ here derives from the most extensive ATI study data currently available. We therefore advise mindfulness of the uncertainty as modeling of viral rebound advances.
While we acknowledge many limitations, our simple model, parametrized with ATI study participant data, offers individualized predictions of time-to-viral rebound following ATI. Our results offer insight into latent cell activation dynamics, can inform future modeling and predictive work, and can be used to inform testing periods in ATI clinical trial design. This study represents first steps towards a model that can make accurate predictions of a person living with HIV (PLWH)’s rebound time distribution based on personal characteristics, and help identify PLWH with expected long viral rebound delays.
|
10.1371/journal.pgen.1006773 | Recurrent promoter mutations in melanoma are defined by an extended context-specific mutational signature | Sequencing of whole tumor genomes holds the promise of revealing functional somatic regulatory mutations, such as those described in the TERT promoter. Recurrent promoter mutations have been identified in many additional genes and appear to be particularly common in melanoma, but convincing functional data such as influence on gene expression has been more elusive. Here, we show that frequently recurring promoter mutations in melanoma occur almost exclusively at cytosines flanked by a distinct sequence signature, TTCCG, with TERT as a notable exception. In active, but not inactive, promoters, mutation frequencies for cytosines at the 5’ end of this ETS-like motif were considerably higher than expected based on a UV trinucleotide mutational signature. Additional analyses solidify this pattern as an extended context-specific mutational signature that mediates an exceptional position-specific vulnerability to UV mutagenesis, arguing against positive selection. We further use ultra-sensitive amplicon sequencing to demonstrate that cell cultures exposed to UV light quickly develop subclonal mutations specifically in affected positions. Our findings have implications for the interpretation of somatic mutations in regulatory regions, and underscore the importance of genomic context and extended sequence patterns to accurately describe mutational signatures in cancer.
| Cancer is caused by somatic mutations that alter cell behavior. While such mutations typically occur in protein-coding genes, recent studies describe individual positions in gene regulatory regions (promoters) that are recurrently mutated in many independent tumors. This suggests that positive selection could be acting on these non-coding mutations, and that they may contribute to carcinogenesis. However, proper interpretation of recurrent mutations requires a detailed understanding of how such mutations arise in the absence of selection pressures, referred to as mutational heterogeneity. In this paper, we describe a distinct sequence signature that characterizes nearly all highly recurrent promoter mutations in melanoma. Additional analyses support that this sequence mediates an exceptional local vulnerability to UV-induced mutagenesis, explaining why mutations are frequently observed in these positions. Importantly, cultured cells exposed to UV light quickly developed mutations specifically in the expected sites. Our results have important implications for the interpretation of recurrent somatic mutation patterns in non-coding DNA.
| A major challenge in cancer genomics is the separation of functional somatic driver mutations from non-functional passengers. This problem is relevant not only in coding regions, but also in the context of non-coding regulatory regions such as promoters, where putative driver mutations are now mappable with relative ease using whole genome sequencing[1,2]. One important indicator of driver function is recurrence across independent tumors, which can be suggestive of positive selection. However, proper interpretation of recurrent mutations requires a detailed understanding of how somatic mutations occur in the absence of selection pressures. Somatic mutations are not uniformly distributed across tumor genomes, and regional variations in mutation rates have been associated with differences in transcriptional activity, replication timing as well as chromatin accessibility and modification[3–5]. Impaired nucleotide excision repair (NER) has been shown to contribute to increased local mutation density in promoter regions and protein binding sites[6,7]. Additionally, analyses of mutational processes and their sequence signatures have shown the importance of the immediate sequence context for local mutation rates[8]. Still, our understanding of mutational heterogeneity is incomplete, and it is not clear to what extent such effects can explain recurrent somatic mutations in promoter regions, which are suggested by some studies to be particularly frequent in melanoma despite several other cancer types approaching melanoma in terms of total mutation load[9,10].
To characterize somatic promoter mutations in melanoma, we analyzed the sequence context of recurrently mutated individual genomic positions occurring within +/- 500 bp of annotated transcription start sites (TSSs), based on 38 melanomas subjected to whole genome sequencing by the Cancer Genome Atlas[10,11]. Strikingly, of 17 highly recurrent promoter mutations (recurring in at least 5/38 of tumors, 13%), 14 conformed to an identical 6 bp sequence signature (Fig 1a and 1b). Importantly, the only exceptions were the previously described TERT promoter mutations at chr5:1,295,228, 1,295,242 and 1,295,250[12,13] (Fig 1c). The recurrent mutations occurred at cytosines positioned at the 5’ end or one base upstream of the motif CTTCCG (Fig 1d), and were normally C>T or CC>TT transitions (Fig 1a). Similar to most mutations in melanoma they thus occurred in a dipyrimidine context and were compatible with UV-induced damage through cyclobutane pyrimidine dimer (CPD) or 6–4 photoproduct formation[8,14]. Out of 15 additional positions recurrently mutated in 4/38 tumors (11%), 13 conformed to the same pattern, while the remaining two showed related sequence contexts (Fig 1a). Many less recurrent sites also showed the same pattern (S1 Table). The signature described here matches the consensus binding sequence of ETS family transcription factors (TFs)[15], and the results are consistent with recent reports showing that ETS promoter sites are often recurrently mutated in melanoma[9] and that such mutations preferably occur at cytosines upstream of the core TTCC sequence[16]. Thus, while recurrent promoter mutations are common in melanoma, they consistently adhere to a distinct sequence signature, which may argue against positive selection as a major causative factor.
The recurrently mutated positions were next investigated in additional cancer cohorts, first by confirming them in an independent melanoma dataset[17] (S2 Table). We found that the identified hotspot positions were often mutated also in cutaneous squamous cell carcinoma (cSCC)[18] (S3 Table) as well as in sun-exposed skin[18,19], albeit at lower variant frequencies (S1 Fig, S4 Table). Additionally, one of the mutations, upstream of DPH3, was recently described as highly recurrent in basal cell skin carcinoma[20]. However, we did not detect mutations in these positions in 13 non-UV-exposed cancer types (S5 Table). The hotspots are thus present in UV-exposed samples of diverse cellular origins, but in contrast to the TERT promoter mutations they are completely absent in non-UV-exposed cancers. This further supports that recurrent mutations at the 5’ end of CTTCCG elements are due to elevated susceptibility to UV-induced mutagenesis in these positions.
Next, we considered additional properties that could support or argue against a functional role for the recurrent mutations. We first noted a general lack of known cancer-related genes among the affected promoters, with TERT as one of few exceptions (Fig 1a and S1 Table, indicated in blue). Secondly, the recurrent promoter mutations were not associated with differential expression of the nearby genes (Fig 1a and S1 Table). This is in agreement with earlier investigations of some of these mutations, which gave no conclusive evidence regarding influence on gene expression[9,16,20], although it should be noted that significant association was lacking also for TERT in this relatively small cohort. Lastly, we found that when comparing different tumors there was a strong positive correlation between the total number of the established hotspot positions that were mutated and the genome-wide mutation load, both in melanoma (Fig 2a; Spearman’s r = 0.88, P = 2.8e-13) and in cSCC (S3 Table; r = 0.78, P = 0.026). This is again compatible with a passive model involving elevated mutation probability in the affected positions. Importantly, this contrasted sharply with most of the major driver mutations in melanoma, which were detected also in tumors with lower mutation load (Fig 2b, S3 Table). These different findings further reinforce the CTTCCG motif as a strong mutational signature in melanoma.
We next investigated whether the observed signature would be relevant also outside of promoter regions. As expected, numerous mutations occurred in CTTCCG sequences across the genome, but notably we found that recurrent mutations involving this motif were always located close to actively transcribed TSSs (Fig 3a, 3b and 3c). We further compared the frequencies of mutations occurring at cytosines in the context of the motif to all possible trinucleotide contexts, an established way of describing mutational signatures in cancer[8]. As expected, on a genome-wide scale, the mutation probability for cytosines in CTTCCG-related contexts was only marginally higher compared to corresponding trinucleotide contexts (Fig 4a). However, close to TSSs, the signature conferred a striking elevation in mutation probability compared to related trinucleotides, in particular for cytosines at the 5’ end of the motif and most notably near highly expressed genes (Fig 4b–4d). Recurrent promoter mutations in melanoma thus conform to a distinct sequence signature manifested only in the context of active promoters, suggesting that a specific binding partner is required for the element to confer elevated mutation probability.
CTTCCG elements have in various individual promoters been shown to be bound by ETS factors such as ETS1, GABPA and ELF1[21], ELK4[22], and E4TF1[23]. This suggests that the recurrently mutated CTTCCG elements could be substrates for ETS TFs. As expected, matches to CTTCCG in the JASPAR database of TF binding motifs were mainly ETS-related (S6 Table). Notably, recurrently mutated CTTCCG sites were evolutionarily conserved to a larger degree than non-recurrently mutated but otherwise similar control sites, further supporting that they constitute functional ETS binding sites (S2 Fig). This was corroborated by analysis of top recurrent CTTCCG sites in relation to ENCODE ChIP-seq data for 161 TFs, which showed that the strongest and most consistent signals were for ETS factors (GABPA and ELF1) (S3 Fig).
The distribution of mutations across tumor genomes is shaped both by mutagenic and DNA repair processes. Binding of TFs to DNA can increase local mutation rates by impairing NER, and strong increases have been observed in predicted binding sites for several ETS factors[6,7]. It is also established that contacts between DNA and proteins can modulate DNA damage patterns by altering conditions for UV photoproduct formation[24–27]. In upstream regions of XPC -/- cSCC tumors lacking global NER, we found that several of the established hotspot sites were mutated (S7 Table) and that the CTTCCG signature still conferred elevated mutation probabilities compared to relevant trinucleotide contexts (Fig 5), although to a lesser extent than in melanomas with functional NER (Fig 4). Transcription-coupled NER (TC-NER) may still be active in XPC -/- tumors, and the signature could thus theoretically arise due to blocking of TC-NER at CTTCCG elements. However, only upstream regions, which should not be subjected to this process, were considered in this analysis. Additionally, TC-NER is strand-specific[14], but the signature was present independently of strand orientation relative to the downstream gene in XPC -/- tumors (Fig 5a and 5b). The signature described here is thus unlikely explained by impaired NER alone, and other mechanisms, such as inhibition of other repair-related processes or favorable conditions for UV lesion formation at the 5’ end of ETS-bound CTTCCG elements, may contribute.
Finally, we sought to experimentally test our proposed model that the observed promoter hotspots are due to localized vulnerability to mutagenesis by UV light. We subjected human melanoma cells and keratinocytes to daily UV doses for a period of 5 or 10 weeks and used an ultrasensitive error-correcting amplicon sequencing protocol, SiMSen-Seq[28], to assay two of the observed promoter hotpots for mutations: RPL13A, the most frequently mutated site in the tumor data, and DPH3[10,20] (Fig 6a). Between 36k and 82k error-corrected reads (>20x oversampling) were obtained for each of 16 different conditions (Fig 6b and 6c). Strikingly, subclonal mutations appeared specifically in expected positions at both time points and in both cell lines at a frequency reaching up to 2.9% of fragments (RPL13A, 10 weeks of exposure), while being absent in non-exposed control cells (Fig 6d and 6e). As predicted by the tumor data, mutations occurred primarily at cytosines upstream of the TTCCG motif, with lower-frequency mutations occurring also in the central cytosines. Few mutations were observed outside of the TTCCG context despite presence of many cytosines in theoretically vulnerable configurations in the two amplicons (Fig 6d and 6e, underscored). Interestingly, an atypical substitution pattern displayed by the DPH3 hotspot in the tumors, involving C>A and C>G in addition to the expected C>T transitions (Fig 1a), was mirrored also in the UV exposure data (Fig 6d). Our results from UV exposure of cultured cells further reinforce that recurrent mutation hotspots in promoters in melanoma arise due to an exceptional vulnerability to UV mutagenesis in these positions.
In summary, we demonstrate that recurrent promoter mutations are common in melanoma, but also that they adhere to a distinct sequence signature in a strikingly consistent manner, arguing against positive selection as a major driving force. This model is supported by several additional observations, including lack of cancer-relevant genes, lack of obvious effects on gene expression, presence of the signature exclusively in UV-exposed samples of diverse cellular origins, and strong positive correlation between genome-wide mutation load and mutations in the affected positions. Crucially, exposing cells to UV light under controlled conditions efficiently induces mutations specifically in affected sites. These results point to limitations in conventional genome-wide derived trinucleotide models of mutational signatures, and imply that extended sequence patterns as well as genomic context should be taken into account to improve interpretation of somatic mutations in regulatory DNA.
Whole-genome sequencing data for 38 skin cutaneous melanoma (SKCM) metastases were obtained from the Cancer Genome Atlas (TCGA) together with matching RNA-seq data (dbGap accession phs000178.v9.p8). Mutations were called using SAMtools[29] (command mpileup with default settings and additional options -q1 and–B) and VarScan[30] (command somatic using the default minimum variant frequency of 0.20, minimum normal coverage of 8 reads, minimum tumor coverage of 6 reads and the additional option –strand-filter 1). Mutations where the variant base was detected in the matching normal were not considered for analysis. Mutations overlapping germline variants included in the NCBI dbSNP database, Build 146, were removed. The genomic annotation used was GENCODE[31] release 17, mapped to GRCh37. The TSS of a gene was defined as the 5’most annotated transcription start. Somatic mutation status for known driver genes was obtained from the cBioPortal[32,33].
RNA-seq data was analyzed with respect to the GENCODE[31] (v17) annotation using HTSeq-count (http://www-huber.embl.de/users/anders/HTSeq) as previously described[34]. Differential gene expression between tumors with and without mutations in promoter regions was evaluated using the two-sided Wilcoxon rank sum test.
The SKCM tumors were analyzed across the whole genome or in regions close to TSS, in which case only mutations less than 500 bp upstream or downstream of TSS were included. For the analysis of regions close to TSS the genes were divided in three tiers of equal size based on the mean gene expression level across the 38 SKCM tumors.
The February 2009 assembly of the human genome (hg19/GRCh37) was downloaded from the UCSC Genome Bioinformatics site. Sequence motif and trinucleotide frequencies were obtained using the tool fuzznuc included in the software suite EMBOSS[35]. The mutation probability was calculated as the total number of observed mutations in a given sequence context across all tumors divided by the number of instances of this sequence and by the number of tumors.
The evolutionary conservation of genome regions was evaluated using phastCons scores[36] from multiple alignments of 100 vertebrate species retrieved from the UCSC genome browser. The analyzed regions were 30 bases upstream and downstream of the motif CTTCCG located less than 500 bp from TSS.
Binding of transcription factors at NCTTCCGN sites was evaluated using normalized scores for ChIP-seq peaks from 161 transcription factors in 91 cell types (ENCODE track wgEncodeRegTfbsClusteredV3) obtained from the UCSC genome browser.
Whole genome sequencing data from sun-exposed skin, eye-lid epidermis, was obtained from Martincorena et al., 2015[19]. SAMtools[29] (command mpileup with a minimum mapping quality of 60, a minimum base quality of 30 and additional option –B) was used to process the data and VarScan[30] (command mpileup2snp counting all variants present in at least one read, with minimum coverage of one read and the additional strand filter option disabled) was used for mutation calling.
Whole genome sequencing data from 8 cSCC tumors and matching peritumoral skin samples was obtained from Durinck et al., 2011[37]. Whole genome sequencing data from cSCC tumors and matching peritumoral skin from 5 patients with germline DNA repair deficiency due to homozygous frameshift mutations (C940del-1) in the XPC gene was obtained from Zheng et al., 2014[18]. SAMtools[29] (command mpileup with a minimum mapping quality of 30, a minimum base quality of 30 and additional option –B) was used to process the data and VarScan[30] (command mpileup2snp counting all variants present in at least one read, with minimum coverage of two reads and the additional strand filter option disabled) was used for mutation calling. For the mutation probability analysis of cSCC tumors with NER deficiency, an additional filter was applied to only consider mutations with a total coverage of at least 10 reads and a variant frequency of at least 0.2. The functional impact of mutations in driver genes was evaluated using PROVEAN[38] and SIFT[39]. Non-synonymous mutations that were considered deleterious by PROVEAN or damaging by SIFT were counted as driver mutations.
A375 melanoma cells were a gift from Joydeep Bradbury and HaCaT keratinocyte cells were a gift from Marica Ericson. Cells were grown in DMEM + 10% FCS + gentamycin (A375) or pen/strep (HaCaT) (Thermo Scientific). Cells were treated in DMEM in 10 cm plates without lids with 36 J/m2 UVC 254 nm (equivalent to 6 hour daily dose at 0.1J/m2/min[40], CL-1000 UV crosslinker, UVP), 5 days a week for 10 weeks. Cells were split when confluent and reseeded at 1:5. Cells were frozen at -20°C.
DNA was extracted based on Tornaletti and Pfeifer [41]. Briefly, cell pellets were lysed in 0.5 ml of 20 mM Tris-HCl (pH 8.0), 20 mM NaCl, 20mM EDTA, 1% (w/v) sodium dodecyl sulfate, 600 mg/ml of proteinase K, and 0.5 ml of 150 mM NaCl, 10 mM EDTA. The solution was incubated for two hours at 37°C. DNA was extracted twice with phenol-chloroform and once with chloroform and precipitated by adding 0.1 vol. 3 M sodium acetate (pH 5.2), and 2.5 volumes of ethanol. The pellets were washed with 75% ethanol and briefly air-dried. DNA was dissolved in 10 mM Tris-HCl (pH 7.6), 1 mM EDTA (TE buffer) (all from Sigma Aldrich). DNA was treated with RNAse for 1 hr at 37°C and phenol-chloroform extracted and ethanol precipitated before dissolving in TE buffer.
To detect and quantify mutations we applied SiMSen-Seq (Simple, Multiplexed, PCR-based barcoding of DNA for Sensitive mutation detection using Sequencing) as described[28]. Briefly, barcoding of 150 ng DNA was performed in 10 μL using 1x Phusion HF Buffer, 0.1U Phusion II High-Fidelity polymerase, 200 μM dNTPs (all Thermo Fisher Scientific), 40 nM of each primer (PAGE-purified, Integrated DNA Technologies) and 0.5M L-Carnitine inner salt (Sigma Aldrich). Barcode primer sequences are shown in S8 Table. The temperature profile was 98°C for 3 min followed by three cycles of amplification (98°C for 10 sec, 62°C for 6 min and 72°C for 30 sec), 65°C for 15 min and 95°C for 15 min. The reaction was terminated by adding 20 μL TE buffer, pH 8.0 (Invitrogen, Thermo Fisher Scientific) containing 30 ng/μL protease from Streptomyces griseus (Sigma Aldrich) at the beginning of the 65°C incubation step. Next, 10 μL of the diluted barcoded PCR products were amplified in a 40 μL using 1x Q5 Hot Start High-Fidelity Master Mix (New England BioLabs) and 400 nM of each sequencing adapter primer. Adapter primers are shown in S8 Table. The temperature profile was 95°C for 3 min followed by 40 cycles of amplification (98°C for 10 sec, 80°C for 1 sec, 72°C for 30 sec and 76°C for 30 sec, with a ramp rate of 0.2°C/sec). The 40 μL PCR products were then purified using Agencourt AMPure XP beads (Beckman-Coulter) according to the manufacturers’ instructions using a bead to sample ratio of 1. The purified product was eluted in 20 μL TE buffer, pH 8.0. Library concentration and quality was assessed using a Fragment Analyzer (Advanced Analytical). Final libraries were pooled to equal molarity in Buffer EB (10 mM Tris-HCl, pH 8.5, Qiagen) containing 0.1% TWEEN 20 (Sigma Aldrich).
Sequencing was performed on an Illumina NextSeq 500 instrument at TATAA Biocenter (Gothenburg, Sweden) using 150 bp single-end reads. Raw FastQ files were subsequently processed as described[28] using Debarcer Version 0.3.0 (https://github.com/oicr-gsi/debarcer). Sequence reads with the same barcode were grouped into families for each amplicon. Barcode families with at least 20 reads, where ≥ 90% of the reads were identical, were required to compute consensus reads. FastQ files were deposited in the Sequence Read Archive under BioProject ID PRJNA375726.
|
10.1371/journal.pgen.1007568 | cis-regulatory architecture of a short-range EGFR organizing center in the Drosophila melanogaster leg | We characterized the establishment of an Epidermal Growth Factor Receptor (EGFR) organizing center (EOC) during leg development in Drosophila melanogaster. Initial EGFR activation occurs in the center of leg discs by expression of the EGFR ligand Vn and the EGFR ligand-processing protease Rho, each through single enhancers, vnE and rhoE, that integrate inputs from Wg, Dpp, Dll and Sp1. Deletion of vnE and rhoE eliminates vn and rho expression in the center of the leg imaginal discs, respectively. Animals with deletions of both vnE and rhoE (but not individually) show distal but not medial leg truncations, suggesting that the distal source of EGFR ligands acts at short-range to only specify distal-most fates, and that multiple additional ‘ring’ enhancers are responsible for medial fates. Further, based on the cis-regulatory logic of vnE and rhoE we identified many additional leg enhancers, suggesting that this logic is broadly used by many genes during Drosophila limb development.
| The EGFR signaling pathway plays a major role in innumerable developmental processes in all animals and its deregulation leads to different types of cancer, as well as many other developmental diseases in humans. Here we explored the integration of inputs from the Wnt- and TGF-beta signaling pathways and the leg-specifying transcription factors Distal-less and Sp1 at enhancer elements of EGFR ligands. These enhancers trigger a specific EGFR-dependent developmental output in the fly leg that is limited to specifying distal-most fates. Our findings suggest that activation of the EGFR pathway during fly leg development occurs through the activation of multiple EGFR ligand enhancers that are active at different positions along the proximo-distal axis. Similar enhancer elements are likely to control EGFR activation in humans as well. Such DNA elements might be ‘hot spots’ that cause formation of EGFR-dependent tumors if mutations in them occur. Thus, understanding the molecular characteristics of such DNA elements could facilitate the detection and treatment of cancer.
| cis-regulatory modules (CRMs) are critical for the development and evolution of all organisms. CRMs integrate the information that a single cell or group of cells receives and, in response, trigger changes in cellular and tissue fate specification [reviewed in 1]. The Drosophila melanogaster leg imaginal disc, which gives rise to the entire leg and ventral body wall of the adult fly, provides an attractive model system for studying the molecular mechanisms of cellular fate integration at the CRM level and the consequent execution of developmental programs that pattern an entire appendage [reviewed in 2]. Leg imaginal discs are initially specified early during embryonic development through the activation of the transcription factors Distal-less (Dll) and Sp1 in distinct groups of cells in each thoracic segment [3–5]. These groups of cells segregate from the embryonic ectoderm to become the leg imaginal discs, the precursors of the adult legs and ventral thorax. During larval stages, the leg discs proliferate, and defined expression domains of the signaling molecules Wg (ventrally expressed) and Dpp (dorsally expressed) activate Dll through the DllLT CRM in the center of the leg imaginal disc, where the wg and dpp expression patterns abut each other (Fig 1A) [6–9]. Slightly later in development, medial leg fates are established by the feed-forward activation of dachshund (dac) by Dll through the dacRE CRM [2, 10]. During subsequent growth of the leg disc, partially overlapping Dll and dac expression domains are maintained by autoregulation. These Dll and dac expression domains, together with the most proximal domain marked by homothorax (hth) expression, create a rudimentary proximal-distal (PD) axis [2] (Fig 1A).
The initial PD axis defined by Dll, Dac, and Hth is further refined by an additional signaling cascade mediated by the Epidermal Growth Factor Receptor (EGFR) pathway [11, 12]. Like Dll, the EGFR pathway is initially triggered by Wg and Dpp, which activate two types of EGFR ligands in the center of the leg disc [11, 12]. One is the neuregulin-related ligand Vein (Vn) and the second is the TGF-α-like ligand Spitz (Spi), which requires metalloproteases of the Rhomboid (Rho) family for processing and secretion [reviewed in 13]. The local activation of vn and rho family members in the center of the leg disc creates an EGFR organizing center (EOC), a local source of secreted Vn and Spi that activate EGFR signaling in neighboring cells. EGFR signaling in turn results in the activation of a series of downstream target genes that are expressed in nested concentric domains that pattern the future tarsus, the distal-most region of the adult leg [11, 12, 14, 15] (Fig 1A).
The mechanism by which EGFR signaling patterns the distal leg is not fully understood. One model suggests that EGFR ligands, produced in the EOC, function as morphogens, acting on neighboring cells to generate distinct transcriptional outputs in a concentration-dependent manner. Consistent with this idea is the observation that gradually reducing EGFR activity by raising flies carrying a temperature-sensitive Egfr allele (Egfrtsla) at increasing temperatures results in gradually more severe leg truncations [11]. However, although consistent with a morphogen model, this result is complicated by the fact that in addition to the EOC, there are other sources of EGFR ligands expressed in rings that appear later in leg development [14]. This additional EGFR activation would also be compromised in Egfrtsla experiments, leaving open the question of the degree to which tarsal PD patterning is due solely to EOC activity. An alternative model posits that the activation of EGFR in the center of the leg disc triggers only local transcriptional outputs, and that alternative sources of EGFR ligands, in combination with indirect transcriptional cascades, are responsible for specifying fates that are further from the EOC.
The EOC/morphogen model predicts that eliminating the production of EGFR ligands from the EOC will have long-range consequences. In contrast, if alternative, non-EOC sources of EGFR ligands play a role in leg patterning, eliminating only the EOC would produce only local defects in distal leg patterning. To distinguish between these models, we searched for CRMs responsible for the expression of EGFR ligands and ligand-processing proteases in the EOC, with the idea that we could specifically eliminate EOC expression by deleting these CRMs. We identified EOC CRMs for vn and rho (vnE and rhoE, respectively) and showed that they are necessary for EOC expression of these genes, respectively. However, although EOC expression is eliminated, simultaneous deletion of these CRMs causes only local PD patterning defects and tarsal truncations comparable to mild Egfr perturbations in the distal tarsus. These results suggest that the EOC is required for activating local EGFR responses in the center of the leg disc, implying that other sources of EGFR ligands, controlled by non-EOC CRMs, further elaborate the tarsal PD pattern. Finally, we also performed rigorous genetic and biochemical analysis of the vn and rho EOC CRMs, and used the discovered regulatory logic to predict additional CRMs, many of which are active in the Drosophila leg. Together, these data reveal a common regulatory logic for gene activation in the distal leg that is used by many genes, in addition to vn and rho.
To understand the molecular mechanism by which the EGFR signaling pathway is activated in the center of leg imaginal discs during larval stages, we searched for leg disc enhancer elements controlling the expression of EGFR ligands and ligand-processing proteases implicated to function in this process [11, 14]. We scanned the genomic regions of vn and rho using in vivo lacZ reporter assays (Fig 1B and 1G and S1 Table) and defined minimal enhancers (vnE– 654 bp and rhoE– 544 bp) that recapitulate the expression pattern of these genes in the center of leg discs during development (Fig 1C and 1H), as well as in the serially homologous antennal discs (S1A and S1B Fig). The vnE- and rhoE-lacZ transgenes exhibited earlier expression (starting at ~71h PEL for vnE and ~82h PEL for rhoE; Fig 1C and 1H) than detected by in situ for vn and rho (Fig 1D and 1I), perhaps because of the greater sensitivity of the anti-ßgal staining, and suggest that the genes might be expressed earlier than previously thought [11, 12, 14, 15]. Our search for leg disc enhancers across the vn locus uncovered only vnE, while in rho we identified two additional rho leg disc enhancers (rhoLLE1 and rhoLLE2 (Fig 1G, LLE stands for ‘late leg enhancer’) that drive expression in ring patterns starting in mid-third instar leg discs (90-92h PEL) (S1C and S1D Fig). Although these enhancers do not participate in EOC formation, they are active at later developmental stages and drive expression in medial/proximal ring patterns and are thus likely to be additional sources of EGFR activity (S1C and S1D Fig).
We also re-examined the expression pattern of additional EGFR ligands and proteases using enhancer-reporter assay (S1E, S1I and S1L Fig; S1 Table), in situ hybridization (S1F, S1J, S1M and S1O Fig; S2 Table) and available enhancer trap lines (S1G Fig) and found that roughoid (ru) (as previously reported [11]) and spitz (spi) (S1G and S1J Fig), but not Keren (Krn) or gurken (grk) (S1M and S1O Fig), were expressed in leg discs during third larval instar. Curiously, ru expression was only detected by an enhancer trap (ruinga-lacZ) and by a newly identified enhancer, ruLLE, that recapitulates the ruinga expression pattern (S1H Fig) but was not detected by in situ hybridization (S1F Fig) (see also Campbell 2002). spi was expressed broadly in leg discs (S1J Fig), and this pattern was recapitulated by a ~10 kb region that includes its promoter and introns (S1K Fig). Although there are five additional rho-family proteases in Drosophila [16], previous genetic analysis suggests that rho and ru are the most relevant [11, 14]. Further, because ru did not show expression in early L3 leg discs (and see below for additional genetic tests), and spi expression was ubiquitous, we focused on vnE and rhoE as the primary CRMs active in the leg disc EOC.
To assess the requirement of the vnE and rhoE CRMs for vn and rho expression, we deleted them from their native genomic loci using CRISPR/Cas9-mediated genome editing ([17–19]; see Materials and Methods) and assessed the phenotypes of these alleles (vnvnE-Df and rhorhoE-Df). We found that these deficiencies abolished the expression of these genes, respectively, only in the EOC of the legs (Fig 1E and 1J). The lack of expression in the enhancer deletion alleles was restored when the wild type enhancers were resupplied in their native genomic positions (Fig 1F and 1K). Therefore, we conclude that vnE and rhoE are necessary and sufficient for vn and rho expression in the EOC, respectively.
Individually, both vnvnE-Df and rhorhoE-Df are viable as homozygotes, exhibit normal leg disc patterning (S2A and S2C Fig), and form morphologically normal and functional legs (S2B and S2D Fig), consistent with previous reports that vn and rho single mutants do not affect the leg disc or adult leg pattern [11, 12]. However, when we examined the combined effect of these deficiencies in rhorhoE-Df vnvnE-Df double mutant flies we found that the expression of EGFR downstream genes C15 and aristaless (al) was abolished in these animals (Figs 2A, 2B, S2E and S2F), and the expression of BarH1/H2, a pair of more proximally expressed PD genes [20], collapsed from a ring pattern to a central circular domain in the leg disc (Fig 2B). In agreement with the leg disc pattern changes, adult rhorhoE-Df vnvnE-Df double mutants exhibited distal leg truncations that lack a pretarsus and parts of tarsal segment 5 (Fig 2N). rhorhoE-Df vnvnE-Df double mutant flies die in late pupal stages most likely because of an inability to exit the pupal case.
A sequence comparison between D. melanogaster and D. virilis, two Drosophila species that diverged from each other ~50 million years ago [21], revealed that vnE is well conserved (45.8% identity over 0.65 kb) and at a similar location upstream of the D. virilis vn transcription start site. In contrast, rhoE could not be identified by sequence homology in D. virilis. These observations prompted us to ask if the orthologous D. virilis vnE (vnE-D.vir) could substitute for the function of D. melanogaster vnE and rescue the rhorhoE-Df vnvnE-Df phenotype. We performed the swap of enhancers (see Materials and Methods) and we found that, indeed, the leg imaginal discs of rhorhoE-Df vnvnE-D.vir flies had normal PD patterning (Fig 2C) and normal adult legs (Fig 2Q). This result suggests that the function of vnE has been maintained over tens of millions of years and this enhancer element plays a conserved role in limb development.
Rho is an EGFR ligand-processing metalloprotease that has the potential to cleave the membrane-bound ligands Spi, Krn, and Grk in order to convert them into active secreted forms, while Vn is expressed as a secreted form that does not require Rho function [reviewed in 13]. Although we did not detect any expression of Krn and grk in leg discs (S1M and S1O Fig), this does not exclude the possibility that these genes function in leg disc development at a level of expression below what is detected in our in situ hybridization experiments. To address this possibility, we performed genetic experiments and found that the single null mutants Krn27-7-B [22] and grkΔFRT [23], and the double mutant grkΔFRT; Krn27-7-B, do not exhibit any leg disc patterning defects (Fig 2D) or adult leg phenotypes (Fig 2R). In addition, rhorhoE-Df vnvnE-Df Krn27-7-B triple mutant (S2G and S2H Fig) and grkΔFRT; rhorhoE-Df vnvnE-Df Krn27-7-B quadruple mutant (Fig 2E) leg discs had similar defects as rhorhoE-Df vnvnE-Df double mutants (Fig 2B), even though the quadruple mutant larvae died at late L3, just before pupation. These results support our conclusion that Krn and Grk are unlikely to be involved in leg development.
The remaining rho-dependent EGFR ligand, Spi, is expressed broadly in leg discs (S1J and S1K Fig) and is a good candidate for participating in EOC activity under the temporal and spatial control of Rho. To confirm the role of Spi, we used RNAi (see Materials and Methods) to examine the phenotypes of animals depleted for both spi and vn in leg discs. We found that, indeed, Spi is the EGFR ligand processed by Rho in the center of leg discs, because spi vn double RNAi (driven by Dll-Gal4) caused loss of expression of the downstream EGFR gene C15, and the near elimination of Bar expression (Fig 2J). This phenotype is stronger than any other combination of EGFR pathway components, similar to Egfrtsla mutants grown at the restrictive temperature of 30°C (Fig 2L and 2P). In addition, in animals depleted for spi and vn using RNAi we observed leg truncations (Fig 2O) similar to those observed in Egfrtsla mutants at 30°C (Fig 2P). Taken together, these results suggest that Vn and Spi are likely the only ligands that activate EGFR signaling during fly leg development.
The triple ru1 rho7M43 vnL6 mutant, but not the rho7M43 vnL6 double mutant, produces a strong leg truncation phenotype, similar to Egfrtsla animals grown at 30°C, suggesting that Ru is involved in patterning the adult leg together with Vn and Rho [11]. vnL6 is a nonsense mutation and a null by genetic criteria [24, 25]. rho7M43 is also a null allele [16], although we, as well as previous studies [16], were unable to identify any amino acid changes in the rho coding sequence of this allele. ru1 is a nonsense mutation that leads to a premature stop codon after residue 55, prior to the Rhomboid domain, suggesting that it is also a null allele [26]. A potential caveat to this conclusion is that ru1/Df (including Dfs ruPLLb and ruPLJc) results in a stronger ‘rough-eye’ phenotype than the ru1 homozygote, implying that ru1 is a hypomorph [16, 26]. However, ru, together with several other genes, is located in the intron of the protein tyrosine phosphatase encoding gene, Ptp61F, which plays a role in EGFR/MAPK signaling (S1E Fig) [27]. Consequently, deficiencies that remove ru could also affect MAPK/EGFR signaling by reducing Ptp61F expression, and could potentially lead to stronger phenotypes compared to the cleaner ru1 allele. Taken together, these observations suggest that ru1 is likely to be a null mutation.
Notably, rho and ru are physically close to each other on chromosome 3L, with rhoE ~55 kb away from the ru promoter, raising the possibility that rhoE could also regulate ru (Figs 1G and S1E). To test this possibility, we examined the lacZ expression pattern driven by the ruinga enhancer trap [28] in the background of the homozygous rhorhoE-Df (see Materials and Methods). We did not detect any effect of rhorhoE-Df on ruinga-lacZ expression in leg discs (S2K and S2L Fig), suggesting that ru is not regulated by rhoE.
Because the triple ru1 rho7M43 vnL6 mutant produces adult leg truncations [11] that are stronger than those observed in our rhorhoE-Df vnvnE-Df double mutant, we carried out additional experiments to address a potential role for ru in leg disc patterning. In the first experiment, instead of examining adult legs we examined ruinga rhorhoE-Df vnL6 triple mutant clones in leg discs (see Materials and Methods). Notably, leg disc patterning in these mutant discs was similar to the pattern observed in the rhorhoE-Df vnvnE-Df double mutant (Fig 2F), and even a small patch of WT tissue in the center of the leg disc could restore a normal PD pattern (Fig 2G). In a second test, we generated ru1 rho7M43 vnvnE-Df triple mutant clones and, as in the previous experiment, we observed the loss of C15 and collapse of BarH1 expression (Fig 2H), similar to the rhorhoE-Df vnvnE-Df double mutant, and a rescue of C15 expression if some distal cells remain wild type (Fig 2I).
Together, these results suggest that ru does not contribute significantly to EOC activity in the early L3 stage to pattern the L3 imaginal disc. Instead, these results suggest a model in which EOC activity is mediated primarily by vnE and rhoE, while the later rings of EGFR activation are controlled by a distinct set of enhancers (e.g. rhoLLE1, rhoLLE2, and ruLLE) (S1C, S1D and S1H Fig), and that this second wave of EGFR activity is important for patterning medial regions of the adult leg. In addition, these data suggest that ru, and perhaps other rho-like family members, plays a role later in leg development through its ring-like expression pattern to ultimately impact adult leg patterning.
Previous studies have underscored the importance of the Wg and Dpp pathways for EGFR activation in the center of leg discs [11, 12]. Using the vnE and rhoE enhancer elements, we have been able to address this question in greater detail. We generated mutant clones of arrow (arr), an obligate co-receptor in Wg signaling, and Mothers against dpp (Mad), a downstream effector of Dpp signaling, at different time points, and assessed the requirement of these pathways for vnE and rhoE activation. Both Wg and Dpp pathways are necessary for the initiation of vnE-lacZ expression in late L2 larval stage (Fig 3A and 3E), while clones made early in L3 stage did not affect vnE-lacZ expression (Fig 3B and 3F). rhoE-lacZ expression was lost when either Wg or Dpp activity was removed during L2 or early L3 (Fig 3C and 3G) but became independent of these pathways later in mid-L3 (Fig 3D and 3H).
In addition to Wg and Dpp, at the early larval stages of leg disc development there are two other factors that are crucial for leg specification and growth–the homeodomain transcription factor Distal-less (Dll) [29] and the Zn finger transcription factor Sp1 [4, 5]. Dll mutant clones induced at any larval stage abolished vnE-lacZ expression (Figs 3I and S3A). In addition, ectopic expression of Dll activated vnE not only in leg discs but in other imaginal discs (S3C, S3D and S3E Fig), as long as Wg and Dpp were available in these tissues at the time of clone induction (S3C, S3D and S3E Fig). These results suggest that Dll is required for vnE activity. Similarly, rhoE-lacZ expression also required Dll at all developmental times (Figs 3K and S3B).
We also examined the requirement of Sp1 for vnE and rhoE activation. We found that vnE activation requires Sp1, either when the entire animal was mutant or in clones (Figs 3J and S3F). This requirement is not mediated by Dll because Dll expression remained intact in mutant clones (Fig 3J) and in leg discs from Sp1 homozygous animals (S3F Fig). In contrast, Sp1 was dispensable for rhoE-lacZ expression (Figs 3L and S3H). In addition, although Sp1 is required for the activation of vnE at L2 larval stage (Figs 3J and S3F), at the beginning of L3 larval stage Sp1 was no longer required for vnE (S3G Fig). We also assessed if vnE and rhoE are regulated by buttonhead (btd), an Sp1 paralog that is co-expressed with Sp1 in leg discs [5]. We found that neither EOC enhancer requires btd (S3I and S3J Fig) and it is unlikely that rhoE requires both Sp1 and Btd redundantly since we did not detect Sp1/Btd binding sites or in vivo binding at rhoE for Sp1 (see below). Together, these results support a model in which vnE activation requires Wg and Dpp together with Dll and Sp1; later, vnE activity becomes independent of Wg, Dpp and Sp1, but still requires Dll (Fig 3Q). Similarly, although the timing differs, rhoE requires initial input from Wg, Dpp, and Dll but later only requires Dll for its maintenance (Fig 3R). The differential onset of expression between the two enhancers might depend on the differential requirement for Sp1.
To investigate if EGFR activity is required for vnE and rhoE, we examined the expression driven by these CRMs in the background of mutants for EGFR pathway components. vnE- and rhoE-driven expression was normal in pntΔ88 [30] mutant clones or Egfrtsla [31] mutant clones at the restrictive temperature (Fig 3M, 3N, 3O and 3P). Capicua (Cic), another downstream component of EGFR [32], is expressed in leg discs (S3K Fig) but was also not required for vnE and rhoE activity (S3L and S3M Fig).
We next carried out epistasis experiments using the MARCM technique [33] in which we overexpressed one vnE or rhoE input and removed another. We excluded Sp1 from this analysis because Sp1 sometimes affects Dll expression making results difficult to interpret [5]. For both vnE and rhoE, we found that while ectopic activation of Dll induced the activity of these enhancers in wildtype tissue (Fig 4A, 4E, 4C and 4G), in clones compromised for either Wg or Dpp signaling neither vnE nor rhoE were activated (Fig 4B, 4F, 4D and 4H). Dll was also unable to induce vnE-lacZ expression in ectopic clones in other imaginal discs when Wg and Dpp signaling was compromised (S3C, S3D and S3E Fig). Further, consistent with previous results [6, 7, 34], ectopic Wg and Dpp pathway activity induced vnE- and rhoE-lacZ expression and created additional EOCs in leg discs when these clones were located close to an endogenous source of Dpp and Wg, respectively (Fig 4I, 4M, 4K and 4O). However, when these clones were also mutant for Dll, these pathways were not able to activate either vnE or rhoE, and hence EGFR signaling (Fig 4J, 4N, 4L and 4P).
Our genetic analysis suggests a complex interplay between the signaling pathways Wg and Dpp and the transcription factors Dll and Sp1 on the vnE and rhoE enhancers. To investigate the configuration of binding sites and the transcription factor grammar of these CRMs, we searched for putative binding motifs using available position weighted matrices (PWMs) [35] and computational methods for identifying consensus Pan (downstream effector of Wg signaling), Mad (downstream effector of Dpp signaling), Dll and Sp1 binding motifs [36]. We performed a comprehensive in vivo mutagenesis analysis for both enhancers (Fig 5A). We mutagenized the enhancer elements by progressively adding (one at a time) mutations (S3 Table) in putative binding sites for each transcription factor (Fig 5A), starting with those that best match consensus binding sites and proceeding to more degenerate binding sites. Because the information from the enhancer bashing experiments (Fig 1B and 1G; S1 Table) revealed that parts of the enhancers containing multiple sites for each of the TFs can not drive intact expression patterns, we inferred that only having the full set of binding sites gives full expression patterns. Based on the combined analysis between the mutagenesis and the enhancer bashing data we found that there are a large number of binding sites important for vnE activation—14 Pan binding sites, 12 Mad sites, and 11 Dll sites (Figs 5A, 5B, 5D, S4A and S4B); mutagenesis of subsets of these binding sites leads only to reduction of enhancer-driven expression (S4A and S4B Fig). In contrast, for each TF, there were fewer binding sites important for rhoE activation—4 Pan, 3 Mad and a single Dll binding site (Fig 5A, 5C and 5E). Curiously, in the case of Dll we found 5 additional putative sites in rhoE that were not required for enhancer activity in optimal laboratory conditions (S4E and S4F Fig; S3 Table). In general, the identified binding sites for the two enhancers had an additive effect on the expression levels of vnE and rhoE because partially mutated enhancers drove patchy expression and progressively diminished levels of reporter expression (S4A, S4B, S4C and S4D Fig). We also confirmed the binding of the TFs involved in vnE and rhoE regulation by in vitro binding assays, suggesting that they act directly to regulate these enhancers (Fig 5F).
It is striking that vnE contains many more binding sites for each TF compared to rhoE. In addition to the differential requirement for Sp1, this difference may also contribute to the earlier timing of vnE activation compared to rhoE, because the larger number of binding sites might render vnE more sensitive to lower TF concentrations.
Consistent with the genetic requirement for Sp1, we identified two putative Sp1 binding sites in vnE. However, when we mutagenized them reporter gene expression was unaltered (Fig 5B and 5D). Therefore, we scanned the enhancer by EMSA using overlapping fragments (S2 Table) in order to identify additional Sp1 binding sites in an unbiased manner. We found that Sp1 binds with low affinity to some Mad binding sites (Fig 5F). Because both Sp1 and Mad can bind to some of the same binding sites, loss of vnE-lacZ expression when Mad sites are mutated may be a consequence of eliminating all Mad and some of the Sp1 inputs.
Because Sp1 and Dll are co-expressed during leg development, we also scanned all of vnE using overlapping oligos (S2 Table) to determine if these proteins might bind cooperatively to DNA. For these experiments we used full-length Dll and nearly full-length Sp1 proteins (see Materials and Methods). Although these experiments confirmed Dll binding to its binding sites, we failed to detect any cooperative binding between Dll and Sp1. Taken together, our results suggest that Sp1 regulates vnE through two Sp1 binding sites and some shared binding sites with Mad.
The vnE and rhoE regulatory inputs that we discovered here resemble one previously characterized in DllLT [9], in that they are all activated by the combinatorial input of Wg, Dpp, Dll and/or Sp1 [5]. These findings prompted us to test if there might be a battery of CRMs that is regulated in the leg disc by these same inputs. To test this idea, we first determined the genome-wide in vivo binding profiles of Dll and Sp1 using chromatin immunoprecipitation followed by deep sequencing (ChIP-seq) in third instar leg discs (Fig 6A). We used either anti-Dll antibody or anti-GFP antibody to ChIP an Sp1-GFP fusion protein expressed from an engineered ~80 kb BAC construct (see Materials and Methods) that drives Sp1-GFP expression identically to Sp1, and can rescue an Sp1 null mutant. Here we focus on genomic loci that show an intersection between 1) Sp1 and Dll binding events, 2) putative Dll, Sp1, Mad and Pan binding sites, and 3) have accessible chromatin as revealed by FAIRE-seq data for leg discs [37]. We found 442 genomic regions that satisfy all six criteria, many of which were close to genes that are expressed in leg discs (S4 Table). In addition, two regions correspond to vnE and DllM, another previously defined CRM of Dll (Fig 6C and 6G). As expected, rhoE was not identified because there was no consensus Sp1 binding site in rhoE. However, this approach identified a fragment that is within rhoLLE1 (rhoLLE1MIN) that, when tested in a reporter gene, drove expression in similar ring pattern as rhoLLE1 (Fig 6E).
To validate the larger set of predicted CRMs, we picked 20 additional genomic fragments (23 together with vnE, DllM and rhoLL1MIN, ~5% of the total 442 intersections) near 11 genes [Antennapedia (Antp), four-jointed (fj), spitz (spi), disconnected (disco), tarsal-less (tal), spineless (ss), Zn finger homeodomain 2 (zfh2), elbow B (elB), no-ocelli (noc), Enhancer of split m3 (E(spl)m3-HLH), and Distal-less (Dll)]. Using this approach, we discovered at least one leg disc enhancer element with a PD bias for each of the genes we tested (Figs 6 and S5; S5 Table), except for disco. In some cases (Antp, fj, spi, Dll, noc) multiple fragments generated leg disc expression patterns. Interestingly, we uncovered two leg disc enhancers for the EGFR ligand Spi. Overall, 18 of the 23 tested fragments (78%) are leg enhancers, suggesting that there is a battery of leg disc gene CRMs that drive expression differentially along the leg disc PD axis and are regulated by the direct input of Wg, Dpp, Dll and Sp1.
We also used genome-wide intersection criteria that excluded Sp1 as a factor, thus following the rhoE regulatory logic. Not surprisingly, this dataset was much larger (3809 loci), making it difficult to validate experimentally. Nevertheless, it also seems to predict enhancer loci because, in addition to rhoE, some of the identified regions corresponded to previously identified leg CRMs such as Dll DKO [38], Dll LL [39], and enhancer elements identified in genome-wide tiling studies [40].
The EGFR signaling pathway is widely used in animal development, and is frequently a target in human disease and developmental abnormalities [reviewed in 41]. Yet despite its importance in animal biology, many questions remain about how this pathway functions. Among these questions is whether secreted ligands that activate this pathway can induce distinct cell fates in a concentration-dependent manner. Here, we test this idea by specifically eliminating a single source of EGFR ligands from the center of the Drosophila leg imaginal disc, which fate maps to the distal-most region of the adult leg. One plausible scenario is that this single source of secreted EGFR ligands, which we refer to as the EOC, activates distinct gene expression responses at different distances from this source. Alternatively, eliminating ligands secreted from the EOC might only affect gene expression locally, close to or within the EOC. Taken together, our data are most consistent with the second scenario (Fig 7). This conclusion is largely supported by our observations that CRM deletions that eliminate vn and rho expression from the EOC have mild developmental consequences, both in the L3 leg imaginal discs and adult legs. These phenotypes are significantly weaker than those generated when the entire EGFR pathway is compromised using a temperature sensitive allele of the EGFR receptor. The difference between these two phenotypes is most likely explained by removing only a single source of EGFR ligands in the enhancer deletion experiments versus affecting EGFR signaling throughout the leg disc in the Egfrtsla experiments. This explanation is further supported by our observation that there are indeed additional CRMs, some of which we define here, that drive EGFR ligand production in more medial ring-like patterns during the L3 stage.
One possible caveat to these conclusions is that there are a total of seven rho-like protease genes in the Drosophila genome that could, in principle, play a role in distal leg development. We focused on rho and ru, based on previous results [11, 14] showing that triple rho ru vn clones generate severe leg truncations that phenocopy strong Egfrtsla truncations. In addition, we note that if other rho family proteases were active in the EOC, we would not expect to see leg truncations and patterning defects in the leg discs of the rhorhoE-Df vnvnE-Df double mutant, because those proteases should be able to produce active Spi. These observations suggest that the remaining five rho-like protease genes play a minor (or no) role in leg development. However, this conclusion will ultimately benefit from further genetic and expression analysis of these additional rho-like genes.
An additional previous observation that contrasts with the suggestion that EOC activity has only a limited role in specifying distal leg fates is the partial rescue of the PD axis when only a small number of distal leg cells were wild type in legs containing large rho ru vn clones [11]. However, we note that even in these ‘rescued’ legs, medial defects in PD patterning were apparent. It is also noteworthy that in these earlier experiments, only adult legs were examined. When we repeated the same experiment, but analyzed L3 discs, we found that rho ru vn clones generated phenotypes that were very similar to those produced by our double vnE rhoE enhancer deletions. Taken together, these observations suggest that timing must be considered in the interpretation of these experiments. When assayed at the late L3 stage, both our enhancer deletion and rho ru vn clone experiments argue that EOC activity is limited to specifying only the most distal fates, marked by the expression of al and C15. Starting in mid L3, and perhaps continuing into pupal development, there are additional sources of EGFR ligands [14] that, when compromised, can affect adult leg morphology. Nevertheless, at least at the L3 stage, these data suggest that EGFR ligands produced from the EOC have a limited and local role in specifying distal leg fates (Fig 7).
Integration of inputs from signaling pathways and organ selector genes at CRMs in order to execute distinct developmental programs is a recurrent theme during animal development (reviewed in [42]). Here, we identified two leg EGFR ligand CRMs that integrate the inputs from the Wg and Dpp signaling pathways and the leg selector genes Dll and/or Sp1 in a manner that is very similar to a previously characterized leg enhancer DllLT [9]. In addition, when we applied the same regulatory logic to the whole genome, we identified a battery of leg enhancer elements (Fig 6). Interestingly, each of these enhancers drives expression in a specific manner with slightly different timing despite the fact that many of the inputs are shared. It is conceivable that the different expression patterns directed by these enhancers are in part a consequence of additional inputs and/or the difference in the TF binding site grammar. In support of this idea, vnE and rhoE differ in the number of binding sites for many inputs and vnE requires Sp1 while rhoE does not. Both of these differences may contribute to the earlier onset of vnE expression compared to rhoE. The remaining enhancer elements identified in this study direct a plethora of PD-biased leg expression patterns–ranging from ubiquitous, to central and ‘ring’ patterns (Fig 6), which likely integrate inputs in addition to the ones described here. Future studies of these CRMs would help reveal the complex network of regulation that orchestrates leg development in the fruit fly. Such detailed understanding of the cis-regulatory architecture of fly leg development would likely give insights into organogenesis and evolution in other animals as well.
The EGFR signaling pathway has tremendous oncogenic potential and understanding the various mechanisms regulating its activation is not only interesting from the point of view of animal development but also has important practical implications. While the core components of the EGFR pathway have been thoroughly studied because of their potent tumorigenic capability in humans [reviewed in 43], little is known about the transcriptional regulation of EGFR ligands that bind the receptor and activate the pathway. The reiterative use of EGFR signaling in many developmental processes implies that different cis-regulatory elements are likely utilized by each EGFR ligand in different organs and tissues in order to correctly read the diverse cues in any specific developmental context. It is conceivable that genomic variation in EGFR pathway CRMs might lead to a predisposition to different types of EGFR-dependent tumors in humans, since such CRMs may respond to potent growth-promoting signaling pathways, such as Wnt and BMP.
In this study, we characterized in detail two Drosophila EGFR CRMs, vnE and rhoE, and showed how they integrate the cues from two transcription factors, Dll and Sp1, and two signaling pathways, Wg and Dpp, in order to execute a leg patterning developmental program. Analogous EGFR CRMs are likely to exist in mammals, especially because complex interactions between BMP, Wnt, Shh, multiple Dlx paralogs and other factors, are implicated in the induction of FGF signaling in mammalian limb development. Consistent with this idea, specific single nucleotide polymorphisms (SNPs) in humans in non-coding loci of genes encoding EGFR ligands have been shown to be associated with different types of cancer [44–46]. Such loci may be enhancer elements analogous to vnE and rhoE. We also note that the regulatory logic uncovered here is likely to be relevant to many CRMs and genes that share spatial and temporal expression programs. Exploiting this regulatory logic in other systems might streamline the identification of enhancer elements that will aid in the discovery of mechanisms that are relevant to EGFR-related human disease and developmental birth defects.
The following mutant alleles and enhancer trap alleles were used in this study: arr2, btdXA, cicQ474X, cicP[PZ]08482, dacp7d23 (dac-Gal4), DllSA1, Dllem212 (Dll-Gal4), Egfrtsla, Egfrf24, grkΔFRT, Krn27-7-B, Mad1-2, pntΔ88, rho7M43, ru1, ruinga, Sp1HR (shared ahead of publication, [47]), spiSC1, spiDf(2L)Exel8041, vnL6, vnGAL4. Transgenic alleles used for in vivo clonal ectopic expression of genes were: UAS-armΔN, UAS-tkvQD, UAS-Dll.
To perform RNAi knockdown of vein and spitz the following strains were used: UAS-vnRNAi (TRiP.HMC04390)/CyO, Dfd:EYFP; UAS-spiRNAi (TRiP.HMS01120) crossed to either Dll-GAL4 (Dllem212), spiDf(2L)Exel8041/ CyO, Dfd-EYFP; vnL6/TM6B, or spiDf(2L)Exel8041/CyO, Dfd-EYFP; vnGAL4/TM6B (vnGAL4 is a null allele [48]). Crosses were raised at 18°C, then shifted to >25°C at the start of L3. For assessment of larval phenotypes, crosses remained at 25°C until fixation and dissection as wandering larvae. For assessment of adult leg phenotypes, crosses were returned to 18°C after 24h until eclosion.
For generation of mutant clones that encompass the entire Dll-expressing leg disc region a yw; Dll-Gal4 (Dllem212), UAS-Flp; Ubi-GFP M- y+ FRT80B/C(2L;3R)Tb strain was crossed to a corresponding FRT80B-containg mutant strain (ruinga rhorhoE-Df vnL6 or ru1 rho7M43 vnvnE-Df). For Flp-FRT inducible mitotic recombination and subsequent mosaic clonal analysis fly larvae were heat-shocked at 48h post egg laying (PEL), 72h PEL or 90h PEL and dissected for staining as crawling stage larvae at around 120h PEL. For generation of Flp-FRT mitotic recombination clones, larvae were heat-shocked for 40 minutes at 37°C. Mitotic recombination clones were generated using the following strains: w hs-Flp1.22 Ubi-RFP FRT19A, yw hs-Flp1.22; Ubi-GFP FRT40A /CyO; E/TM6B, yw hs-Flp1.22; FRT42D Ubi-GFP/CyO; E/TM6B, yw hs-Flp1.22; E/CyO; Ubi-GFP FRT80B /TM6B, yw hs-Flp1.22; E/CyO; FRT82B Ubi-GFP/TM6B, yw hs-Flp1.22; FRT42D M- hs-GFP/CyO; E/TM6B, yw hs-Flp1.22; E/CyO; Ubi-GFP M- FRT80B/TM6B. The corresponding strains carrying mutant alleles were used in crosses for generation of mutant clones in the resulting progeny. E in these genotypes designates either vnE-lacZ or rhoE-lacZ inserted in landing sites 51D or 86Fa on chromosome II and III, respectively. To induce GFP expression in larvae marked with hs-GFP, an additional heat-shock was given 1 h before dissection for 20 min to 1 hour at 37°C.
The following strains were used for MARCM experiments where E designates either vnE-lacZ or rhoE-lacZ inserted in site 86Fa: yw hs-Flp1.22 tub-Gal4 UAS-GFP; tub-Gal80 FRT40A/CyO; E/TM2, yw hs-Flp1.22 tub-Gal4 UAS-GFP; FRT42D tub-Gal80/CyO; E/TM2, yw; Mad1-2 FRT40A; UAS-Dll/C(2L;3R)Tb, yw; FRT42D arr2; UAS-Dll/C(2L;3R)Tb, yw; FRT42D DllSA1; UAS-armΔN/C(2L;3R)Tb, yw; FRT42D DllSA1; UAS-tkvQD/ C(2L;3R)Tb, yw; y+ FRT40A; UAS-Dll/C(2L;3R)Tb, yw; FRT42D y+; UAS-Dll/C(2L;3R)Tb, yw; FRT42D y+; UAS-armΔN/C(2L;3R)Tb, yw; FRT42D y+; UAS-tkvQD/ C(2L;3R)Tb.
For all in vivo clonal experiments, at least 20 examples of discs with clones of the correct genotype were examined, which is typical for experiments of this type, and more than one independent experiment was carried out for each tested genotype.
All wildtype and mutagenized enhancer-reporter transgenic constructs were made using the lacZ reporter vector pRVV54 as an acceptor vector [49]. Coordinates of the genomic fragments PCR-amplified in the enhancer bashing experiments are listed in S1 and S5 Tables. The ΦC31 system was used for transgenesis and plasmids were introduced in landing sites 51D or 86Fa [50].
Site-directed mutagenesis of the vnE and rhoE enhancers was performed according to the QuikChange II protocol (Agilent Technologies). vnE and rhoE enhancers were first introduced in pBluescript SK+ vector for site-directed mutagenesis and the resulting mutated enhancers were consequently transferred to pRVV54 for in vivo analysis in the fruit fly. Primers used for mutagenizing of putative binding site are listed in S3 Table.
Plasmids for recombinant protein production were made by introducing cDNA sequences into pET21 series vectors (Novagen-EMD Millipore) and their derivatives, resulting in C-terminally tagged His proteins. Primers used to generate Dll-His (full-length Dll), Sp1Zn-finger-His (only the Zn-finger domains; used for confirming in vitro binding to Sp1 sites), Sp1424AA-His (used to examine cooperativity with Dll), MadMH1-His (only the MH1 domain) and PanHMG-His (only the HMG domain) vectors are listed in S2 Table.
The vnE and rhoE CRISPR/Cas9 alleles were generated by using pCFD4 vector for driving gRNA expression [18] and a germline-expressing Cas9 donor strain for plasmid mix injection [19]. The following sequences were used as gRNAs for generation of the vnEDf allele: CGATTTTAATGCGAAAGCTA and TTTGGCTTTCAACGCTTAAT. The following sequences were used as gRNAs for generation of the rhoEDf allele: GAGCCGAGGGCACAAATTGA and ATGATGATGATGTATTGCCC. We created a vector containing a cassette with P3-RFP [50] and FRT(F5)-hs-neo-FRT(F5) selectable markers flanked by minimal inverted ΦC31 [51] attP sites (pRVV613) [52]. This vector was used for insertion of upper and lower homologous arms for generation of donor vectors for creation of platforms for cassette-exchange. Primers used for PCR-amplification of the homologous arms are listed in S2 Table. vnE and rhoE pCFD4-based gRNA vectors (250ng/μl) were co-injected with the corresponding vnE and rhoE homologous arm donor cassette vectors (500ng/μl) and resulting flies were screened for P3-RFP expression. To generate rhoE deletion allele in the background of ruinga, injections to generate the rhoEDf were repeated in a nos-Cas9/CyO; ruinga/TM3 strain. Positive fly lines were verified by PCR for correct insertion of the donor cassettes. Deletion alleles without P3-RFP were generated through RMCE by injection with an empty multiple cloning site vector containing inverted ΦC31 attB sites (pRVV578) [52]. The P3-RFP-containing and -non-containing enhancer deletion alleles exhibited identical expression patterns and phenotypes. The WT vnE, rhoE and the D. virilis vnE enhancers were cloned into pRVV578 and resupplied by RMCE in a similar manner (primers are listed in S2 Table).
Recombinant proteins were expressed in BL21 (DE3) cells (Agilent Technologies) through IPTG induction for 4h. Proteins were subsequently purified through Cobalt chromatography with TALON Metal Affinity Resin (Clontech, #635501). EMSA gels were performed as previously described [53].
Immunostainings of fly imaginal discs was performed by standard protocol. The following antibodies were used in this study: rabbit anti-β-galactosidase (Cappel), mouse anti-β-galactosidase (Sigma-Aldrich, #G4644), guinea pig anti-Dll [9], rat anti-Sp1, guinea pig anti-Hth [54], mouse-anti-GFP (ThermoFisher Scientific, #A11121), rat anti-C15 [15], rat anti-Al [6], rat anti-BarH1 [55], rabbit anti-BarH1 [56], mouse anti-Dac [57]. AlexaFluor488-, AlexaFluor555-, and AlexaFluor647-conjugated secondary antibodies from ThermoFisher Scientific or Jackson ImmunoResearch Laboratories were used at 1⫶500 dilution.
Adult legs were dissected, mounted, and analyzed by light microscopy. All adults of the relevant genotype that eclosed within an 8-hour period were scored. Roman numerals in the figure legends indicate the tarsal segments present in each phenotypic class (with the distal most segment perturbed). For example, a truncation designated as I-III means that tarsal segments I, II and III were present, with segment III partially defective (e.g. Fig 2P). n refers to the number of individual legs scored. The number of legs examined for each genotype is reported in the figures and figure legends.
To generate vectors for in situ probes vn, ru, spi, Krn, and grk DNA sequences were amplified from genomic DNA and rho DNA sequence was amplified from cDNA clone (LD06131; DGRC clone #3528) using primers listed in S2 Table. DNA fragments were cloned into pBluescript SK+ (Agilent Technologies).
RNA antisense probes were transcribed with either T3 or T7 RNA polymerase (depending on the cDNA sequence orientation in the vectors listed in S2 Table) and labeled using DIG UTP mix (Sigma, #11175025910). Sense RNA probes were used as negative controls. rho probes were then hydrolized for 30 minutes at 60°C as previously described [58]. Third instar larvae were dissected in cold 1xPBS and fixed for 16h at 4°C in 4% PFA + 2mM EGTA. In situ hybridization was then performed as previously described [58] and signal was developed in BM-Purple AP substrate (Sigma #11442074001) after staining with anti-DIG -AP antibody at a concentration of 1:2000 (Roche #1093274). Multiple (≥10) discs were examined for each time point, probe, and genotype.
Mid-third instar larvae carrying wild-type or mutant vnE- or rhoE-lacZ reporter constructs were raised, fixed, stained and imaged in parallel according to standard immunohistochemical protocols. Average fluorescence was measured for the area within the central/tarsal domain of all unobstructed leg imaginal discs using ImageJ software (http://rsb.info.nih.gov/ij) and reported as the ratio of β-gal:Dll (staining control) in arbitrary units (AU). Ordinary one-way ANOVA adjusted for multiple comparisons (Dunnett's test) were performed and graphed in Prism software (graphpad.com) to compare wild-type fluorescence to mutant enhancer genotypes where ns = not significant, * = p≤ 0.0332, ** = p≤ 0.0021, *** = p≤ 0.0002 and **** = p<0.0001 (adjusted p-values). n refers to the number of individual leg discs scored. The number of leg discs scored for each genotype is reported in the figure legends.
Triplicate pools of 100 yw and 100 Sp1-GFPBAC L3 wandering larvae were used to perform independent chromatin IPs as previously described [59]. The Sp1-GFPBAC is a GFP-tagged Sp1 in BAC clone CH321-64M02 inserted in landing site VK00033 (gift from Dr. Rebecca Spokony). All 6 leg discs from each larva were used as material for each IP. Chromatin from the yw larvae pools was immuno-precipitated with goat anti-Dll antibody (sc-15858, Santa Cruz Biotechnology, 1.5 μg/ml for IP) while chromatin from the Sp1-GFPBAC larvae pools was immuno-precipitated with rabbit anti-GFP antibody (ab290, Abcam, 1∶300 dilution for IP). DNA from non-immunoprecipitated 10% chromatin input was isolated from each pool as reference control. Both control and immunoprecipitated DNA samples were prepared for Illumina sequencing using the Epicentre Nextera DNA Sample Preparation Kit and sequenced on an Illumina HiSeq 2000 according to the manufacturer's specifications. Experiments were performed in duplicate and peak calling was based on merged reads for duplicate ChIPs. Sequences were aligned to the Drosophila genome using the Burrows-Wheeler Aligner and ChIP-seq peaks were called using MACSv2 [60, 61]. Peak regions were defined using a p-value cutoff of 1.00e-02, but only those peaks passing a more stringent q-value cutoff of 1.00e-04 were used for further analysis. Datasets generated in this study are available at the Gene Expression Omnibus (GEO): accession number GSE113574.
PWMs for Dll, Sp1, Pan, and Mad were extracted from The Fly Factor Survey Database using the command grep within the MotifDb Bioconductor/R package. To generate BED files containing position information for each of the above PWMs, the matchPWM command from the Biostrings Bioconductor/R package was used. In-house code was used to run the command iteratively through the chromosomes (using DM3 build). Only hits above a minimum score of 80% were retained. IGVtools within the Integrative Genomics Viewer (IGV) was used to sort and index the BED files prior to intersection. Intersections of all BED files (derived from PWM analysis and ChIP-seq and FAIRE peak calling analysis) were done using Bedtools2 run locally from the command line. ChIP-seq peaks for Dll and Sp1 were first intersected with the FAIRE peaks. The product of this intersection was then sequentially intersected with each of the PWM files, always returning the peak coordinates from the initial file. The command intersectBed was used with options: -wa, -F 1.0, -u. To determine the gene nearest to each of the intersected ChIP peaks, packages within R/Bioconductor were used. The annotation package TxDb.Dmelanogaster.UCSC.dm3.ensGene was downloaded and annotated transcripts extracted. The distanceToNearest function was used to find the nearest annotated transcript to each of the ChIP Peaks. In-house R script was then used to generate the table containing the coordinates of the ChIP peaks, as well as the nearest annotated gene (S4 Table).
|
10.1371/journal.pcbi.1006549 | Simulations of blood as a suspension predicts a depth dependent hematocrit in the circulation throughout the cerebral cortex | Recent advances in modeling oxygen supply to cortical brain tissue have begun to elucidate the functional mechanisms of neurovascular coupling. While the principal mechanisms of blood flow regulation after neuronal firing are generally known, mechanistic hemodynamic simulations cannot yet pinpoint the exact spatial and temporal coordination between the network of arteries, arterioles, capillaries and veins for the entire brain. Because of the potential significance of blood flow and oxygen supply simulations for illuminating spatiotemporal regulation inside the cortical microanatomy, there is a need to create mathematical models of the entire cerebral circulation with realistic anatomical detail. Our hypothesis is that an anatomically accurate reconstruction of the cerebrocirculatory architecture will inform about possible regulatory mechanisms of the neurovascular interface. In this article, we introduce large-scale networks of the murine cerebral circulation spanning the Circle of Willis, main cerebral arteries connected to the pial network down to the microcirculation in the capillary bed. Several multiscale models were generated from state-of-the-art neuroimaging data. Using a vascular network construction algorithm, the entire circulation of the middle cerebral artery was synthesized. Blood flow simulations indicate a consistent trend of higher hematocrit in deeper cortical layers, while surface layers with shorter vascular path lengths seem to carry comparatively lower red blood cell (RBC) concentrations. Moreover, the variability of RBC flux decreases with cortical depth. These results support the notion that plasma skimming serves a self-regulating function for maintaining uniform oxygen perfusion to neurons irrespective of their location in the blood supply hierarchy. Our computations also demonstrate the practicality of simulating blood flow for large portions of the mouse brain with existing computer resources. The efficient simulation of blood flow throughout the entire middle cerebral artery (MCA) territory is a promising milestone towards the final aim of predicting blood flow patterns for the entire brain.
| The brain’s astonishing cognitive capacity depends on the coordination between neurons and the cerebral circulation, a system known as the neurovascular unit. The spatial and temporal coupling between these two networks is the object of intense research. However, the concise anatomical description of the cerebral circulation has so far been intractable. This paper introduces a methodology for the in silico creation of realistic models for the entire cerebral circulation. This innovation incorporates topological data from several neuroimaging modalities covering three lengths scales as input into a computer algorithm, which assembles anatomically accurate circulatory networks. When simulating blood flow as red blood cells suspended in plasma for experimental and synthetic cortical network models, we discovered that red blood cells tend to be more concentrated in deeper layers of the cortex compared to the surface. RBC fluxes are more homogenous in deeper layers. The phenomenon of depth dependent red blood cell supply supports the notion that the intricate architecture of the cortical microcirculation serves a self-regulating function to maintain uniform oxygen perfusion to neurons. We also demonstrate the practicality of predicting blood flow patterns for the entire brain with existing computer power.
| Metabolic activity of the brain is controlled by a complex system of neuroreceptors, small molecular regulators such as nitric oxide, hormones and proteins. The supply, clearance, and balance of metabolites, oxygen, glucose and waste are controlled by the cerebral circulation which is coupled with the cerebrospinal and interstitial fluid (CSF/ISF) subnetworks [1,2]. The coordination between oxygen extraction and increased cerebral blood flow after neuronal firing has garnered intense research interest in blood oxygen-level dependent (BOLD) signal, which is the basis of functional magnetic resonance imaging (fMRI). Recent work [3] has begun to quantify the microvascular origin of the BOLD fMRI signal in a microsection of a mouse brain. The study integrated state-of-the-art neuroimaging of anatomical spaces, tissue oxygen tension measurements and a mechanistic model of blood-bound oxygen supply to convert changes in cerebral blood flow and oxygen extraction into synthetic BOLD signals using Monte Carlo simulations. The main achievement was a successful first principles correlation between measured oxygen and cerebral blood flow (CBF) levels generating fMRI signals.
A recent paper from our group [4] aimed at widening the spatial coverage of coupled blood flow and oxygen simulations. Our model also offered detailed saturation and dissociation kinetics of plasma and red blood cell-bound oxygen, endothelial mass transfer and tissue oxygen extraction. Our study quantified vascular network effects by coupling biphasic (= suspension of red blood cells in plasma) hemodynamics and nonlinear blood rheology with oxygen kinetics. In addition, the number, distribution and position of neuronal and glial cell nuclei were acquired in a sizable section (~1x1x1 mm3) of vibrissa primary sensory cortex. We also predicted oxygen saturation in arterioles, capillaries and veins within experimental error bounds. By adopting a probabilistic approach to account for mitochondria respiration associated with specific neuronal and glial somata, the model was used to compute subcellular oxygen gradients between the extracellular matrix, the cytoplasm and individual neuronal/glial mitochondria. The remaining open question concerns the spatiotemporal coordination inside the neurovascular unit.
There is agreement that the neurovascular unit locally controls the cerebral blood flow response. Yet, oxygen supply exceeds the metabolic demand of neuronal activation for reasons that still remain uncertain [5]. Because of the massive size of the mammalian brain with its immense number of neurons and capillaries, the precise temporal and spatial coordination among cellular components still eludes exact physiological description. For example, studies suggest that functional hyperemia causes local neuronal metabolism increase of 5%, which in turn augments local blood flow by 30% up to almost 130% of base line perfusion [6]. However, the exact timing, regulation, and extent of dilation in individual spatially distributed vascular compartments during functional hyperemia are still being investigated [3,7–9].
The cerebral circulation also exercises a second blood flow control mechanism known as cerebral autoregulation [10–17]. Clinical observation [11] suggests that the total cerebral blood flow (CBF) remains constant over a wide range in perfusion pressure (±50 mmHg, ±6666 Pa). Many excellent contributions [18–20] correctly attribute the constancy of cerebral blood supply to global resistance adjustments. Yet, the involvement of specific vascular compartments, speed and spatial coverage of local vasodilatory/vasoconstrictive districts remains elusive [11,18–20]. Moreover, quantification of network effects and control principles among vascular compartments requires an anatomically accurate mathematical model of the cerebral circulation.
Propelled by the advances in neuroimaging, several groups have begun to integrate medical image data with large-scale computer models [3,4,9,21–23]. Generally, these efforts fall into two types. One type adopts a reductionist approach using simplified networks to highlight global blood flow distribution patterns [7,24–28]. The second type follows a bottom-up strategy which aims at replicating relevant microcirculatory components down to the level of the cellular ensemble. Noteworthy examples include quantifying the neurovascular coupling in functional hyperemia [3], analysis of pressure drop dependence on cortical depth [22], predictions of blood flow control by intra-cortical arterioles [9], and cortical oxygen distribution [29,30]. The ultimate goal of bottom-up models is a hemodynamic simulation of the entire brain, yet virtual circulation models of the whole brain have been perceived as intractable due to size and nonlinearity of the mathematical coupling between blood flow and oxygen kinetics [24].
This manuscript will introduce a computational procedure that integrates multimodal neuroimage data covering different length scales into a unified virtual representation of the murine cortical circulation. Two-photon imaging provides data for the reconstruction of capillary networks. High resolution micro computed tomography (μCT) imaging is used to capture the connectivity between main arterial branches and pial blood vessels. The morphometrics of the micro, meso- and macro-scale vascular models have been statistically analyzed in order to synthesize virtual blood flow networks with anatomically equivalent statistics, but without being confined to the limited field-of-view or resolution of imaging modalities.
The aim of this paper was to quantify network effects of uneven red blood cell distribution in the cerebral circulation. Although uneven red blood cell distribution also known as plasma skimming can be observed in single bifurcations, neuroimaging of the entire cerebral circulation has so far not been accomplished. To overcome this shortcoming, we integrated physiological data from several neuroimaging modalities covering three different lengths scales. Massive computer simulations of large microcirculatory networks of the murine primary cortex revealed a trend of depth-dependent hematocrit, which is a significant finding indicating that the intricate architecture of the cortical microcirculation serves a self-regulating function to maintain uniform oxygen perfusion.
We first assessed morphometrics of experimental data obtained from murine primary somatosensory cortex samples (N = 4, E1.1-E4.1). The indexing and naming scheme for the data sets is listed in Table 1. The total microvascular segment count was 24,669±9,594 splines. An important property is that all four original two-photon laser scanning microscopy (2PLSM) data sets contained blood vessels that divide into more than two daughter branches (multifurcations). Specifically, the four data sets contained 654, 725, 1686, 1440 multifurcations, respectively.
Statistics on cumulative metrics including vascular surface area, path length and luminal volume are compiled in Table 1. Although the data originate from the same cortical region, there are subject-specific variations between different specimens. There was higher variability in the low end of the vascular diameter spectra, because unavoidable uncertainty affects the thinnest vessels close to image resolution threshold as observed previously [31]. We also estimated the surface area to tissue volume ratio of the blood-brain-barrier (BBB) of the microvascular network as 8.8±1.1 mm2 vasculature/mm3 tissue. This number was obtained by summing the (endothelial) surface area of the capillary bed; this estimate compares to experimental values of the BBB surface of about 10-17 mm2/mm3 in humans [32–34].
A modified constructive growth algorithm (mCCO [30]) was used to create 60 synthetic data sets (S1.1-S4.15) of the murine primary sensory cortex. For each of the four experimental data sets, 15 clones with statistics matching closely their experimental original were created, so that the S1.1–15 series matched the original E1.1, and S4.1–15 matched data set E4.1. Artificial networks smoothly connect arterial vessels through the capillary bed to the veins without gaps or the need to insert artificial segments as observed with other methods [27]. In addition, since blood vessels are not exactly straight, realistic tortuosity values were imposed by a Bezier spline-based technique described previously [30]. Moreover, at the boundaries of the synthetic data sets neither pial surface vessels, nor deeper laying arterioles, capillaries, or venules were severed or had to be pruned thanks to the precise geometric control of the vasogenic growth algorithm. Artificial network growth took less than five minutes for each dataset on a personal computer.
We also compared morphometrics of experimental (N = 4) against synthetic vibrissa primary somatosensory cortex data sets (N = 60, S1.1-S4.15). No discernible feature differences can be inferred from visual inspection as shown for three experimental (E2.1, E3.1, E4.1) and six synthetic data sets (S2.5, S3.3, S4.5, S2.3, S3.4, S4.8) in Fig 1A. Total count amounted to 24,679 ± 8389 spline segments and 16,451 bifurcations. Spline segments were defined as tubular connections (splines) between branching points (bifurcations or multifurcations). This counting method ensured that the final tally is independent of image grid resolution or number of segment sub-partitions. The comparison of cumulative properties and probability density functions shows excellent agreement between the experimental and synthetic networks as seen in the plots Fig 1B and 1C. The synthetic networks are different realizations, but statistically equivalent replica (clones) of the original image samples.
The nonlinear biphasic blood flow, pressure and hematocrit equations for all four experimental networks and all sixty synthetic networks converged within five minutes [29]. Results were visualized with 3D rendering software Walk-In Brain developed at our institution [36,37]. Path analysis was conducted based on flow trajectories traversing the network along streamlines. Biphasic blood flow and network effects determining blood pressure and hematocrit distribution through large experimental (N = 4) and synthetic (N = 60) networks perfusing a large portion of the cortex were studied.
Typical pressure distributions along the microcirculatory network hierarchy are shown in Fig 2. Pressure drop trajectories through the microcirculation showed patterns consistent with experimental data [38–40]. Results of the path analysis in Fig 2 also depict the wide variations of hemodynamic states when blood traverses the dense microcirculatory network from the pial surface vessels through penetrating arterioles into the capillary bed and finally back to the collecting veins. The trajectories of individual paths (green, blue, magenta and yellow) display wide variability of hemodynamic states along the flow direction. Flow analysis reflected that a perfusion pressure drop in the microcirculatory networks from 120 to 5 mmHg (15,999-667 Pa) resulted in a mean tissue perfusion of 68.9 ml/100g/min (11 ∙ 10−6 m3/kg/s) which is within experimentally observed ranges [41,42].
We further inspected the RBC flux distribution as a function of network hierarchy (= vascular) and position inside the cortical hierarchy (= neuronal). The results were acquired for both empirical and synthetic data sets. Two representative specimens are highlighted in Fig 3A and 3B; eight more examples are displayed in Fig 3C. Typical paths belonging to different cortical layers are color coded in Fig 3. Flow paths were generated by tracing the flow from arterial inlet nodes downstream through the capillary bed until reaching a venous outlet. Paths were sorted according to their tissue supply function as follows: a path depth label equal to the cortical depth of the deepest segment was assigned to each flow path. Thus, all paths were uniquely ordered within a spectrum of shallow to deep reaching paths according to the neuronal layer (I-VI) hierarchy in agreement with previously reported values [35,43,44]. Fig 3 depicts hematocrit values along representative paths in shallow (layer I-green) and deeply penetrating paths (layer V/VI-yellow). Along each path and between different paths there is high variability along the flow direction. For example, discharge hematocrit in data set S1.1 reaches values as high as hmax~0.7, and as low as hmin~0.18. However, there is an overall trend of higher hematocrit being carried to lower cortical levels (layer-V/VI paths). The trend of relatively higher hematocrit, h, conveyed to deeper tissue layers (p-value<0.01, using one-way ANOVA test in MatLab) was observed consistently in all experimental and synthetic data sets. The bulk flow, Q, showed the opposite trend; it was reduced in segments of deeper layers which are connected by longer paths as is summarized in Fig 4. In contrast to bulk flow and hematocrit, the RBC flux (= volumetric flow rate of the RBC phase) exhibited weak layer dependency, it was almost constant irrespective of the cortical depth. We also observed that the variance of capillary RBC fluxes decreased with cortical depth, thus RBC fluxes in deeper layers show lower variability than paths on the surface. Taken together, biphasic blood rheology and network effects seem to induce depth dependent hematocrit supply to the cerebral cortex which leads to more homogenized RBC fluxes in deeper layers (= lower variance in RBC fluxes). Further analysis of diameter dependence on hematocrit confirmed the high degree of hematocrit variability across the diameter spectra as previously observed [4] (S1 Supplement).
The agreement between the simulation results obtained for experimental and synthetic data confirms that the synthetic networks are hemodynamically equivalent to the experimental networks. The satisfactory match in morphometrics and hemodynamics between experimental and synthetic data justifies the extension of network synthesis to large anatomical regions as described next.
Vascular networks covering the circulation of the entire MCA territory were generated with the help of our modified CCO (mCCO) algorithm as described in Gould et. al [30]. The mCCO algorithm was launched with the MCA M1 as the first segment. The location of the MCA territory within the context of the mouse cortex is shown in Fig 5 top-row. Sequentially, more segments were added at the cortical surface depicted in Fig 5 top-row, while minimizing the vascular tree volume subject to blood flow constraints. Thus, gradually the algorithm generated all arterial branches of the pial network on the cortical surface. Then, it was directed to proceed with penetrating arterioles and microcirculatory growth to a depth of approximately 1 mm below the pial surface, until a preset vessel density was reached. At each step of the segment generation, connectivity and bifurcation position were optimized to obtain minimum tree volume. The diameters of the network branches were recursively recomputed in accordance with hemodynamically-inspired principles [45]. The total number of splined segments in the artificial MCA territory was 993,185. This was roughly 60 times the number of segments in the cortical samples.
The topology of the synthetic MCA territory resembled maps available in mouse atlases [46,47]. Branching density and pattern of the pial arteries as well as the number of penetrating arterioles was within ranges of the reconstructed sets of μCT images as listed in Table 2. Detailed views in Fig 5 show pial, microcirculatory and individual capillary scales illustrating different aspects of the massive network model covering three length scales ranging from the MCA M1 segment with a diameter [48] of 142 µm down to the capillary bed [35], d<6 µm. Morphometrics of the synthetic MCA networks are summarized in Table 2. Fig 5A–5C depicts the pressure, flow and hematocrit field from the outflow of the Circle of Willis (MCA M1), down to the smallest capillaries in the microcirculation. The anatomical detail and branching pattern is depicted for the highly irregular, tortuous microcirculatory network.
The simulation of the entire MCA territory included the compartments of pial arteries, penetrating arterioles, pre-capillaries, capillaries, post-capillaries, ascending venules and pial veins. To complete the MCA circulation, the venous tree including venules was synthesized in reverse and connected to the capillary bed as described previously [30]. Fig 6 depicts the distribution of pressure, flow and hematocrit throughout the MCA territory. Fig 6A shows comprehensive three-dimensional maps of the anatomical hierarchy, pressure distribution, blood flow in the MCA territory, and uneven biphasic hematocrit. Fig 6B–6E highlights the anatomical grouping, pressure, flow, and hematocrit distribution throughout individual compartments. In these views, explosion diagrams separating the anatomical groups (pial arteries, penetrating arterioles, pre-capillaries, capillaries, post-capillary venules, venules and pial veins) were used to better delineate the hemodynamic states in each group. Visual inspection of the microcirculatory compartments (pre-capillaries, capillaries, and post-capillaries) depicted in Fig 6E reveal higher hematocrit levels in deeper cortical layers than on the surface.
Simulations conducted for the entire circulation on the MCA territory required boundary conditions at only two points; MCA M1 arterial blood pressure (p = 120 mmHg, 15,999 Pa), hematocrit level (h = 0.35), and venous outlet pressure (p = 5 mmHg, 667 Pa). The solution encompassed blood pressure, flow and hematocrit for 5452 pial vessels, 27,374 segments perpendicular to the pial surface, and 960,359 capillaries of the entire center MCA territory, for a total of 993,185. In total, the proposed iterative method succeeded in bringing to convergence a total of 2,648,853 equations for biphasic blood flow.
The predicted perfusion rate for the MCA territory was 50 ml/100g/min (= 8.3 ∙ 10−6 m3/kg/s) which is in agreement to literature ranges [41,42] of 40–163 ml/100g/min (= 6.7–27.2 ∙ 10−6 m3/kg/s). The trend of higher hematocrit levels in deeper cortical layers seen in the smaller cortical samples was also confirmed in the massive simulations for the MCA territory as shown in Fig 7.
It should be noted that the simulations showed virtually no boundary effects in the center of the MCA territory where the primary sensory cortical samples were located. The suppression of boundary effects that can be achieved by large-scale simulation is extremely important for simulating hemodynamic blood flow control such as it occurs in functional hyperemia or under autoregulatory control. A full simulation of the entire MCA territory (arterial and venous side) required 65 iterations and ~2 hours on multicore workstations.
We performed multiscale morphometric analysis of the cerebral circulation in mouse over three length scales. On both the macro and the mesoscale, statistical data for the Circle of Willis, the middle cerebral artery and its pial arterial network were extracted from high quality micro-CT (µCT) data [56]. Microcirculatory morphometrics were acquired by two-photon imaging (2PLSM) delineating the micro-angioarchitecture down to the level of individual capillaries for sizable sections (~1x1x1 mm3) of the vibrissa primary sensory cortex. There were statistical differences between the 2PLSM microcirculatory data sets especially in the diameter information as can be expected from a high resolution analysis of cortical microcirculatory networks. However, these variations did not significantly alter hemodynamic flow patterns. The morphometrics (arterial, capillary and venous segment number, connectivity and branching patterns, probability density functions for length, diameter and surface area spectra) informed a synthetic vascular growth algorithm. Because the statistics (e.g. segment numbers) could directly be input into the mCCO algorithm, we were able to create 15 synthetic replica for each of the four data sets. In total, we synthesized artificial vascular networks (N = 60) with morphometrics and blood perfusion patterns that are statistically equivalent to the experimental data. The wealth of experimental and synthetic data used in this study provided a testbed for hemodynamic analysis of biphasic blood flow through the cortical microcirculation.
Hemodynamic simulations were performed using computer algorithms described and tested extensively [29]. We performed biphasic blood flow simulations on both experimental (N = 4) and synthetic microcirculatory networks (N = 60). Simulation results predicted patterns of blood flow, pressure and hematocrit within ranges currently known from experiments. Even though our blood flow computations are deterministic [4,29], computed hemodynamics states varied widely within the labyrinth of paths traversing the microcirculation. We pinpointed randomness of the angioarchitecture as the origin of the wide range of predicted hemodynamic states. The finding of variability in hemodynamic states due to network architecture is significant, because it suggests that there are no characteristic properties (e.g. average hematocrit, mean capillary pressure) that would justifiably represent a typical physicochemical state of a microvascular compartment (arterioles, capillary bed, venules). It also explains why idealized trees such as binary ordered hierarchical graphs [26] are unsuitable surrogates for microcirculatory flow networks, because their regular and symmetric branching patterns lack the randomness in network topology seen in the murine anatomy. Specifically, ordered trees have equal states in all branches of a given hierarchy, which leads to even hematocrit splits due to symmetry in daughter branching diameters.
Variability in hemodynamic states reported previously [4] has implications for neuroimaging research. Specifically, even exact measurements at an individual point within the limited neuroimaging field of view (e.g. ~1 mm2 surface in two-photon images) would be prone to exhibit wide variations. The patchiness (variability) obtained by image acquisition at a single point cannot be overcome by more accurate imaging. Instead, an effective response to counteract variability due to network randomness is to adopt imaging protocols that emphasize spatially distributed samples over point measurements. In other words, measurements intended to infer global trends necessitate spatially distributed samples. Specifically, point observations acquired for single blood vessels can be expected to exhibit wide variations due to network effects, even if measurements are precise.
Our large-scale computer simulations suggest a depth dependent hematocrit gradient in the cortical blood supply as summarized conceptually in Fig 8. Detailed analysis of the spectrum of individual microcirculatory blood flow paths illuminated a clear trend; namely that deeply penetrating microvessels convey more red blood cells than paths running closer to the pial surface. The observation of higher hematocrit in deeper paths was observed in all simulation experiments for the primary sensory sets (experimental data sets, N = 4; synthetic microcirculatory networks, N = 60 as seen in Fig 4) as well as for the large-scale blood flow simulations covering the entire MCA territory shown in Fig 7. The predicted homogenization effect results in more uniform RBC fluxes, because shorter superficial paths tend to have higher bulk flow, Q, but carry less hematocrit, h. On the other hand, longer deeper penetrating paths have to overcome higher resistance leading to lower flows, but enjoy increased hematocrit as summarized in Fig 4 and Fig 7. As a consequence, this phenomenon also suggests that shorter surface paths which tap into fresh arterial oxygen supply have fewer RBCs, while deeper paths have higher concentrations of RBCs which on average carry lower O2 saturation. Another effect of hematocrit gradient is that net oxygen fluxes conveyed to different cortical layers are more evenly balanced than would be the case if RBCs distributed uniformly (no plasma skimming). We also noticed that the variance of RBC fluxes decreased with cortical depth. Accordingly, the distribution of RBC fluxes in deeper layers is more homogeneous than in surface layers. Random network architecture together with non-uniform hematocrit distribution due to the complex biphasic blood rheology seems to be two synergetic factors for ensuring homogenous oxygen supply irrespective of the cortical tissue depth. Since this homogenization effect needs no external feedback, it is plausible to infer that layer dependency of hematocrit and reduction of RBC flux variance serves a self-regulatory mechanism to balance oxygen supply to all cortical layers.
The plasma skimming effect describes a phenomenon seen in microvascular bifurcations (d<300 µm) [57,58] in which thinner side branches syphon disproportionately large amounts of plasma from the parent segment than thicker daughter branches. Our mechanistic simulations illustrate how plasma skimming phenomena apply over thousands of bifurcations and multifurcations in a tortuous vessel network, effectively overcoming the geometrical unavoidability of path length differences as shown in Fig 8.
Our recently developed kinetic plasma splitting model (KPSM) was our choice for computing large-scale network effects in this study. The main critical reasons include: (i) the KPSM split rule is able to handle multifurcations that occur in the murine microcirculatory anatomy (7.1%, 5.9%, 8.9%, 6.7% of all segments had multifurcations in experimental data sets), (ii) its predictions fall within physiologically meaningful property ranges. Specifically, it does not lead to predictions of zero or excessive hematocrit, and (iii) its linear and differentiable mathematical properties guarantee convergence of massive network computations. A full account documenting the KPSM model can be found in S3 Supplement.
The previously introduced network synthesis used a modified constrained constructive optimization (mCCO) [30] algorithm. The mCCO algorithm originally conceived by Schreiner [45] deploys two very simple principles: (i) minimization of vascular volume, and (ii) hemodynamic flow principle constraints which enforce that the total blood flow entering the network discharges in exactly equal amounts through the terminal outflow segments. Remarkably, this approach builds network structures whose topology resembles vascular network anatomy observed in vivo. One major task consisted of testing whether realistic network representations with arterial-capillary-venous closures could be synthesized with morphometric and hemodynamic properties matching networks acquired with neuroimaging modalities. The results showed that synthetic data (N = 60) created with a modified mCCO algorithm were statistically and hemodynamically equivalent to experimental cortical data sets (N = 4).
The hemodynamically inspired vascular growth procedure enabled the construction of realistic representations of the cortical blood supply of the entire MCA territory spanning multiple length scales from the large arteries (mm range) to the smallest capillaries (µm range), and draining through the pial veins (mm range) or three orders of magnitude in length scales. It allowed us to seamlessly integrate state-of-the-art topological data acquired from two entirely different imaging modalities (µCT and 2PLSM) into a single, coherent multiscale representation of the entire MCA territory with unprecedented anatomical detail that includes both the arterial and the venous side of the cerebrocirculation. Because simple, blood flow inspired construction principles are applied at all length scales, the resulting MCA circulation has no discontinuities or gaps between the main cerebral arteries, the pial arterial network, or the microcirculation. Morphometrics, anatomical details such as the shape of the cortical surface and hemodynamic principles, are incorporated at each stage of the growth algorithm. Thus, our proposed methodology may serve as an alternative to the practice of merely stitching together data from different locations or length scales.
The application of biphasic blood flow simulations for the entire MCA territory shows that large-scale blood flow and hematocrit simulations are feasible with existing computer resources. The large-scale simulations confirmed the trend of hematocrit layer dependence predicted for the smaller cortical samples. The massive simulations also elucidate the spatiotemporal coordination between different vascular compartments at different length scales (arteries vs arterioles vs capillary bed). The anatomical detail achieved with the MCA model may serve as a starting point for dynamic simulations that elucidate the involvement of different vascular components in regulating functional hyperemia, autoregulation or collateral blood supply in stroke. Because the network extended over a sizable portion of the mouse cortex, predictions for the center of the primary sensory cortex were free of boundary effects.
The synthetic MCA circulatory network also has the critical advantage that boundary conditions, which have been reported to hamper simulations on thin data sets [9], are applied very far away from the area of investigation. For example, Fig 5 displays typical subsections comparable in size to the 2PLSM data sets which are located far away from the MCA boundaries (MCA M1 segment and veins of the superior sagittal sinus). Thus, in samples situated at the center of the MCA territory, boundary conditions have negligible impact on hemodynamic predictions. The blood flow simulation for the entire MCA territory required only the arterial inlet pressure at the M1 segment and the blood pressure at the venous side.
We point out three additional reasons why the ability to synthesize morphologically and hemodynamically equivalent data sets is significant. (i) Artificial networks continuously connect the arterial side and the venous side without gaps. In 3D neuroimages assembled from two-dimensional image stacks, it is easy to miss segment connections or segments running between two slices. (ii) No segments are severed nor is there a need to prune dangling segments at domain bounds (this cleanup is unavoidable in image reconstructions [3,23]). In particular, fragmentation to pial arteries and many microcirculatory segments running perpendicular to the pial surface lead to boundary effects that can substantially affect predictions [9]. (iii) The most important benefit is the ability to expand the scope of data acquired by neuroimages without being confined to the bounded field-of-view or limited resolution of the imaging modality.
The ability to conduct brain-wide simulations would free the modeler from the burden of making uncertain assumptions at the boundaries of the artificial domain (edge of the image or simulation domain boundary). Because our algorithm succeeded in converging blood flow computations with hematocrit split for the entire MCA circulation in about two hours of CPU time, our group is confident that the proposed computational approach will enable blood flow simulations and oxygen transport on a brain-wide level in the near future.
Despite the evidence for trends such as depth dependent hematocrit, it should be emphasized that individual flow paths may experience substantially weaker or even reverted trends, as can be expected from the inherent randomness of the microcirculatory network architecture.
The 2PLSM technique provided a very detailed inventory of the cortical microcirculation. The four data sets did not include information about the subcortical blood supply to the white matter. White matter subcortical circulation is physiologically separated from the cortical blood supply. Accordingly, we assumed that the white matter supply is hydraulically separated from the cortical blood supply. However, certainty about this point would require a model of both the cortical and the subcortical networks (white matter blood supply). This task is intriguing, but is currently beyond the reach of 2PLSM, which is limited to ~1 mm depth. This is clearly a point for future research, but is currently outside the scope of this paper.
The main finding of depth dependency of hematocrit supply to the cortical layers is the result of a model prediction whose basis rests on experimental observations about plasma skimming and uneven hematocrit splits observed in capillaries outside the brain [59–61]. Therefore, the next logical step is to experimentally verify layer dependent hematocrit with deep imaging such as adaptive optics (AO) two-photon imaging [62]. If experiments confirm depth dependence and homogenization of RBC flux distribution, it would constitute a remarkable mathematical modeling contribution, which actually predicted, instead of merely explained, cortical blood supply. In the adverse case, the model would have prompted the need to revise our understanding of biphasic blood flow rheology as it relates to the cortical microcirculation (= diameter and hematocrit dependent viscosity laws, and hematocrit split rules), since so far it has been assumed that plasma skimming is active in capillaries throughout the entire circulatory system including the brain.
The conclusions about oxygen supply also need to be verified experimentally and computationally. The methods presented previously might be a first step in this direction [4]. However, oxygen predictions require discretization of the extracellular space which can be done in principle using the methods presented in Gould et. al [29], but is beyond the scope of this paper.
We predicted uneven depth dependent hematocrit distribution due to the complex biphasic blood rheology. Because our simulation did not include external factors such as gravity, we conclude that the result of depth dependent hematocrit arises from the combination of structural and hemodynamic properties of the network. Our findings suggest that network effects due to biphasic blood rheology and randomness of the network architecture are a controlling factor for ensuring adequate oxygen supply irrespective of the cortical depth. Since the observed homogenization of RBC variability requires no feedback, depth dependent hematocrit gradient may serve an important self-regulatory mechanism to balance oxygen supply to all cortical layers.
Uneven distribution of hemodynamic states in the microcirculation as well as the notion of layer-dependent hematocrit also have implications on the interpretation of the fMRI BOLD signal where it is usually assumed that hemodynamic states and hematocrit are homogeneous and evenly distributed throughout the microcirculation. The predictions in this work suggest that focal analysis of the fMRI BOLD signal would be more relevant than assuming global constants for the entire cortex.
We demonstrated that the modified constrained constructive optimization algorithm (mCCO) is successful in synthesizing artificial microcirculatory networks with topological and hemodynamic properties that are statistically equivalent to experimental data sets from different imaging modalities and length scales.
Simulations of the entire MCA circulation, which until recently would have to be considered intractable, are now becoming accessible to rigorous numerical analysis due to stable, efficient and physiologically consistent plasma skimming algorithms implemented on existing computer hardware. The synthesis of anatomically faithful cerebrocirculatory networks with desired topology closes the gap between large-scale blood flow simulations performed on image-derived data sets on one hand, and simulations on purely synthetic data sets on the other.
The successful synthesis of the entire MCA territory with biphasic blood flow simulation constitutes a step towards the ultimate goal of first principle simulations of cerebrocirculatory blood and oxygen distribution patterns for the entire brain.
An overview of the data structures used in this study is presented in Fig 9.
Nine female C57BL/6 mice were imaged for pial vascular network structures following intravascular injection of a lead pigment contrast agent as described elsewhere [56,63–65]. The mice were perfusion fixed prior to micro computed tomography (µCT) imaging with 7–20 µm isotropic resolution of the cerebral angioarchitecture. The resulting 3D images were filtered and the vascular lumen reconstructed as previously described [66–68]. Fig 9A shows raw µCT samples of the mouse vasculature from a 20 µm resolution image. The pial network statistics such as penetrating arteriole density and vessel diameter were compiled with results summarized in Table 2.
Four volumes (N = 4) that encompassed the murine vibrissa primary sensory cortex [35] were imaged using two-photon laser scanning microscopy (2PLSM) and are shown in Fig 9C. 2PLSM was employed to extract the spatial arrangement, length and orientation of blood vessels in the vibrissa primary sensory cortex [31,35,69]. Blood vessels in four data sets (~1x1x1 mm3) were labeled as pial arteries, penetrating arterioles, capillaries, ascending venules, or pial veins. Categorization was based on size and branching level according to Strahler order rather than physiological markers. No effort was made to differentiate pre-capillary arterioles from post-arteriole capillaries because it requires differential labeling of smooth muscle and pericytes. Capillaries were differentiated from ascending venules by a diameter cutoff of 6 µm and penetrating venules were differentiated from pial veins for vessels within a depth of 100 µm below the pia and a diameter less than 12 µm. Diameter information was also derived from images. The network information was stored using sparse connectivity matrices. Length, diameter, and tortuosity spectra are depicted in Fig 1. More details on image acquisition [31,35,69], image reconstruction [70], as well as the formulation of the network equations [29] can be found elsewhere.
Artificial microvascular networks (N = 60) for large sections of the cortex (~1x1x1 mm3) were synthesized using a previously described vascular growth algorithm [30]. Four examples are displayed in Fig 9D. The algorithm preserved dimensions of the experimentally acquired cortical samples, pattern and dimension of pial arteries, number, orientation and connectivity of penetrating arterioles, and morphometrics of the capillary bed, draining venules and pial veins, as listed in Table 1. Statistics and morphometric comparisons of experimental and synthetic data sets are displayed in Fig 1.
The arterial network of the entire MCA territory spanning three orders of magnitude in length from large arteries (~1 mm range) down to the entire capillary bed (~1 µm) was synthesized based on morphometric statistics of source data from multimodal images (µCT and 2PLSM).
Microcirculatory blood flow was modeled as a biphasic suspension comprised of red blood cells and plasma. Bulk blood flow was described by Poiseuille law relating volumetric flow to pressure drop as a function of resistance which in turn depends on diameter, d, and hematocrit-dependent viscosity [71]. In addition, a kinetic plasma skimming model (KPSM) presented previously [29] accounted for the uneven RBC distribution, known as plasma skimming.
This model has only one adjustable parameter, the skimming coefficient, m. It was set to value of m = 8 in all microcirculatory models, although this parameter could be refined as shown recently [72–74]. The nonlinear systems of conservation balances in system (1) were solved iteratively to calculate blood pressures, p, flow, Q, and hematocrit, h. Here, R is the resistance matrix, C1 and C2 are fundamental connectivity matrices [75] and C3 is the advection flux matrix. Boundary conditions are summarized in Table 3. More details on the mathematical background are given in S2 Supplement; implementation details are discussed elsewhere [29].
|
10.1371/journal.pgen.1007095 | A MIG-15/JNK-1 MAP kinase cascade opposes RPM-1 signaling in synapse formation and learning | The Pam/Highwire/RPM-1 (PHR) proteins are conserved intracellular signaling hubs that regulate synapse formation and axon termination. The C. elegans PHR protein, called RPM-1, acts as a ubiquitin ligase to inhibit the DLK-1 and MLK-1 MAP kinase pathways. We have identified several kinases that are likely to form a new MAP kinase pathway that suppresses synapse formation defects, but not axon termination defects, in the mechanosensory neurons of rpm-1 mutants. This pathway includes: MIG-15 (MAP4K), NSY-1 (MAP3K), JKK-1 (MAP2K) and JNK-1 (MAPK). Transgenic overexpression of kinases in the MIG-15/JNK-1 pathway is sufficient to impair synapse formation in wild-type animals. The MIG-15/JNK-1 pathway functions cell autonomously in the mechanosensory neurons, and these kinases localize to presynaptic terminals providing further evidence of a role in synapse development. Loss of MIG-15/JNK-1 signaling also suppresses defects in habituation to repeated mechanical stimuli in rpm-1 mutants, a behavioral deficit that is likely to arise from impaired glutamatergic synapse formation. Interestingly, habituation results are consistent with the MIG-15/JNK-1 pathway functioning as a parallel opposing pathway to RPM-1. These findings indicate the MIG-15/JNK-1 pathway can restrict both glutamatergic synapse formation and short-term learning.
| We explored the molecular mechanisms that govern synapse formation in vivo using the nematode C. elegans. Our results have identified a conserved MIG-15/JNK-1 MAPK pathway that restricts formation of glutamatergic, neuron-neuron synapses in the mechanosensory neurons, but does not restrict synapse formation in motor neurons. This could have important implications because synapses made by the mechanosensory neurons are reminiscent of synapses in the mammalian central nervous system, and relatively little is known about the signals that specifically influence central synapse formation in vivo. Interestingly, our results are consistent with the MIG-15/JNK-1 pathway opposing RPM-1, a signaling hub and ubiquitin ligase that inhibits different JNK and p38 signaling pathways. This suggests RPM-1 might inhibit specific MAPK pathways, such as the DLK-1 pathway, rather than simply acting as a general inhibitor of JNK and p38 signaling in neurons. Our results are particularly interesting given emerging links between JNK MAPK signaling and neurodegenerative diseases, such as Alzheimer’s disease.
| Information is relayed throughout the nervous system via chemical synapses, and the process of synapse formation is essential for construction of neural circuitry [1, 2]. In the central nervous system, most synaptic connections are glutamatergic and often referred to as central synapses.
Different neurons in the nematode C. elegans have proven extremely valuable in identifying conserved regulators of synapse formation, including motor neurons, HSN neurons, and mechanosensory neurons [3, 4]. Of these, only the mechanosensory neurons form glutamatergic, neuron-neuron synapses that are reminiscent of central synapses. Genetic screens using mechanosensory neurons have revealed several molecules that regulate glutamatergic synapse formation including the intracellular signaling hub Regulator of Presynaptic Morphology 1 (RPM-1) [5], the F-box protein MEC-15 [6], the transcription factor SAM-10 [7], the focal adhesion protein ZYX-1/Zyxin [8], and the microRNA LIN-4 [9].
C. elegans has two PLM mechanosensory neurons, which sense posterior gentle touch [10]. The PLM neurons form electrical synapses and glutamatergic chemical synapses with postsynaptic interneurons [10–12]. Initial touch sensation is thought to rely upon both electrical synapses and glutamatergic transmission. Glutamatergic transmission is also required for more complex touch-response behaviors, such as habituation to repeated tap and arousal from a sleep-like state called lethargus [13, 14].
RPM-1 is the C. elegans ortholog of mouse Phr1 and Drosophila Highwire, which are collectively referred to as Pam/Highwire/RPM-1 (PHR) proteins. RPM-1 is a relatively broad regulator of synapse formation and axon termination affecting these processes in mechanosensory neurons and motor neurons [5, 15–17]. Studies in flies and mice have shown these are conserved RPM-1 functions [18–22]. Previous work showed RPM-1 functions in the mechanosensory neurons during development to regulate habituation to repeated tap stimulus, a behavioral deficit that likely results from abnormal glutamatergic synapse formation in rpm-1 mutants [23].
PHR proteins are enormous intracellular signaling hubs that regulate numerous downstream pathways [15]. Proteomic screens identified several RPM-1 binding proteins that mediate RPM-1 function during development including: the RCC1-like protein GLO-4 [24], the microtubule binding protein RAE-1 [25, 26], the PP2C phosphatase PPM-2 [27], and the Nesprin ANC-1 [28]. Suppressor genetics revealed that RPM-1 inhibits p38 and JNK MAPK signaling [29–31]. Studies from flies and mammals have shown this is a conserved PHR protein function [32–34]. RPM-1 inhibits MAPK signaling by ubiquitinating MAP3Ks, such as DLK-1 and MLK-1, thereby targeting them for degradation by the proteasome [29, 35, 36]. At present, it remains unclear whether RPM-1 is a general inhibitor of MAP3Ks during neuronal development, or a potentially selective regulator of certain p38 and JNK pathways.
JNK MAP kinase pathways play a conserved role in synapse formation in the developing nervous system [37]. Studies in worms [29, 30], flies [32, 38] and vertebrates [39, 40] have shown JNK signaling needs to be restricted for proper synapse formation. The role of JNK signaling in synapse formation has been primarily explored using the neuromuscular junction (NMJ) in different organisms. Much less is known about how JNK signaling impacts the formation of neuron-neuron, central synapses in vivo. It also remains unclear if JNK isoforms have differential roles in mediating synapse formation in different types of neurons. The importance of addressing these gaps in our knowledge is highlighted by emerging links between altered JNK signaling and synaptic dysfunction in Alzheimer’s disease [41–44].
JNK activation is regulated by upstream kinase cascades that include MAP2Ks, MAP3Ks, and in some cases MAP4Ks. In vitro biochemistry and genetic studies on fly embryogenesis have shown Misshapen (Msn), called MIG-15 in worms and NIK (HGK/MAP4K4) in mammals, can function as a MAP4K that activates JNK signaling [45–47]. While MIG-15 and Msn regulate axon guidance in the nervous system, they do so through the cytoskeletal regulators ERM-1 and Bifocal [48, 49]. To date, functional links in the nervous system between JNK signaling and MIG-15 or Msn have remained stubbornly elusive.
Using suppressor genetics, we identified several kinases that are likely to be part of a new JNK pathway that regulates glutamatergic, neuron-neuron synapse formation in the mechanosensory neurons of C. elegans. This JNK pathway is composed of MIG-15 (MAP4K), NSY-1 (MAP3K), JKK-1 (MAP2K), and JNK-1. Loss of function mutations in kinases of the MIG-15/JNK-1 pathway specifically suppress defects in synapse formation in the mechanosensory neurons of rpm-1 mutants. Consistent with suppression results, transgenic overexpression of kinases in the MIG-15/JNK-1 pathway in wild-type animals results in impaired synapse formation. Further biochemical support for the MIG-15/JNK-1 pathway comes from our observation that NSY-1 can bind both MIG-15 and JKK-1 when expressed in HEK 293 cells.
Based on prior work, the most likely signaling model that might explain our results is RPM-1 acting as a ubiquitin ligase to inhibit an upstream kinase in the MIG-15/JNK-1 pathway, similar to RPM-1 effects on the DLK-1 and MLK-1 pathways. However, our behavioral analysis of habituation indicated the MIG-15/JNK-1 pathway is more likely to function as a parallel opposing pathway to RPM-1. These results provide the first potential link in the nervous system between the MIG-15 MAP4K and a JNK isoform, and indicate MIG-15/JNK-1 activity can affect synapse formation and short-term learning.
Previous studies showed that loss-of-function mutations in kinases of the DLK-1/PMK-3 p38 MAPK pathway and the MLK-1/KGB-1 JNK pathway suppress synapse formation defects in rpm-1 mutants [29, 30]. A different pathway consisting of JNK-1 and an MKK7 ortholog JKK-1 functions in the nervous system to regulate locomotion, and synaptic vesicle trafficking in GABAergic motor neurons [50, 51]. jkk-1 and jnk-1 also suppress synaptic position defects in cholinergic NMJs of arl-8 mutants [52]. These observations prompted us to test whether jkk-1 and jnk-1 could suppress defects in synapse formation and axon termination caused by rpm-1 loss of function (lf) in the mechanosensory neurons.
C. elegans has two PLM mechanosensory neurons, one on each side of its body. Each PLM neuron extends a single axon that terminates growth prior to the cell body of the ALM mechanosensory neuron (Fig 1A and 1B). Glutamatergic chemical synapses are formed en passant with interneurons by a collateral branch that extends from the primary axon (Fig 1A and 1D). The transgene muIs32, which expresses GFP in the mechanosensory neurons, can be used to assess PLM neurons for axon termination, as well as formation of the synaptic branch and synaptic boutons [24]. Anatomical separation of axon termination and chemical synapse formation makes the PLM neurons an ideal system to assess these two developmental processes within the same neuron.
Consistent with prior studies [5, 53], we observed that rpm-1 mutants display two phenotypes in the PLM neurons. The first is failed axon termination, which results in the PLM axon extending beyond the ALM cell body and hooking towards the ventral side of the animal, which we refer to as a hook defect (Fig 1B). The second phenotype is retraction of the synaptic branch, which results from impaired synapse formation (Fig 1D) [5]. Quantitation of axon termination defects, and synaptic branch defects in rpm-1 mutants are shown in Fig 1C and 1E, respectively.
To test if jkk-1 and jnk-1 affect axon termination and synapse formation of PLM neurons, we generated double mutants with previously described null alleles for jkk-1 (km2) and jnk-1 (gk7) [50, 54]. rpm-1; jkk-1 and rpm-1; jnk-1 double mutants did not show changes in the frequency of axon termination defects (Fig 1B and 1C). In contrast, synaptic branch defects were significantly suppressed in both double mutants (Fig 1D and 1E).
Given that both jkk-1 and jnk-1 suppress synaptic branch defects, but not axon termination defects, we tested whether they function in the same genetic pathway. To do so, we generated rpm-1; jkk-1; jnk-1 triple mutants. We observed no further suppression of synaptic branch defects in triple mutants (Fig 1E), and again saw no suppression of axon termination defects (Fig 1C). These results are consistent with jkk-1 and jnk-1 functioning in the same genetic pathway to specifically suppress synapse formation defects in the PLM neurons of rpm-1 mutants.
Next, we explored MAP3Ks and MAP4Ks that might exist in the JKK-1/JNK-1 pathway. We began by testing the MAP3K NSY-1 because it regulates olfactory neuron fate during development [55], and its mammalian ortholog ASK activates JNK signaling in vitro [56]. To analyze genetic interactions between rpm-1 and nsy-1, we used a previously described nsy-1 allele, ok593, that deletes the kinase domain and is likely to be a null. Similar to outcomes with jkk-1 and jnk-1, we observed suppression of synaptic branch defects in rpm-1; nsy-1 double mutants, but did not see changes in axon termination defects (Fig 2A and 2D). Quantitation showed that suppression of synaptic branch defects in rpm-1; nsy-1 double mutants was significant compared to rpm-1 single mutants (Fig 2E). No significant decrease in axon termination defects occurred in rpm-1; nsy-1 double mutants (Fig 2B). Importantly, no increased suppression occurred in rpm-1; jnk-1; nsy-1 triple mutants compared to either double mutant (Fig 2E). These results are consistent with nsy-1 functioning in the same pathway as jnk-1 to suppress synapse formation defects.
Previous studies on MAP kinases regulating neuronal development in C. elegans have often identified three component pathways (MAP3K, MAP2K and MAPK) [29–31]. At present, potential MAP4Ks that regulate JNK signaling in neuronal development remain unknown. MIG-15 is orthologous to mammalian NIK and fly Msn, which can act as MAP4Ks in cultured cells and during embryogenesis [46, 47]. In the developing C. elegans nervous system, MIG-15 regulates cell migration and axon guidance, but functional genetic links to JNK signaling in neurons remain absent [57, 58]. Therefore, we tested if loss of MIG-15 function specifically suppresses synapse formation defects in rpm-1 mutants similar to nsy-1, jkk-1 and jnk-1. We used a strong loss of function allele, mu342, that is a point mutation in the MIG-15 kinase domain and likely to disrupt kinase activity [58]. Axon termination defects were not suppressed in rpm-1; mig-15 double mutants (Fig 2A and 2C). However, synaptic branch defects were significantly suppressed in rpm-1; mig-15 double mutants (Fig 2D and 2F). It should be noted that in some mig-15 mutants, the PVM cell body did not migrate normally, which is consistent with previously described cell migration defects (Fig 2D). Similar to results with nsy-1, further suppression of branch defects was not observed in rpm-1; jnk-1; mig-15 triple mutants compared to rpm-1; mig-15 double mutants. This result is consistent with mig-15 functioning in the same genetic pathway as jnk-1.
The logic of how MAPK modules operate suggests the most likely order for this putative kinase pathway from most upstream to downstream kinase would be: MIG-15/NSY-1/JKK-1/JNK-1. However, our genetic results do not rule out the alternative possibility that these kinases are part of multiple MAPK pathways that function in parallel to suppress synapse formation defects, but not axon termination defects, caused by rpm-1 (lf).
In addition, we observed a low but significant level of synaptic branch defects in mig-15 single mutants (Fig 2F). This observation, and suppression of synaptic branch defects caused by rpm-1 (lf) suggest that a balance of MIG-15 activity is required for proper synapse formation in the PLM neurons.
Previous studies suggested impaired synapse formation in rpm-1 mutants leads to retraction of the PLM synaptic branch [5]. Therefore, the absence of the synaptic branch served as a proxy for assessing synapse formation in different double mutants of rpm-1 and kinases of interest. However, there are two possible cellular explanations for why synaptic branch defects are suppressed. The first is that synapse formation is improved. Alternatively, MIG-15, NSY-1, JKK-1 and JNK-1 kinases might be required for synaptic branch retraction.
To differentiate between these two possibilities, we simultaneously labeled PLM neurons with two transgenes: 1) jsIs1114 which expresses the synaptic vesicle marker GFP::RAB-3, and 2) jsIs973 which expresses mRFP and allows the morphology of the axon, synaptic branch and synaptic boutons to be visualized. RAB-3 is a small G-protein that associates with synaptic vesicles (SV), and has been used to assess synaptogenesis in PLM neurons [8]. Consistent with prior work, we observed GFP::RAB-3 at the presynaptic boutons of wild-type animals (Fig 3A). In confocal projections, it is possible to see synaptic boutons from both the left and right PLM neurons (PLML and PLMR) as they reside in the same z-plane in the ventral nerve cord. However, the morphology of only a single PLM axon can be visualized (Fig 3A). We observed two phenotypic groups of rpm-1 mutants. The first, and most common, lacked one of the two PLM synaptic branches. The example shown in Fig 3A (rpm-1 left absent) is missing the left PLM synaptic branch and accompanying GFP::RAB-3 accumulations, but GFP::RAB-3 from PLMR is still present. Less frequently, we observed rpm-1 mutants in which both PLM neurons lack a synaptic branch and show no presynaptic GFP::RAB-3 accumulation (Fig 3A, rpm-1 both absent). Quantitation showed accumulation of RAB-3 always corresponded with the presence of a synaptic branch in wild-type animals (Fig 3C). The frequency of complete synaptic branches, defined as branches with presynaptic boutons and GFP::RAB-3 accumulation, were reduced in rpm-1 mutants (Fig 3C). In rpm-1; mig-15 and rpm-1; jkk-1 double mutants, we observed an increase in the number of complete synaptic branches and corresponding accumulation of GFP::RAB-3 at presynaptic terminals (Fig 3B and 3C). Thus, mig-15 and jkk-1 suppression of rpm-1 results in improved synapse formation.
Loss of function in NSY-1, JKK-1 and JNK-1 had minimal effects on PLM synapse formation, as assessed by the presence of the synaptic branch and presynaptic boutons (Figs 1E and 2E). Nonetheless, it was possible more subtle changes might exist in morphology or number of presynaptic terminals when these kinases are perturbed.
To test this, we utilized transgenic worms that express either the synaptic vesicle marker GFP::RAB-3, or the active zone marker UNC-10::tdTOMATO in the mechanosensory neurons (Fig 4A and 4B). Transgenes with an appropriate cell fill marker (GFP or mRFP) were evaluated simultaneously to anatomically identify PLM presynaptic terminals. Both presynaptic RAB-3 and UNC-10 puncta size and number showed no significant differences when wild-type animals were compared to jnk-1, jkk-1 or nsy-1 mutants (Fig 4C and 4D). Data for nsy-1 and GFP::RAB-3 is not shown because recombinants between nsy-1 and the RAB-3 transgene could not be obtained.
To further evaluate whether these kinases might affect synapse formation outside of rpm-1 suppression, we turned to the microtubule destabilizing drug colchicine. It was previously shown that treating wild-type worms with colchicine causes synapse formation defects in PLM neurons [59]. This provided us with a pharmacological manipulation to further examine the role of the MIG-15/JNK-1 pathway in synapse formation. While no effect was observed with the vehicle DMSO, we found that low frequency synaptic branch defects caused by colchicine were enhanced in jnk-1, jkk-1 or nsy-1 mutants (Fig 4E and 4F). These observations indicate altering microtubule stability can also unveil a role for NSY-1, JKK-1 and JNK-1 in glutamatergic synapse formation.
Our genetic analysis showed mig-15, nsy-1, jkk-1 and jnk-1 are likely to function in a linear genetic pathway to suppress synapse formation defects in rpm-1 mutants. Next, we wanted to determine if these kinases regulate synapse formation by functioning cell autonomously within PLM neurons, or non-cell autonomously in surrounding tissue. We addressed this with transgenic rescue experiments using double mutants of rpm-1 and different kinases. Transgenic expression of kinases was driven by the mec-7 promoter, which is expressed in mechanosensory neurons including the PLM neurons, and is not expressed in the postsynaptic interneurons or tissues that surround the PLM synapse, such as muscles or hypodermis. NSY-1, JKK-1 and JNK-1 were expressed using transgenic extrachromasomal arrays. MIG-15 was expressed using a MosSCI single-copy integrated transgene because mig-15 mutants have small body and brood size, which made deriving extrachromosomal arrays extremely difficult. Expression of each kinase in PLM neurons gave significant, robust rescue of suppression in kinase double mutants with rpm-1 (Fig 5A and 5B). For example, transgenic expression of JNK-1 (but not the negative control protein mCherry) in rpm-1; jnk-1 double mutants significantly rescued suppression of synaptic branch defects (Fig 5A and 5B). Our findings support the conclusion that MIG-15, NSY-1, JKK-1 and JNK-1 function cell autonomously in PLM mechanosensory neurons, and are consistent with these kinases functioning in a linear cascade to regulate synapse formation.
Our observation that loss of MIG-15, NSY-1, JKK-1 or JNK-1 improves synapse formation defects in rpm-1 mutants suggests these kinases might inhibit synapse formation. To test this hypothesis, we transgenically overexpressed individual kinases in wild-type animals using the pan-neuronal rgef-1 promoter, which has been used previously for transgenic overexpression experiments with PLM neurons [30, 53]. PLM synaptic branches were not altered in wild-type animals carrying transgenic extrachromosomal arrays that overexpress mCherry, which was used as a negative control (Fig 6A and 6B). In contrast, transgenic overexpression of MIG-15, NSY-1 or JKK-1 in wild-type worms resulted in synaptic branch defects (Fig 6A and 6B). Transgenic overexpression of JNK-1 did not lead to defects (Fig 6B). This result with JNK-1, which is likely at the bottom of a possible MIG-15/JNK-1 pathway, is not surprising given the prior finding that kinases such as the p38 MAPK PMK-3 and the JNK isoform KGB-1 generate relatively low frequency defects when transgenically overexpressed [30]. Nonetheless, our results show that overexpressing MIG-15, NSY-1 or JKK-1 is sufficient to impair synapse formation. This indicates that these kinases need to be opposed functionally or restricted for proper PLM synapse formation to occur. Furthermore, synapse formation defects caused by overexpression of MIG-15, NSY-1 and JKK-1 occurred at similar levels, which provides further evidence these kinases could function in the same MAPK pathway.
Our genetic experiments suggested that MIG-15, NSY-1, JKK-1 and JNK-1 potentially function in a novel MAPK pathway that regulates synapse formation in PLM neurons. We wanted to further evaluate the biochemistry that underpins these findings.
Previous in vitro experiments showed MAP3Ks often bind to both MAP4Ks and MAP2Ks in the same pathway [60]. For example, MEKK1 (MAP3K1) binds to both the MAP4K NIK and the MAP2K MKK4 [46, 61], while NSY-1 binds to the MAP2K SEK-1 [62]. To test whether similar interactions occur between the NSY-1 MAP3K and the MIG-15 MAP4K or JKK-1 MAP2K, HEK 293 cells were transfected with FLAG or HA tagged kinases and binding was assessed by coimmunoprecipitation. When NSY-1 was immunoprecipitated from transfected cell lysates, we detected coprecipitating MIG-15 (Fig 7A) and JKK-1 (Fig 7B). NSY-1 binding to both MIG-15 and JKK-1 provides biochemical evidence that these kinases could function in a linear MAPK pathway.
Because MIG-15, NSY-1, JKK-1 and JNK-1 affect synapse formation, it is plausible these kinases might localize to presynaptic terminals. To explore this, we generated transgenic extrachromosomal arrays that simultaneously expressed mCherry and GFP::JNK-1. As shown in Fig 8A, GFP::JNK-1 localized to the presynaptic boutons of PLM neurons. JNK-1 was also observed in the axon and synaptic branch. GFP::NSY-1 similarly localized to presynaptic boutons (Fig 8B). We tried to determine if MIG-15 and JKK-1 localize to presynaptic terminals, but these constructs were not readily detected in PLM neurons.
To further define presynaptic localization, we simultaneously expressed GFP::JNK-1 and the active zone marker UNC-10::tdTOMATO. Confocal microscopy showed colocalization between GFP::JNK-1 and UNC-10::tdTOMATO at presynaptic terminals (Fig 8C). This demonstrates that JNK-1 localizes to the presynaptic active zone of PLM mechanosensory neurons. Collectively, these results are consistent with NSY-1 and JNK-1 regulating glutamatergic synapse formation.
RPM-1 and its orthologs, Drosophila Highwire and mouse Phr1, regulate synaptogenesis at NMJs [16, 18, 19]. Loss of function in the DLK-1/PMK-3 or MLK-1/KGB-1 kinase pathway suppresses synapse formation defects in the GABAergic motor neurons of rpm-1 mutants [29, 30]. Therefore, we tested whether NSY-1 and JKK-1 affect defects in GABAergic NMJs caused by rpm-1 (lf).
The GABAergic DD motor neurons of C. elegans form synapses with dorsal body wall muscles (Fig 9A). To visualize GABAergic NMJs we used juIs1, a transgene that expresses the synaptic vesicle marker Synaptobrevin-1 fused to GFP (SNB-1::GFP). We observed regularly spaced presynaptic puncta in wild-type animals labeled with SNB-1 (Fig 9A). Consistent with previous studies, rpm-1 mutants had synapse formation defects in which large sections of the dorsal cord lacked SNB-1 puncta, and SNB-1 puncta aggregated (Fig 9A). Interestingly, synapse formation defects were unchanged in rpm-1; jkk-1 and rpm-1; nsy-1 double mutants (Fig 9A). Quantitation of SNB-1 puncta confirmed synapse formation defects in GABAergic motor neurons of rpm-1; jkk-1 or rpm-1; nsy-1 double mutants were not suppressed compared to rpm-1 single mutants (Fig 9B). We observed a small, significant defect in NMJ formation in jkk-1 single mutants (Fig 9B). This suggests JKK-1 might regulate GABAergic synapse formation. Overall, these results demonstrate that NSY-1 and JKK-1 do not affect RPM-1 regulation of synapse formation in GABAergic motor neurons the way they do in PLM mechanosensory neurons. This indicates that loss of NSY-1 and JKK-1 specifically suppresses synapse formation defects at glutamatergic, neuron-neuron connections.
The mechanosensory neurons mediate gentle touch responses, including the response to non-localized mechanical stimuli such as vibrations caused by tapping the plate on which worms grow. The mechanosensory neurons, including the PLM neurons, propagate sensory activity primarily via electrical synapses at sites along the primary axon (Fig 10A) [10, 63]. Glutamatergic chemical synapses are formed in an anatomically distinct location on the collateral synaptic branch, and contribute to a lesser extent to touch sensation (Fig 10A) [10, 11]. Glutamatergic transmission has been implicated in short-term learning, specifically habituation to repeated tap stimuli [13, 23].
Previous work showed rpm-1 mutants respond normally to touch [5, 23], which is consistent with these animals having normal electrical synapses in their mechanosensory neurons [64] (Borgen and Grill, in press). In contrast, rpm-1 mutants have dramatic defects in habitation to repeated tap, which result from loss of RPM-1 function in the mechanosensory neurons [23]. This suggests that habituation defects in rpm-1 mutants are likely to result from defective chemical synapse formation. Thus, tap habituation is a reasonable behavioral readout to further test the genetic relationship between rpm-1 and kinases in the MIG-15/JNK-1 pathway. Tap habituation also provided us with a whole-animal, physiological assay that is sensitive to impacts on chemical synapse formation in the mechanosensory neurons.
As shown in Fig 10B, wild-type animals displayed habituation to repeated tap, in which responses diminish with increasing tap stimuli over time. Consistent with prior findings [23], rpm-1 mutants were defective in habituation, and maintained a high probability of response over repeated tap stimuli (Fig 10B). Tap habituation defects were suppressed in double mutants of rpm-1 with jnk-1, jkk-1 or nsy-1 (Fig 10B, 10C and 10D). Defects in body morphology and locomotion prevented us from testing mig-15 mutants. These behavioral findings show that loss of NSY-1, JKK-1 and JNK-1 suppress rpm-1 in the context of tap habituation. This finding correlates nicely with our findings on chemical synapse formation in PLM neurons. It also provides further behavioral evidence that is consistent with NSY-1, JKK-1 and JNK-1 functioning in the same pathway.
Our observation that impairing kinases in the MIG-15/JNK-1 pathway suppresses rpm-1 (lf) could be explained by two different signaling models. First, RPM-1 might function as a ubiquitin ligase to inhibit the MIG-15/JNK-1 pathway (Fig 10E, Model 1), similar to what occurs with the DLK-1/PMK-3 and MLK-1/KGB-1 pathways [29, 30]. Alternatively, RPM-1 could function in a parallel opposing pathway to MIG-15/JNK-1 (Fig 10E, Model 2). Our unexpected observation that jnk-1 and jkk-1 single mutants have more rapid habituation, the opposing phenotype to rpm-1, provided us with an ideal phenotypic relationship to test genetic epistasis (Fig 10B and 10C). rpm-1; jnk-1 and rpm-1; jkk-1 double mutants had an intermediate habituation phenotype similar to wild-type animals (Fig 10B and 10C). This suggests it is likely the MIG-15/JNK-1 pathway functions in a parallel opposing pathway to RPM-1 (Fig 10E, Model 2). It is not clear to us why nsy-1 mutants did not have increased habituation. One possibility is another, unknown MAP3K functions redundantly with NSY-1 to regulate JKK-1. Nonetheless, we also observed an intermediate habituation phenotype in rpm-1; nsy-1 double mutants (Fig 10D). Collectively, these habituation results with JNK-1, JKK-1 and NSY-1 are consistent with the MIG-15/JNK-1 pathway functioning as a parallel opposing pathway to RPM-1 (Fig 10E, Model 2).
Previous studies have shown the C. elegans PHR protein RPM-1 functions through multiple mechanisms to regulate synapse formation and axon termination [15]. One mechanism is RPM-1 acting as a ubiquitin ligase to inhibit p38 and JNK signaling pathways. It remains unclear how broad a regulator of p38 and JNK signaling RPM-1 is. Furthermore, it is unknown if there are pathways that interact with RPM-1 to specifically affect synapse formation.
Here we identify several kinases, that are likely to form a MAP kinase pathway, that specifically suppresses synapse formation defects caused by rpm-1 (lf) in the mechanosensory neurons. This pathway is composed of the MIG-15 MAP4K, the NSY-1 MAP3K, the JKK-1 MAP2K and the JNK isoform JNK-1 (Fig 10E). Mutations in kinases of the MIG-15/JNK-1 pathway suppress synapse formation defects in the mechanosensory neurons of rpm-1 mutants, and suppress accompanying deficits in habituation. Furthermore, behavioral analysis is consistent with the MIG-15/JNK-1 pathway most likely functioning as a parallel opposing pathway to RPM-1. Thus, our results suggest the MIG-15/JNK-1 pathway needs to be opposed by RPM-1 signaling for proper glutamatergic synapse formation and short-term learning (Fig 10E, Model 2). Our results also address an important missing functional link between the MIG-15 MAP4K and JNK in the nervous system.
Biochemical studies and genetics on fly embryogenesis showed the orthologs of MIG-15 (Msn and NIK) activate JNK signaling [46, 47]. In worms and flies, MIG-15 and Msn regulate axon guidance by functioning through the cytoskeletal regulators, ERM-1 and Bifocal, rather than acting on JNK signaling [48, 49]. At present, it is unknown if MIG-15, Msn or NIK function through JNK signaling in the nervous system.
We now provide evidence that MIG-15 is likely to function in a JNK pathway to regulate synapse formation in mechanosensory neurons. If this is the case, in the developing worm nervous system MIG-15 regulates axon guidance through the actin regulator ERM-1, and regulates synapse formation through a kinase pathway that includes the JNK isoform JNK-1. Our results show mig-15 mutants have mild defects in synapse formation, and suppress the strong synapse formation defects caused by rpm-1 (lf) (Figs 2 and 3). This suggests a balance of MIG-15 signaling is important for proper glutamatergic synapse formation. The importance of our finding is accentuated by recent work showing NIK potentially mediates synapse loss during neurodegeneration in ALS [65].
Numerous observations collectively support the conclusion that mig-15, nsy-1, jkk-1 and jnk-1 function in the same pathway. 1) Suppression of synapse formation defects caused by rpm-1 is not increased when multiple kinases are eliminated compared to elimination of a single kinase (Figs 1 and 2). 2) The MIG-15/JNK-1 pathway has a specific suppressor phenotype: These kinases suppress synapse formation defects, but not axon termination defects, in the mechanosensory neurons of rpm-1 mutants (Figs 1 and 2). 3) Kinases in the MIG-15/JNK-1 pathway function cell autonomously in the mechanosensory neurons (Fig 5). 4) Transgenic overexpression of MIG-15, NSY-1 or JKK-1 causes synapse formation defects at a similar frequency (Fig 6). 5) nsy-1, jkk-1 and jnk-1 suppress defects in tap habituation in rpm-1 mutants (Fig 10). 6) Finally, our biochemical results show that NSY-1 binds to MIG-15 and JKK-1 (Fig 7). Our findings on the MIG-15/JNK-1 pathway and the order we have assigned to kinases in the pathway agrees with prior in vitro biochemistry, genetic results on fly embryogenesis, and is consistent with how MAPK cascades generally operate [45]. However, our results do not definitively rule out the alternative possibility that theses kinases could function in multiple, parallel MAPK pathways. This is important to note, as prior work has shown that NSY-1 can stimulate activation of p38 MAPK [55].
Prior work showed JKK-1 and JNK-1 function in the same pathway to regulate locomotion, synaptic vesicle trafficking and ARL-8 regulation of synaptic position [51, 52]. Our results now identify a new function for JKK-1 and JNK-1 in glutamatergic synapse formation, and suggest MIG-15 and NSY-1 are likely to be further kinases in this pathway. It is unlikely impaired synaptic vesicle trafficking in jnk-1 and jkk-1 mutants would explain our findings, as we observe the opposite phenotype, an improvement in synapse formation defects in rpm-1; jkk-1 and rpm-1; jnk-1 double mutants. However, if synaptic vesicle trafficking to glutamatergic presynaptic terminals is increased in jnk-1 and jkk-1 mutants, which remains unknown, it might explain our results. Whether MIG-15 and NSY-1 are involved in ARL-8 regulation of synaptic position like JKK-1 and JNK-1 remains an open question.
We have found that the MIG-15/JNK-1 pathway specifically suppresses synapse formation defects in the mechanosensory neurons of rpm-1 mutants. This differs from prior work showing the DLK-1/PMK-3 and MLK-1/KGB-1 pathways suppress both axon termination and synapse formation defects in mechanosensory neurons, as well as synapse formation defects in motor neurons of rpm-1 mutants [29, 30]. Because mechanosensory neurons form glutamatergic, neuron-neuron synapses that are reminiscent of mammalian central synapses, it is reasonable to speculate that the MIG-15/JNK-1 pathway might regulate central synapse formation in other organisms.
Recent studies suggested axon termination defects are a remodeling phenotype that occurs due to retraction of the PLM synaptic branch [59, 66]. This is one possible explanation for why mutants, such as rpm-1, affect both axon termination and synapse formation. However, if this were true, we would expect genetic changes to affect both axon termination and synapse formation in all cases. Our discovery of the MIG-15/JNK-1 pathway demonstrates distinct signals exist within the mechanosensory neurons that differentially influence chemical synapse formation compared to axon termination.
Accompanying differing effects of the MIG-15/JNK-1 pathway and the DLK-1/PMK-3 pathway on synapse development are differing results on habituation to tap, a form of short-term learning, that is dependent upon chemical synapse function in the mechanosensory neurons. Prior work showed rpm-1 mutants have dramatic defects in habituation to repeated tap stimulus, which is suppressed by dlk-1 [23]. Defects in habituation stem from RPM-1 function in the mechanosensory neurons. We have found nsy-1, jkk-1 and jnk-1 also suppress habituation defects in rpm-1 mutants. However, there were important differences between these kinases and DLK-1. First, jkk-1 and jnk-1 single mutants have increased habituation (Fig 10B and 10C), which was not observed in dlk-1 mutants [23]. Second, our results show that double mutants of rpm-1 with jnk-1, jkk-1 or nsy-1 are suppressed to an intermediate habituation phenotype with neither phenotype from single mutants dominating. The most likely explanation for these results is that the MIG-15/JNK-1 pathway functions as a parallel opposing pathway to RPM-1 (Fig 10E, Model 2). This differs from the DLK-1/PMK-3 pathway, which is inhibited by RPM-1 [29]. While we favor the parallel opposing pathway interpretation of our data, it is important to note our genetic results do not rule out that RPM-1 might ubiquitinate and inhibit a kinase in the MIG-15/JNK-1 pathway, in which case the MIG-15/JNK-1 pathway would function downstream of RPM-1 (Fig 10E, Model 1).
Our finding that the MIG-15/JNK-1 pathway is likely to function as a parallel opposing pathway to RPM-1 indicates that RPM-1 could be a relatively specific inhibitor of certain p38 and JNK pathways. This is consistent with prior work showing RPM-1 binds to the PP2C phosphatase PPM-2 to specifically inhibit DLK-1 [27]. Thus, our findings continue to support the concept that RPM-1 is a relatively sophisticated regulator of intracellular signaling, and not simply a general silencer of MAPK signaling [15, 27].
Our discoveries that the MIG-15/JNK-1 pathway inhibits glutamatergic synapse formation, and that impairing certain kinases in the MIG-15/JNK-1 pathway increases short-term learning have potentially important implications for neurodegenerative disease. This is particularly noteworthy, as a JNK inhibitor was shown to reduce synaptic dysfunction and improve cognitive outcomes in mouse models of Alzheimer’s disease [41–43], and recent work showed genetically impairing JNK activation improves outcomes in ALS and Alzheimer’s models [44]. The specificity of the MIG-15/JNK-1 pathway compared to the DLK-1/PMK-3 and MLK-1/KGB-1 pathways in worms suggests inhibitors of specific JNK isoforms could have differing efficacy in neurodegenerative disease models.
The N2 isolate of C. elegans was used for all experiments and worms were maintained using standard procedures. The following mutant alleles were used in this study: rpm-1 (ju44), jnk-1 (gk7), jkk-1 (km2), nsy-1 (ok593), and mig-15 (mu342). Genotyping was done by PCR and restriction digestion. nsy-1 (ok593) was maintained as a balanced strain with mIn1. mig-15 mutants were maintained at 20°C and shifted to 23°C for experiments. The MosSCI insertion strain EG6699, tTi5605 (Mos1 transposon on chromosome II); unc-119 (ed3), was used to generate the integrated transgene bggSi1 [Pmec-7MIG-15) that was used for mig-15 rescue experiments.
Other integrated transgenes used included: muIs32 [Pmec-7::GFP] II, jsIs1114 [Pmec-7GFP::RAB-3] II, jsIs973 [Pmec-7mRFP] III, juIs1 [Punc-25SNB-1::GFP] IV and bggIs28 [Pmec-7::UNC-10::tdTOMATO] III.
cDNA encoding jnk-1, jkk-1, nsy-1 or mig-15 were cloned with iProof High-Fidelity DNA polymerase (BIO-RAD) PCR from an N2 cDNA pool using standard conditions. cDNAs were Taq polished and TOPO cloned into pCR8 Gateway entry vector (Invitrogen). cDNA entry vectors were validated by sequencing and recombined into the necessary Gateway destination vectors to create final plasmids for microinjection or transfection into HEK 293 cells.
For construction of Pmec-7::UNC-10::tdTOMATO (pBG-GY757), tdTOMATO cDNA was initially subcloned by PCR, adding PpuMI and NheI sites to 5’ and 3’ ends, respectively. These enzyme sites were used to insert tdTOMATO into a Pmec-7 Gateway destination vector to create the C-terminal tagging vector pBG-GY747. unc-10 genomic DNA was TOPO cloned using pJH430 (kind gift of Dr. Mei Zhen, University of Toronto) as a PCR template, and recombined into pBG-GY747.
To build a targeting vector for mig-15 MosSCI insertion, the pCR8 mig-15 Gateway entry vector (pBG-GY596) was recombined with a Pmec-7 destination vector (pBG-GY119). The mec-7 promoter, mig-15 cDNA, and unc-54 3’UTR was amplified by PCR as a single sequence using iProof DNA polymerase. SpeI sites were included in the primers and added to the 5’ and 3’ ends of this construct. PCR product was TOPO cloned into pCR8 to create pBG-GY616 and verified by sequencing. pBG-GY616 plasmid was cut with SpeI and subcloned into the pCFJ350 MosSCI targeting vector.
Transgenic strains were derived using standard procedures. For mechanosensory neuron rescue experiments, plasmid DNA mixtures were injected in double mutants of interest with muIs32 to derive transgenic extrachromosomal arrays that used the mec-7 promoter to drive expression from jnk-1, jkk-1 or nsy-1 cDNA. mig-15 rescue was done using a MosSCI single copy integrated transgene, bggSi1, that was crossed onto rpm-1 (ju44); mig-15 (mu342) double mutants. MosSCI was necessary due to technical limitations of injecting into mig-15 mutants, which have small body and brood size. MosSCI insertion was carried out with peel-1 negative selection [67].
For overexpression experiments, kinase cDNAs were expressed using the pan-neuronal rgef-1 promoter and extrachromasomal arrays were derived using wild-type animals. PCR amplification was performed using the Roche Expand Long Template PCR system.
For localization experiments, transgenic extrachromosomal arrays were generated in wild-type animals by injecting plasmids that used the mec-3 or mec-7 promoter to express nsy-1 or jnk-1 cDNA with N-terminal GFP fusions. For colocalization, GFP::JNK-1 was injected into bggIs28. The bggIs28 transgene was generated by TMP/UV genomic integration of a transgenic array containing Pmec-7::UNC-10::tdTOMATO (injected at 25 ng/μL). We generated N and C-terminal GFP and tdTOMATO fusions with MIG-15 and JKK-1, but these constructs were not readily detected in PLM neurons across different injection concentrations using either the mec-3 or mec-7 promoters.
All injection mixes used to make extrachromasomal arrays included the co-injection marker Pmyo-2::mCherry at 1 ng/μL, and were made up to total DNA concentration of 100 ng/μL using pBluescript. S1 Table contains information for DNA concentrations for all individual injections. DNA was microinjected with using Zeiss AxioObserver.A1 microscope, Shutter Instrument XenoWorks Digital Microinjector and NARISHIGE MO-202U micromanipulator and standard injection procedures.
PLM axon termination and synaptic branch morphology was analyzed and quantified using the transgene muIs32. L4 parent worms were grown at 23°C and F1 progeny were picked to separate plates as L4 animals and maintained at 23°C. Animals were scored or imaged 16–24 hours later as young adults. This procedure was used to ensure consistent developmental staging across genotypes. Animals were anesthetized in 5 μM levamisole or 1% (v/v) 1-phenoxy-2-propanol in M9 buffer on a 2% agarose pad, and mounted and visualized with a Leica DM5000 B (CTR5000) epifluorescence microscope using 40x or 63x oil immersion objectives. Images were acquired with a CCD Leica DFC345 FX camera.
The synaptic branch was scored as a proxy for synapse formation in the PLM neurons. Defects were scored by following PLM synaptic branches down to the ventral nerve cord (VNC). Branches with terminal GFP varicosities, representing presynaptic boutons, were scored as wildtype. The absence of terminal varicosities and incomplete branches, or total absence of branches were scored as mutant. A subset of mig-15 mutants had multiple collateral PLM branches, a phenotype previously documented in the DD motor neurons for mig-15 mutants [48]. If at least one complete branch reached the VNC to form a varicosity the neuron was not counted as mutant for synaptic branch. The position of the vulva was also used to assess which branch was the normal, wild-type synaptic branch in these animals. A small proportion of mig-15 mutants displayed severe migration and guidance problems; the PLM neurons from these neurons were not included in our analysis.
To score PLM neurons simultaneously expressing Pmec-7::GFP::RAB-3 (jsIs1114) and Pmec-7::mRFP (jsIs973), the same criteria for the synaptic branch was used as above, but the presence of GFP::RAB-3 puncta was scored simultaneously. Whenever GFP::RAB-3 puncta were observed at the presynaptic terminal, the PLM synaptic branch and presynaptic bouton morphology were also normal.
Axon termination was scored as mutant if the axon overextended beyond the normal termination point (prior to the ALM cell body) and curved towards the VNC. These defects were referred to as “hook” defects.
To evaluate synapse formation in the GABAergic motor neurons, we used the transgene Punc-25::SNB-1::GFP (juIs1). Analysis of SNB-1::GFP puncta was done on young adult worms grown at 25°C. Worms were prepared for analysis as described for the PLM neurons, and anesthetized with 1% (v/v) 1-phoxy-2-propanol. Images of the dorsal cord were collected using a 40x oil immersion objective. Stretches of dorsal cord were measured using Leica Application Suite AF software, and the number of SNB-1::GFP puncta were manually counted.
L4 worms were picked to a separate plate and young adults were imaged 18–24 hours later. Slides were prepared with the same procedure described for PLM analysis with cytosolic GFP using 5 μM levamisole. Coverslips were sealed with BIOTIUM CoverGrip Coverslip Sealant. Confocal images were collected with Leica SP8 confocal microscope under 25x or 40x objectives. The following acquisition conditions were used: Bidirectional, 1.00 AU pinhole, 2.5–3.0x scan zoom factor, 400-600Hz, HyD detectors, 150–200 gain, 512x512 format, between lines sequential acquisition. Analysis conditions were kept identical across genotypes. Leica Application Suite (LAS) software was used to define Z-stacks that were collected at 1.0 to 1.5μm intervals. Maximum intensity projections of Z-stacks were made using LAS software.
Exported TIFF files of confocal stacks were analyzed in FIJI (ImageJ). Regions of interest (ROI) were defined manually for GFP::RAB-3 and UNC-10::tdTOMATO puncta within PLM presynaptic boutons. Mean area and number of puncta were calculated based on ROIs, and data was pooled from three separate experiments for each genotype.
Colchicine (0.25 mM from a 0.5 M stock solution dissolved in DMSO) was added while pouring NGM plates. Control NGM plates were made by adding an equal volume of DMSO. OP50 bacteria lawns were seeded onto colchicine or DMSO plates and used within 10 days of pouring. P0 young adult worms were placed on colchicine or DMSO plates, and F1 young adults were analyzed.
6-cm dishes of confluent HEK 293-T cells were transfected with 10 μL of Lipofectamine using the designated amount of plasmid: pBG-GY711 (FLAG-NSY-1, 2 μg), pBG-GY738 (MIG-15-HA, 4 μg) and pBG-GY784 (HA-JKK-1, 4 μg). pBluescript was added to reach 8 μg of total DNA transfected per plate. 22–26 hrs after transfection, cells were lysed with 0.1% NP-40 lysis buffer (50 mM Tris, pH 7.5, 150 mM NaCl, 10% glycerol, 1 mM DTT, and 1x Pierce HALT protease inhibitor cocktail). 0.25 to 0.5 mg of total protein from transfected cells was used for coIPs. Lysates were incubated with primary antibody for 30 min and precipitated for 4 h with 10 μL protein G agarose (Roche Applied Science) at 4°C. Precipitates were boiled in Laemmli sample buffer (Bio-Rad) with ß-mercaptoethanol (Sigma) and run on a 4–12% Bis-Tris gel (Invitrogen). Gels were transferred to PVDF membranes using Tris acetate transfer buffer and immunoblotted. Blots were visualized with HRP conjugated secondary antibodies, ECL (1:5 dilution of Supersignal West Femto (Thermo Scientific) in TBS) and x-ray film.
FLAG tagged proteins were immunoprecipitated with a mouse monoclonal anti-FLAG antibody (M2, Sigma), and immunoblotted with rabbit polyclonal anti-FLAG antibodies (Cell Signaling). HA tagged proteins were immunoprecipitated with rabbit polyclonal anti-HA antibodies (Invitrogen) and immunoblotted with a rabbit monoclonal anti-HA antibody (C29F4, Cell Signaling). For anti-HA blots, anti-rabbit light chain specific secondary antibodies were used.
Tap habituation experiments were performed as described previously with minor modifications [23]. Behavioral recordings were collected using a modified Multi-Worm Tracker. Age-synchronized animals (~50–100) were cultivated from egg until gravid adult (~3 days) at 23°C, and assayed on 5 cm NGM plates with 50 μl of OP50 E. coli. Behavior of animals was recorded for 550 seconds, and animals were mechanically stimulated by tapping the side of the plate with an automated linear solenoid after the first 100 seconds. Plates were tapped 45 times with a 10 second inter-stimulus interval. Response was measured as reversal probability, which was estimated by the proportion of worms that reversed within 2 seconds of each tap stimulus. Because reversal probabilities to the initial response differed in some genotypes (mean ± SEM, Fig 10B: wildtype = 0.90 ± 0.01, rpm-1 = 0.98 ± 0.01, jnk-1 = 0.78 ± 0.03, rpm-1; jnk-1 = 0.53 ± 0.03; Fig 10C: wildtype = 0.89 ± 0.02, rpm-1 = 0.96 ± 0.01, jkk-1 = 0.81 ± 0.02, rpm-1; jkk-1 = 0.75 ± 0.01; and Fig 10D: wildtype = 0.90 ± 0.01, rpm-1 = 0.93 ± 0.01, nsy-1 = 0.80 ± 0.02, rpm-1; nsy-1 = 0.82 ± 0.02), responses were standardized as the percent mean of initial response. For each plate, exponential curves were fit to the responses across stimuli, and habituation level was measured as the value of the fit at the final stimulus. All strains used contained the muIs32 transgene.
|
10.1371/journal.pcbi.1004888 | Identifying Network Perturbation in Cancer | We present a computational framework, called DISCERN (DIfferential SparsE Regulatory Network), to identify informative topological changes in gene-regulator dependence networks inferred on the basis of mRNA expression datasets within distinct biological states. DISCERN takes two expression datasets as input: an expression dataset of diseased tissues from patients with a disease of interest and another expression dataset from matching normal tissues. DISCERN estimates the extent to which each gene is perturbed—having distinct regulator connectivity in the inferred gene-regulator dependencies between the disease and normal conditions. This approach has distinct advantages over existing methods. First, DISCERN infers conditional dependencies between candidate regulators and genes, where conditional dependence relationships discriminate the evidence for direct interactions from indirect interactions more precisely than pairwise correlation. Second, DISCERN uses a new likelihood-based scoring function to alleviate concerns about accuracy of the specific edges inferred in a particular network. DISCERN identifies perturbed genes more accurately in synthetic data than existing methods to identify perturbed genes between distinct states. In expression datasets from patients with acute myeloid leukemia (AML), breast cancer and lung cancer, genes with high DISCERN scores in each cancer are enriched for known tumor drivers, genes associated with the biological processes known to be important in the disease, and genes associated with patient prognosis, in the respective cancer. Finally, we show that DISCERN can uncover potential mechanisms underlying network perturbation by explaining observed epigenomic activity patterns in cancer and normal tissue types more accurately than alternative methods, based on the available epigenomic data from the ENCODE project.
| Certain genes can regulate other genes’ expression and activity levels to perform key biological processes in a cell. Understanding how genes affect expression levels among one another is a fundamental goal in molecular biology. New statistical techniques that analyze genome-wide mRNA expression data obtained from different individuals can improve our understanding of gene interactions. In this paper, we focus on identifying the genes involved in changes in the topology of gene networks between two distinct biological status, for example, between diseased tissue and normal tissue. Our method, called DISCERN, takes two expression datasets, each from a different condition, for example, cancer and normal tissue, and identifies the genes that are likely to change their neighbors in the estimated gene networks between disease and normal conditions. Most analysis methods that compare gene expression datasets from two conditions address the question of which genes are significantly differentially expressed between conditions. The DISCERN method addresses a distinct question concerning which genes are significantly rewired in the inferred gene-regulator network in disease tissues. We show that DISCERN successfully identifies known disease specific genes, for example, genes whose expression levels are significantly associated with survival time in cancer, genes that have known disease specific driver mutations, as well as genes that have epigenomic evidence of regulatory rewiring between cancer and normal tissue types.
| Genes do not act in isolation but instead work as part of complex networks to perform various cellular processes. Many human diseases including cancer are caused by dysregulated genes, with underlying DNA or epigenetic mutations within the gene region or its regulatory elements, leading to perturbation (topological changes) in the network [1–7]. This can ultimately impair normal cell physiology and cause disease [8–11]. For example, cancer driver mutations [12–19] on a transcription factor can alter its interactions with many of the target genes that are important in cell proliferation (Fig 1A). A key tumor suppressor gene can be bound by different sets of transcription factors between cancer and normal cells, which leads to different roles [6, 20–23] (Fig 1B). Recent studies stress the importance of identifying the perturbed genes that create large topological changes in the gene network between disease and normal tissues as a way of discovering disease mechanisms and drug targets [8, 9, 24–27]. However, most existing analysis methods that compare expression datasets between different conditions (e.g., disease vs. normal tissues) focus on identifying the genes that are differentially expressed [28–30]. For example, a recent review paper on biological network inference [31] emphasized that there is a lack of methods that focus on inferring the differential network between different conditions (e.g., distinct species, and disease conditions).
Several recent studies compare gene networks inferred between conditions based on expression datasets [1, 32–37]. They fall into three categories: 1) Network construction based on prior knowledge: West et al. (2012) computes the local network entropy, based on the protein interaction network from prior knowledge and expression datasets from cancer and normal tissues [1]. 2) Pairwise correlation-based networks: Guan et al. (2013) [35] proposed the local network similarity (LNS) method to compare the pairwise Pearson’s correlation matrices of all genes between two conditions. Still other authors compared pairwise correlation coefficients for all gene pairs between conditions with different correlation measures including t-test p-values [32, 33, 36]. 3) Learning a condition-specific conditional dependency network for each condition and comparing the networks between conditions: Gill et al. (2010) proposed a method, called PLSNet, that fits a partial least squares model to each gene, computes a connectivity scores between genes, and then calculates the L1 distance between score vectors to estimate network perturbation [34]. Zhang et al. (2009) proposed a differential dependency network (DDN) method that uses lasso regression to construct networks, followed by permutation tests to measure the significance of the network differences [37].
There have been approaches to identify dysregulated genes in cancer by utilizing multiple types of molecular profiles, not based on network perturbation across disease states estimated based on expression data. Successful examples use a linear model to infer each gene expression model based on copy number variation, DNA methylation, ChIP-seq, miRNAs or mRNA levels of transcription factors [38–40]. The advantages of the aforementioned methods that take only expression datasets as input to identify perturbed genes are in their applicability to diseases for which only expression data are available. In this paper, we focus on identifying perturbed genes purely based on gene expression datasets representing distinct states, and compare our method with existing method, LNS, D-score and PLSNet.
We present a new computational method, called DISCERN (DIfferential SparsE Regulatory Network), to identify perturbed genes, i.e. the genes with differential connectivity between the condition specific networks (e.g., disease versus normal). DISCERN takes two expression datasets, each from a distinct condition, as input, and computes a novel perturbation score for each gene. The perturbation score captures how likely a given gene has a distinct set of regulators between conditions (Fig 1A). The DISCERN method contains specific features that provide advantages over existing approaches: 1) DISCERN can distinguish direct associations among genes from indirect associations more accurately than methods that focus on marginal associations such as LNS; 2) DISCERN uses a penalized regression-based modeling strategy that allows efficient inference of genome-wide gene regulatory networks; and 3) DISCERN uses a new likelihood-based score that is more robust to the expected inaccuracies in local network structure estimation. We elaborate on these three advantages below:
First, DISCERN infers gene networks based on conditional dependencies among genes—a key type of probabilistic relationship among genes that is fundamentally distinct from correlation. If two genes are conditionally dependent, then by definition, their expression levels are still correlated even after accounting for (e.g., regressing out) the expression levels of all other genes. Thus, conditional dependence relationship is less likely to reflect transitive effects than mutual correlation, and provides stronger evidence that those genes are functionally related. These functional relationships could be regulatory, physical, or other molecular functionality that causes two genes expression to be tightly coupled. As a motivating example, assume that the expression levels of genes ‘3’ and ‘5’ are regulated by gene ‘1’ in a simple 7-gene network (Fig 1A). This implies that the expression level of gene ‘1’ contains sufficient information to know the expression levels of genes ‘3’ and ‘5’. In other words, genes ‘3’ and ‘5’ are conditionally independent from each other and from the rest of the network given gene ‘1’.
Second, DISCERN uses an efficient neighborhood selection strategy based on a penalized regression to enable the inference of a genome-wide network that contains tens of thousands of genes. Penalized regression is a well established technique to identify conditional dependencies [41]. Inferring the conditional dependence relationships from high-dimensional expression data (i.e., where the number of genes is much greater than the number of samples) is a challenging statistical problem, due to a very large number of possible network structures among tens of thousands of genes. Unlike pairwise correlation, the conditional dependence between ‘1’ and ‘2’ cannot be measured based on just the expression levels of these two genes. We should consider the possible networks among all genes and find the one that best explains the expression data. This involves both computational and statistical challenges. To make this process feasible, DISCERN uses a sparse regression model for each gene to select neighbors in the network [41, 42]. The use of a scalable method to infer a genome-wide conditional dependence network is a key distinguishing feature of the DISCERN method.
Finally, one of the most novel features of DISCERN is the ability to avoid the overestimation of the degree of network perturbation due to dense correlation among many genes. Revisiting the 7-gene network example (Fig 1A), assume that genes ‘5’ and ‘7’ are highly correlated to each other, in which case a penalized regression that imposes a sparsity penalty, such as the lasso method, may arbitrarily select one of them. This can result in a false positive edge between genes ‘1’ and ‘7’ instead of ‘1’ and ‘5’. This may lead to overestimation of the perturbation of gene ‘1’ (Fig 1A). Our network perturbation score overcomes this limitation by measuring the network differences between conditions based on the likelihood when the estimated networks are swapped between conditions—not based on the differences in topologies of the estimated networks. We demonstrate the effectiveness of this feature by comparing with methods based on the topology differences of the estimated networks.
We evaluated DISCERN on both synthetic and gene expression data from three human cancers: acute myeloid leukemia (AML), breast cancer (BRC), and lung cancer (LUAD). Integrative analysis using DISCERN on epigenomic data from the Encyclopedia of DNA Elements (ENCODE) project leads to hypotheses on the mechanisms underlying network perturbation (Fig 1C). The resulting DISCERN score for each gene in AML, BRC and LUAD, the implementation of DISCERN, and the data used in the study are freely available on our website http://discern-leelab.cs.washington.edu/.
Here, we describe the DISCERN method, referring to the Methods for a full description. We postulate that a gene can be perturbed in a network largely in two ways: A gene can change how it influences other genes (Fig 1A), for example, a driver mutation on a transcription factor can affect cell proliferation pathways [12–19]. A gene can change the way it is influenced by other genes, a common example being when a mutated (genetically or epigenetically) gene acquires a new set of regulators, which occurs frequently in development and cancer [6, 20–23] (Fig 1B). Identifying the genes that are responsible for large topological changes in gene networks could be crucial for understanding disease mechanisms and identifying key drug targets [8, 9, 24–27]. However, most current methods for identifying genes that behave differently in their expression levels between diseased and normal tissues focus on differential expression [28–30], rather than differential connection with other genes in a gene expression network (Fig 1A).
We model each gene’s expression level using a sparse linear model (lasso regression): let y i ( s ) be expression levels of gene i in an individual with state s, cancer (s = c) or normal (s = n), modeled as: yi(s)≈∑r=1pwir(s)xr(s). Here, x1, …xp denote candidate regulators, a set of genes known to regulate other genes, including transcription factors, chromatin modifiers or regulators, and signal transduction genes, which were used in previous work on network reconstruction approaches [43–46] (S1 Table). Linear modeling allows us to capture conditional dependencies efficiently from genome-wide expression data containing tens of thousands genes. Naturally, a zero weight wir indicates that a regulator r does not affect the expression of the target gene i. Sparsity-inducing regularization helps to select a subset of candidate regulators, which is a more biologically plausible model than having all regulators, and makes the problem well-posed in our high-dimensional setting (i.e., number of genes ≫ number of samples).
To determine the regulators for any given gene, we use a lasso penalized regression model [47] with the optimization problem for each lasso regression defined as: argmin w i 1 ( s ) , … , w i p ( s ) ∑ j = 1 n ( y i j ( s ) − ∑ r = 1 p w i r ( s ) x r j ( s ) ) 2 + λ ∑ r = 1 p | w i r ( s ) | , where y i j ( s ) means the expression level of the ith gene in the jth patient in the sth state, and x r j ( s ) similarly means the expression level of the rth regulator in the jth patient in the sth state. The second term, the L1 penalty function, will zero out many irrelevant regulators for a given gene, because it is known to induce sparsity in the solution [47]. We normalize the expression levels of each gene and each regulator to be mean zero and unit standard deviation, a process called standardization, which is a standard practice before applying a penalized regression method [47–50]. The difference in the weight vector between conditions, w i ( n ) and w i ( c ), can indicate a distinct connectivity of gene i with p regulators between the conditions. However, simply computing the difference of the weight vectors is unlikely to be successful, due to the correlation among the regulators. The lasso, or other sparsity-inducing regression methods, can arbitrarily choose different regulators between cancer and normal. Examining the difference in the weight vectors between conditions would therefore lead to overestimation of network perturbation.
Instead, DISCERN adopts a novel network perturbation score that measures how well each weight vector learned in one condition explains the data in a different condition. This increases the robustness of the score to correlation among regulators, as demonstrated in the next section. We call this score the DISCERN score, defined as DISCERN i = per sample -log-likelihood computed using learned weights w ( s ) in a different condition s′ per sample -log-likelihood computed using learned weights w ( s ) in the same condition s. This is equivalent to DISCERN i = e r r i ( c , n ) + e r r i ( n , c ) e r r i ( c , c ) + e r r i ( n , n ), where e r r i ( s , s ′ ) = 1 n s | | y i ( s ) - ∑ r = 1 p w i r ( s ′ ) x r ( s ) | | 2 2. Here ns is the number of samples in the data from condition s. The numerator measures the error of predicting gene i’s expression levels in cancer (normal) based on the weights learned in normal (cancer). If gene i has different sets of regulators between cancer and normal, it is likely to have a high DISCERN score. The denominator plays an important role as a normalization factor, which is demonstrated by comparing with an alternative score, namely the D0 score (Fig 2A), that uses only the numerator of the DISCERN score. We also compare with existing methods, such as LNS [35] and PLSNet [34], that compare the weight vectors between cancer and normal models where we demonstrate the advantages of the likelihood-based model that DISCERN uses.
In order to systematically compare DISCERN with alternative methods in a controlled setting, we performed validation experiments on 100 pairs of synthetically generated datasets representing two distinct conditions. Each pair of datasets contains 100 variables drawn from the multivariate normal distribution with zero mean and covariance matrices Σ1 and Σ2. We divided 100 variables into the following three categories: 1) variables that have different sets of edge weights with other variables across two conditions, 2) variables that have exactly the same sets of edge weights with each other across the conditions, and 3) variables not connected with any other variables in the categories 2) and 3) in both conditions. For example, in Fig 1A, ‘1’ is in category 1). ‘2’, ‘4’, ‘6’, and ‘7’ are in category 2), and ‘3’ and ‘5’ is in category 3). We describe how we generated the network edge weights (i.e., elements of Σ 1 - 1 and Σ 2 - 1) among the 100 variables in more detail in Methods.
We compared DISCERN with 4 alternative methods to identify perturbed genes: LNS [35], D-score [36], PLSNet [34], and D0 that uses only the numerator of the DISCERN score. Here, we do not compare with the methods to identify differentially expressed genes, such as ANOVA, because the synthetic data were generated from a zero mean Gaussian distribution. We note that the PLSNet method uses empirical p-values as the network perturbation scores, where the empirical p-value for each gene is estimated from permutation tests that generate the null distribution of the gene’s score [34]. All the other methods, such as DISCERN, LNS, and D-score, do not require permutation tests (see Methods for details). To show that DISCERN outperforms existing methods and those that use the empirical p-values obtained through permutation tests as the network perturbation scores, we developed the following methods for comparison: LNS, D-score, and D0 followed by permutation tests to compute the empirical p-values, called pLNS, pD-score, and pD0, respectively.
The average receiver operating characteristic (ROC) curves across 100 pairs of datasets for these methods (Fig 2A) show that DISCERN significantly outperforms all the other 7 methods—3 existing methods (LNS, D-score, and PLSNet), and 4 methods we created for comparison (D0, pD0, pLNS, and pD-score). Except DISCERN, PLSNet performs the best among all existing methods. However, its run time grows too quickly as the number of variables increases, which makes it two to three orders of magnitude slower than DISCERN when run on larger data (Fig 2B). PLSNet was too slow to run on genome-scale data and therefore we did not use it for the subsequent experiments on genome-wide gene expression data from cancer patients.
We note that DISCERN does not need permutation tests to generate the null distribution of the score for each gene. All other methods improve when the empirical p-values from permutation tests are used, which indicates that the gene-level bias on the magnitude of the raw scores hurts their performance to identify perturbed genes. DISCERN significantly outperforms D0 that uses only the numerator of the DISCERN score, which indicates that the denominator of the DISCERN score plays a role to normalize the score such that the scores of different genes can be compared to each other. Computing the empirical p-value for each gene based on the gene-specific null distribution obtained through permutation tests is not feasible on genome-wide data. To obtain a p-value of 0.05 after Bonferroni correction, we need at least (1/0.05 × p) permutation tests per gene, where p is the total number of genes, and (1/0.05 × p2) permutation tests in total. When p = 20,000, this number is (4 × 109) permutation tests, which is not feasible even when using multiple processors at a reasonable cost. This is demonstrated in Fig 2B that shows the run time of PLSNet, a permutation test-based method, when applied to data containing a varying number of genes (p).
We used genome-wide expression datasets consisting of 3 acute myeloid leukemia (AML) datasets, 3 breast carcinoma (BRC) datasets and 1 lung adenocarcinoma (LUAD) dataset (Table 1). Details on the data processing are provided in Methods. To evaluate the performance of the DISCERN method, we compared DISCERN with existing methods that scale to over tens of thousands of genes: LNS [35] and D-score [36] that aim to estimate network perturbation, and ANOVA that measures differential expression levels between cancer and normal samples.
We first computed the DISCERN, LNS, D-score, and ANOVA scores in the 3 cancers based on the following datasets that contain normal samples: AML1, LUAD1 and BRC1 (Table 1). Then, we used the rest of the datasets to evaluate the performance of each method at identifying genes previously known to be important in the disease, for example, the genes whose expression levels are significantly associated with survival time in cancer. The value of the sparsity tuning parameter λ was chosen via cross-validation tests, a standard statistical technique to determine the value of λ [47]. For the chosen λ values, the overall average regression fit measured by cross-validation test R2 was 0.493.
To remove any potential concern of the effect of standardization on genes with very low expression level, we first show that genes with low mean expression do not tend to have high enough DISCERN score to be considered in our evaluation in the next sections (S1 Fig). The Pearson’s correlation between the mean expression before standardization and the DISCERN score ranges from 0.08 and 0.43 across datasets. Positive correlation is induced because genes with low mean expression tend to have lower DISCERN scores, indicating that genes whose expression are likely essentially noise would not be selected as high-scoring genes. To further reduce the potential concern of genes with low expression in RNA-seq data (LUAD), we applied the voom normalization method that is specifically designed to adjust for the poor estimate of variance in count data, especially for genes with low counts [51].
We assessed the significance of the DISCERN scores through a conservative permutation testing procedure, where we combined cancer and normal samples, and permuted the cancer/normal labels among all samples (more details in Methods). Unlike the gene-based permutation test described in the previous section, here, we generate a single null distribution for all genes, which requires a significantly less number of permutation tests (one million in this experiment). After applying false discovery rate (FDR) correction on these p-values, there are 1,351 genes (AML), 2,137 genes (BRC), and 3,836 (LUAD) genes whose FDR corrected p-values are less than 0.05. We consider these genes to be significantly perturbed genes (S2 Table). The difference in these numbers of significant perturbed genes identified by DISCERN is consistent with a prior study that showed that lung cancer has a larger number of non-synonymous mutations per tumor than breast cancer, which has a larger number than AML [52].
The 1,351 genes that were predicted to be significantly perturbed between AML samples and normal non-leukemic bone marrow samples were enriched for genes causally implicated previously in AML pathogenesis (S2 Table). This include a number of genes that we and others have previously identified as being aberrantly activated in leukemic stem cells such as BAALC, GUCY1A3, RBPMS, and MSI2 [55–57]. This is consistent with over-production of immature stem-cell like cells in AML, which is a major driver of poor prognosis in the disease. Prominent among high-scoring DISCERN genes were many HOX family members, which play key roles in hematopoietic differentiation and in the pathogenesis of AML [58]. HOX genes are frequently deregulated by over-expression in AML, often through translocations that result in gene fusions. The highest ranked gene in AML by DISCERN is HOXB3 which is highly expressed in multipotent hematopoietic progenitor cells for example. Thirteen (out of 39 known) HOX genes are in the 1,351 significantly perturbed genes (p-value: 5.99 × 10−6).
When compared to known gene sets from the Molecular Signature Database (MSigDB) [59] in an unbiased way, the top hit was for a set of genes (VERHAAK_AML_WITH_NPM1_MUTATED_DN; p-value: 2 × 10−86) that are down-regulated by NPM1 (nuclephosmin 1) mutation in AML (S1 File). NPM1 is one of three markers used in AML clinical assessment; the others are FLT3 and CEBPA that are significantly perturbed genes identified by DISCERN as well. Mutation leads to aberrant cytoplasmic location of itself and its interaction partners, leading to changes in downstream transcriptional programs that are being captured by DISCERN. Also highly significant were genes highly expressed in hematopoietic stem cells [60] (JAATINEN_HEMATOPOIETIC_STEM_CELL_UP; p-value: 6 × 10−74). Among these were key regulators of hematopoietic system development such as KIT, HOXA3, HOXA9, HOXB3 (with the latter homeobox genes also implicated in AML etiology), as well as FLT3 which plays a major role in AML disease biology, with its mutation and constitutive activation conferring significantly worse outcomes for patients [61]. Comparison to Gene Ontology (GO) categories identified dysregulation of genes involved in hemostasis and blood coagulation, a key clinical presentation of AML. Furthermore, GTPase activity/binding and SH3/SH2 adaptor activity were enriched among high-scoring DISCERN genes. These are pertinent to AML due to previously noted high expression in AML leukemic stem cells of GUCY1A3 and SH3BP2, both identified as perturbed genes by DISCERN [55]. However, their function has not been examined in detail, suggesting that they are potential targets for further investigation as to their role in AML disease mechanisms. Several other highly significant enrichments were for AML subtypes that are driven by specific translocations, including MLL (mixed lineage leukemia) translocation with various partners, as well as t(8;21) translocations. The latter is of particular interest, since it is primarily a pediatric AML, whereas our network analysis uses purely adult AML samples—indicating the potential to uncover putative mechanisms that generalize beyond the context of the immediate disease type.
There are 3,836 significantly perturbed genes identified by DISCERN in lung cancer (LUAD) (S2 Table). The 3rd and 4th highest ranked genes are ICOS (inducible costimulator) and YWHAZ (14-3-3-zeta). Both genes have known roles in disease initiation or progression in lung cancer. Polymorphisms in ICOS have been associated with pre-disposition to non-small cell lung cancer [62], while over-expression of YWHAZ is known to enhance proliferation and migration of lung cancer cells through induction of epithelial-mesenchymal transitions via beta-catenin signaling [63]. GIMAP5 (GTPase IMAP Family Member 5), another high scoring LUAD gene (11th), is consistently repressed in paired analyses of tumor vs normal lung tissue from the same patient, and encodes an anti-apoptotic protein [64]. Down-regulation of GIMAP5 in lung tumors therefore potentially facilitates their evasion of programmed cell death, one of the hallmarks of cancer.
Several of the GO biological categories enriched in 3,836 high-scoring DISCERN genes in LUAD (FDR-corrected p-value <0.05) reflected metabolic and proliferative processes that are commonly de-regulated in solid tumors such as lung adenocarcinoma. Among these were cellular response to stress, mitotic cell cycle, amino acid metabolism, and apoptosis (S1 File). In fact the top-ranked gene was MCM7 (minichromosome maintenance protein 7), an ATP-dependent DNA helicase involved in DNA replication which has been implicated in carcinogenesis previously due to its function as a binding partner of PRMT6 [65]. Moreover, it was specifically identified as being a potential therapeutic target due to its over-expression in solid tumors relative to normal tissues. The high ranking of genes associated with apoptosis is consistent with the fact that there is often high rate of tumor cell death. Although the highly-ranked CARD6 (caspase recruitment domain family member 6) functions in apoptotic processes, it is also known as a regulator of downstream NF-κβ signaling. Indeed, consistent with this, we found enrichment for NF-κβ signaling pathway genes among high DISCERN-scoring genes in LUAD including NFKBIB (NF-κβ inhibitor β) which inhibits the NF-κβ complex by “trapping” it in the cytoplasm, preventing nuclear activation of its downstream targets. Although the role of NFKBIB in lung cancer has not been studied extensively, its related family member NFKBIA is known to be a silencer in non-small-cell lung cancer patients with no smoking history, suggesting that it could play some role in LUAD that arises through inherent genetic influences, or environmental insults other than smoking [66]. Levels of β-catenin have been known for some time to influence progression and poor prognosis in LUAD, potentially through its role in differentiation and metastasis from primary tumor sites [67]. We found that components of β-catenin degradation pathways—including most notably CTNNBIP1 (β-catenin interacting protein 1)—ranked among the most significant DISCERN genes in our LUAD analysis.
When comparing to other sets of genes in MSigDB, we also found targets of transcription factors including MYC, which is often de-regulated in solid tumors (either by mutation or copy number variation), and targets of the polycomb repressive complex gene EZH2. The developmental regulator EZH2 functions through regulation of DNA methylation [68], and has been implicated in B-cell lymphomas through somatic mutations [69], promotion of transformation in breast cancer [70], as well as progression in prostate cancer [71]. Interestingly, the most highly dys-regulated gene set identified by comparison to GO categories in LUAD was one related to NGF (nerve growth factor)-TrkA signaling. There are a few reports on the relevance of this axis to cancers including neuroblastoma, ovarian cancer, and a possible role in promoting metastasis in breast cancer. However, its striking appearance as the most significant hit for high-ranking DISCERN genes suggests that it merits study in lung cancer.
Here, we did the functional enrichment analysis with 2,137 genes identified by DISCERN to be significantly perturbed in breast cancer (BRC) (S2 Table). BRC showed perturbation of distinct genes and sets of genes in comparison to LUAD, as well as similarities. Again, these included GO biological processes that one would generically expect to be over-activated in a solid tumor, such as translation intiation, cell cycle, proliferation, and general cellular metabolic processes. As with LUAD, targets of MYC were enriched in high-scoring DISCERN genes in BRC. Another high-scoring group in BRC was comprised of genes that are highly correlated with each other, but with this relationship de-regulated by BRCA1 mutation [72]. Additional significant overlaps were identified with luminal A, luminal B, HER2-enriched, and basal-like breast cancer subtype-specific genes that are associated with clinical outcomes [73], and genes associated with ER-positive breast cancer [74]. The 3rd highest ranked DISCERN gene was BRF2 (TFIIB-related factor 2). BRF2 is a known oncogene in both breast cancer and lung squamous cell carcinoma, and a core RNA polymerase III transcription factor that senses and reacts to cellular oxidative stress [75]. A GO category associated with NGF (nerve growth factor)-TrkA signaling shows the highest overlap with DISCERN genes in BRC (p-value: 3.16 × 10−104). NGF-TrkA signaling is upstream of the canonical phosphatidylinositol 3-kinase (PI3K) –AKT and RAS –mitogen-activated protein kinase (MAPK) pathways, both of which impinge on cell survival and differentiation. In the context of breast cancer, over-expression of TrkA has been connected to promoting growth and metastasis, as an autocrine factor, presumably due to its influence on PI3K-AKT and RAS/MAPK [76]. TrkA is reportedly over-expressed in breast carcinoma relative to normal breast tissue in a majority of cases [77], supporting the high-ranking of genes in this pathway by DISCERN. Taken together, these results indicate that DISCERN highly ranks genes that are connected to known phenotypic and survival-associated processes in breast cancer. However, intriguingly the top DISCERN gene was CLNS1A (chloride nucleotide-sensitive channel 1A). This chloride channel gene has not, to our knowledge, been implicated in pathogenesis in any cancer, although it is a member of the BRCA1-related correlation network noted above. In fact there appear to have been few studies of its function although Entrez gene notes that it performs diverse functions.
In this section, we focus on the quantitative assessment of DISCERN and the comparison with LNS and D-score in terms of how much the identified genes are enriched for genes implicated to be important in the disease. Specifically, genes whose expression levels are significantly associated positively or negatively with survival time are often considered to be associated with tumor aggression. Identifying such genes has been considered as an important problem by a number of authors, where breast cancer was one of the first cancers to show promise in terms of identifying clinically relevant biomarkers [78, 79]. Here, we evaluated DISCERN based on how well it reveals survival-associated genes identified in an available independent dataset.
We chose the datasets with measures of patient prognosis: AML2, BRC2, and LUAD1. AML2 and BRC2 were not used for computing any scores (DISCERN, LNS, D-score, and ANOVA). For each of these datasets we computed the survival p-values based on the Cox proportional hazards model [80] measuring the association between each gene’s expression level and survival time. We defined survival-associated genes as the genes whose expression levels are associated with survival time based on the Cox proportional hazards model (p-value < 0.01) (S3 Table).
We considered the genes whose DISCERN scores are significantly high at FDR corrected p-value < 0.05 in each cancer: 1,351 genes (AML), 2,137 genes (BRC), and 3,836 genes (LUAD). We first computed the Fisher’s exact test p-values to measure the statistical significance of the overlap between these significantly perturbed genes and survival-associated genes in each of three cancers. For each cancer, we compared with existing methods to detect network perturbation—LNS and D-score—when exactly the same number of top-scoring genes were considered (Fig 3A–3C). Since these numbers of genes were chosen specifically for DISCERN, there is a chance that LNS and D-score would show a higher enrichment for survival-associated genes if different numbers of top-scoring genes were considered. As discussed in the previous section, performing the gene-based permutation tests to estimate the confidence of each gene’s score in genome-wide data is not feasible. Instead, we compared the Fisher’s exact test p-values of the three methods across a range of numbers of top-scoring genes from 0 to N Fig 3D–3F. It is pretty clear that neither LNS nor D-score would be better than DISCERN in revealing survival-associated genes, even when different numbers of top-scoring genes were considered across all cancer types.
ANOVA is a well-established method to identify differentially expressed genes across distinct conditions; DISCERN LNS, and D-score are methods to identify differentially connected genes across conditions. Therefore, the purpose of the comparison with ANOVA is not to evaluate DISCERN in identifying survival-associated genes as perturbed genes. The purpose is to compare between differentially expressed genes (that are commonly considered important) and perturbed genes estimated by the three methods (DISCERN, LNS, and D-score), in terms of the enrichment for genes with potential importance to the disease. For ANOVA, in Fig 3A–3C, we considered 8,993 genes (AML), 7,922 genes (BRC) and 13,344 genes (LUAD) that show significant differential expression between cancer and normal samples at FDR corrected p-value < 0.05. The perturbed genes identified by DISCERN are more associated with survival than differentially expressed genes captured by ANOVA in AML and LUAD (Fig 3).
In addition to the comparison with other methods—LNS and D-score—we also compare with frequently mutated genes and genes annotated to be involved in the respective cancer. We considered the following three gene sets: 1) a gene set constructed based on the gene-disease annotation database, Malacards [81], 2) genes known to have cancer-causing mutations based on the Cancer Gene Census [82], and 3) genes predicted to have driver mutations identified by MutSig [5] applied to The Cancer Genome Atlas (TCGA) data for the respective cancer type. The Malacards (gene set #1) and TCGA driver gene sets (#3) are generated for each cancer type—AML, breast cancer, or lung cancer. For example, for Malacards, we used the genes that are annotated to be involved in AML in Malacards to compare it with DISCERN genes identified in AML. Similarly, for the TCGA driver gene sets (#3), we used the AML TCGA data to identify the frequently mutated genes that are likely driver genes, and compared with high DISCERN-scoring genes in AML. We used the breast cancer TCGA data for BRC, and lung cancer TCGA data for LUAD. The Cancer Gene Census (CGC) gene set is a rigorously defined set of genes with multiple sources of evidence that its genes are cancer drivers in a single or multiple cancers.
For each cancer type, we compared these three sets of genes with the perturbed genes identified by DISCERN—1,351 (AML), 2,137 (BRC), and 3,836 (LUAD) genes with high DISCERN scores—on the basis of the significance of the enrichment for survival-associated genes. S2 Fig shows that the perturbed genes identified by DISCERN are more significantly enriched for survival-associated genes.
In this section, we evaluated the DISCERN score based on how well it identifies genes that are predictive of patient prognosis. Here, we test the possibility of using the network perturbed genes identified by DISCERN as prognostic markers. For the cancer types with at least three data sets (AML and BRC; see Table 1), we construct a survival time prediction model using the genes with significant DISCERN scores (AML: 1,351 genes, BRC: 2,137 genes) identified based on one data set (Data # 1: AML1 and BRC1) as described in the previous subsection. Then, we trained the prediction model using one of the other datasets (Data #2: AML2 and BRC2) not used for the computation of the DISCERN score. Finally, we tested the prediction accuracy on the third data set (Data #3: AML3 and BRC3).
We controlled for clinical covariates whose data are available—age in case of AML and age, grade and subtype in case of BRC—by adding them as unpenalized covariates into our elastic net Cox regression model. We trained the Cox regression model using Data #2 and tested the survival prediction model on Data #3. Since we evaluated the survival prediction in separate data (AML3 and BRC3) that were not used when training the survival prediction model, using more predictors, e.g., by adding clinical covariates, does not necessarily improve the prediction performance. Adding more predictors often leads to a higher chance of overfitting. Our survival prediction model based on the high DISCERN-scoring genes works at least as well as models based on the genes contained in the previously established prognosis markers, such as Leukemic Stem Cell score (LSC) [54] for AML and MammaPrint signature (with ∼70 genes) [83] for BRC, as shown in Fig 4. The c-index in AML is 0.669 with standard error (se) being 0.031 (Fig 4B); in BRC, the c-index is 0.668 (se: 0.027) (Fig 4D). The DISCERN-based expression marker with clinical covariates makes better predictions than when clinical covariates alone are used.
One of the possible mechanisms underlying network perturbation identified in gene expression datasets representing different conditions (e.g., cancer and normal) is the following: A transcription factor (TF) ‘X’ binds to a gene ‘Y”s promoter or its enhancer region in cancer but not in normal (or vice versa). Then, ‘X’ or its co-regulator could be an expression regulator for ‘Y’ in cancer but not in normal (or vice versa), and Y is identified as a perturbed gene (i.e., a high DISCERN-scoring genes). It is possible that ‘X”s binding information is not available and ‘X”s protein level is not reflected in its mRNA expression level; thus we cannot expect the DISCERN score of a gene inferred from expression data to be perfectly correlated with whether the gene has a differential biding of a certain TF, inferred from ChIP-seq or DNase-seq data. However, the degrees of correlation between the network perturbation score (DISCERN, LNS or D-score) of a gene and whether a TF differentially binds to the gene can be a way to evaluate the network perturbation scoring methods.
To determine whether or how much our statistical estimates of network perturbation reflects perturbation of the underlying TF regulatory network, we queried epigenomic data from ENCODE project. Two of the ENCODE cell lines—NB4 (an AML subtype [84]) and CD34+ (mobilized CD34 positive hematopoietic progenitor cells)—are closest to AML and normal conditions, and the DNase-seq data from these cell lines are available. We used the DNase-seq data from NB4 and the position weight matrices (PWMs) of 57 TFs available in the JASPAR database [85] to find the locations of the PWM motifs that are on the hypersensitive regions. This is a widely used approach to estimate active binding motifs using DNase-seq data, when ChIP-seq data are not available. We identified the locations of these PWM motifs on the hg38 assembly by using the FIMO [86] method (p-value ≤10−5). We then intersected these motif locations with hypersensitive regions identified by the DNase-seq data for each TF. We repeated for the other cell line CD34+.
For each TF, we measured how well the DISCERN score of a gene can predict the differential binding of the TF in active enhancer regions (marked by H3K27Ac) within 15kbs of the transcription start site (TSS) of the gene (Fig 5A–5C) and 5kb of the gene between blood cancer and normal cell lines (NB4 and CD34+) (S3A–S3C Fig). We show that the DISCERN score can reflect differential binding of most of the TFs better than existing methods to identify network perturbation (LNS and D-score) and a method to identify differentially expressed genes (ANOVA). As a way to summarize these results across all 57 TFs, we computed the Pearson’s correlation between the score of each gene and the proportion of TFs that differentially bind to that gene out of all TFs that bind to that gene. Fig 5D shows that DISCERN detects genes with many TFs differentially bound between cancer and normal better than the other network perturbation detection methods (LNS and D-score) and ANOVA.
Considering hypersensitive sites identified by DNase-seq data as the indication of “general” binding of TFs or other DNA-associated proteins, we assume that a gene is differentially bound if there is a DNase signal within a 150bp window around its TSS in one condition (cancer or normal), but not in the other condition. We observe that the DISCERN scores of the genes that are differentially bound are significantly higher than those of the genes that are not (Fig 5E). These results suggest that DISCERN identifies possible regulatory mechanisms underlying network perturbation more accurately than existing network perturbation detection methods (LNS and D-Score) and a method for identifying differential expression levels (ANOVA).
As a specific example, STAT3 has been shown to differentially regulate the mRNA expression of BATF in myeloid leukemia but not in normal condition [87]. We found that STAT3 differentially binds to BATF in the AML cell line but not in the normal cell line based on our differential binding analysis using the DNase-seq/motif data, as described above (S4 Table). Interestingly, DISCERN identifies BATF as a perturbed gene in AML (FDR corrected p-value < 0.05). DISCERN also identifies STAT3 as the strongest regulator for BATF in AML expression data, but STAT3 is not selected as an expression regulator in normal expression data (S1 File). Interestingly, LNS and D-Score detect STAT3 as an expression regulator of BATF in both conditions, not as a differential expression regulator.
Two of the Tier 1 ENCODE cell lines—K562 (chronic myeloid leukemia cell line) and GM12878 (a lymphoblastoid cell line)—correspond to blood cancer and normal tissues as well [88]. Tier 1 data contain the largest number of TFs with ChIP-seq datasets, which allows us to perform this kind of analyses using ChIP-seq datasets for these TFs. We repeated the same analysis with these cell lines and showed similar results (see S4 and S5 Figs).
Additionally, we investigated whether one can use DISCERN as a filtering step to increase the power in a pathway enrichment analysis. We consider hypersensitive sites identified by DNase-seq data as the indication of “general” binding of TFs or other DNA-associated proteins, and important regulatory events. As describe above, we identified differentially regulated genes between AML and normal cell lines (NB4/ CD34+) by identifying gene that have DNase-seq peaks within 150bp around the TSS in one condition (cancer or normal), but not in the other condition. There are 3,394 differentially regulated genes selected based on the DNase-seq data, of which 339 are significant DISCERN genes (S1 File). Presumably, these disease specific targets should be enriched for pathways or categories that will help us understand mechanisms underlying the disease. Alternatively, some targets may be spurious, especially considering the use of cell lines that are not a perfect match to healthy and diseased bone marrow samples and experimental noise.
Here we attempt to identify differentially regulated genes between AML and normal samples, by integrating the information on the DNase-seq data (i.e., differentially bound genes) and significantly perturbed genes identified by DISCERN based on the expression datasets from AML samples and normal non-leukemic bone marrow samples. To show that combining these two pieces of information helps us to identify pathways that are specifically active in one condition not in the other, we compared the significance of the enrichment for Reactome pathways measured in fold enrichment between 1) 339 differentially bound DISCERN genes (intersection of 3,394 differentially bound genes and high DISCERN-scoring genes), and 2) 3,394 differentially bound genes. S6 Fig shows that for most of the pathways, using the intersection of differentially bound and perturbed genes increases the fold enrichment compared to when differentially bound genes were used (Wilcox p-value <7 × 10−5).
Among the pathways, ‘platelet activation signalling and aggregation’ shows significant improvement in fold enrichment: 1) when differentially bound DISCERN genes were used (f = 2.9; FDR q-value = 0.01), compared to 2) when differentially bound genes were used (f = 1.03). It has been shown that the interactions between platelets and AML cells have considerable effects on metastasis, and the various platelet abnormalities have been observed in AML and other leukemias [89]. G-alpha signalling-related pathways also show significant boost in fold enrichment when DISCERN was used as a filtering mechanism for differentially bound genes. ‘Gq signalling pathway’ shows significant increase in fold enrichment: 1) when differentially bound DISCERN genes were used (f = 2.16; FDR q-value = 0.05), compared to 2) when differentially bound genes were used (f = 0.92). ‘G12/13 signalling pathway’ shows significant improvement in fold enrichment: 1) when differentially bound DISCERN genes were used (f = 3.4; q-value <0.03), compared to 2) when differentially bound genes were used (f = 1.5). These pathways have been implicated in leukemias [90].
We present a general computational framework for identifying the perturbed genes, i.e., genes whose network connections with other genes are significantly different across conditions, and tested the identified genes with statistical and biological benchmarks on multiple human cancers. Our method outperforms existing alternatives, such as LNS, D-score, and PLSNet, based on synthetic data experiments and through biological validation performed using seven distinct cancer genome-wide gene expression datasets, gathered on five different platforms and spanning three different cancer types—AML, breast cancer and lung cancer. We demonstrated that DISCERN is better than other methods for identifying network-perturbation in terms of identifying genes known to be or potentially important in cancer, as well as genes that are subject to differential binding of transcription factor according to the ENCODE DNase-seq data. We also demonstrated a method to use DISCERN scores to boost signal in the enrichment test of targets of differential regulation constructed using DNase-seq data available through the ENCODE Project.
Raw cell intensity files (CEL) for gene expression data in AML1, AML2, and AML3 were retrieved from GEO [91] and The Cancer Genome Atlas (TCGA). Expression data were then processed using MAS5 normalization with the Affy Bioconductor package [92], and mapped to Enztrez gene annotations [93] using custom chip definition files (CDF) [94], and batch-effect corrected using ComBat [95] implemented in package sva from CRAN.
BRC1 expression data were accessed through Broad Firehose pipeline (build 2013042100). We checked whether BRC1 processed by Firehose shows evidence of batch effects. We confirmed that the first three principal components are not significantly associated with the plate number (which we assumed to be a batch variable), which indicates no strong evidence of batch effects. BRC2 and BRC3 were accessed through Synapse (syn1688369, syn1688370). All probes were then filtered and mapped using the illuminaHumanv3.db Bioconductor package [96]. Probes mapped into the same genes were then collapsed by averaging if the probes being averaged were significantly correlated (Pearson’s correlation coefficient greater than 0.7).
LUAD1 expression data were accessed through Broad Firehose pipeline (build 2015110100). Genes which had a very weak signal were filtered out of the LUAD1 data. We then applied the voom normalization method that is specifically designed to adjust for the poor estimate of variance in count data, especially for genes with low counts [51]. The voom algorithm adjusts for this variance by estimating precision weights designed to adjust for the increased variance of observations of genes with low counts. This would stabilize the estimated distribution of RSEM values in the LUAD data, making it more normally distributed. Since LUAD data comes from different tissue source sites, we have applied batch-effect correction using ComBat.
For all datasets, only probes that are mapped into genes that have Entrez gene names were considered. Table 1 shows the number of samples and genes used in each dataset. For AML1, BRC1, and LUAD1 that were used for score computationa, we splitted each dataset into two matrices, one with only cancerous patients and one with normal patients. These matrices are normalized to 0-mean, unit-variance gene expression levels for each gene, before each network perturbation score (DISCERN, LNS, and D-score) was computed, which is a standard normalization step for accurately measuring the difference in the network connectivity. For methods that measure the differential expression levels (ANOVA), such normalization was not applied.
Lastly, candidate regulators are identified from a set of 3,545 genes known to be transcription factors, chromatin modifies, or perform other regulatory activity, which have been used in many studies on learning a gene network from high-dimensional expression data [43–46] (S1 Table).
DISCERN uses a likelihood-based scoring function that measures for each gene how much likely the gene is differently connected with other genes in the inferred network between two conditions (e.g., cancer and normal). We model each gene’s expression level based on a sparse linear model. Let y i ( s ) be a standardized expression levels of gene i in an individual with a condition s (cancer or normal) modeled as: yi(s)≈∑r=1pwir(s)xr(s), where x 1 ( s ) , … x p ( s ) denote standardized expression levels of candidate regulator genes in a condition s. Standardization is a standard practice of normalizing expression levels of each gene to be mean zero and unit stadard deviation before applying penalized regression method [47–50]. To estimate weight vector w i ( s ) lasso [47] optimizes the following objective function: argmin w i 1 ( s ) , … , w i p ( s ) ∑ j = 1 n ( y i j ( s ) − ∑ r = 1 p w i r ( s ) x j r ( s ) ) 2 + λ ∑ r = 1 p | w i r ( s ) | , where the subscript j in the formula iterates over all patients, used as training instances for lasso. Here, y i j ( s ) corresponds to the expression level of the ith gene in the jth patient in the sth state and x i j ( s ) similarly corresponds to the expression level of the ith regulator in the jth patient in the sth state. The second term, the L1 penalty function, will zero out many irrelevant regulators for a given gene, because it is known to induce sparsity in solution [47].
After estimating w i ( s ) for each s, the DISCERN score measures how well each weight vector learned on one condition explains the data in the other condition, by using a novel model selection criteria defined as:
DISCERN i = per sample -log-likelihood based on w i ( s ) on data in the other condition s′ per sample -log-likelihood based on w i ( s ) on data in the same condition s = e r r i ( c , n ) + e r r i ( n , c ) e r r i ( c , c ) + e r r i ( n , n ) , (1)
where e r r i ( s , s ′ ) = 1 n s | | y i ( s ) - ∑ r = 1 p w i r ( s ′ ) x r ( s ) | | 2 2. Here ns is the number of samples in the data from condition s. The numerator in Eq (1) measures the error of predicting gene i’s expression levels in cancer (normal) based on the weights learned in normal (cancer). If gene i has different sets of regulators between cancer and normal, it would have a high DISCERN score. The denominator plays an important role as a normalization factor. To show that, we defined an alternative score, namely the D0 score that uses only the numerator of the DISCERN score, Eq (1):
Di0=erri(c,n)+erri(n,c)=1nc| | yi(c)−∑r=1pwir(n)xr(c) ||22+1nn || yi(n)−∑r=1pwir(c)xr(n) ||22 (2)
The first step of calculating the DISCERN score and D0 score is to fit a sparse linear model (such as lasso [47]) for each gene’s expression level. We used the scikit-learn Python package (version 0.14.1) to calculate these scores with the values of the sparsity tuning parameters λ chosen by using the 5-fold cross-validation tests.
Analysis of Variance (ANOVA) is a standard statistical technique to measure the statistical significance of the difference in mean between two or more groups of numbers. For each gene, the 1-way ANOVA test produces a p-value from the F-test, which measures how significantly its expression level is different between conditions (e.g., cancer and normal). The ANOVA score was computed as negative logarithm of a p-value, obtained from 1-way ANOVA test using f_oneway function in scipy.stats Python package.
PLSNet score attempts to measure how likely each gene is differently connected with other genes between conditions. It was computed using dna R package version 0.2_1 [34]. The network perturbation score for each gene is computed based on the empirical p-value from 1,000 permutation tests.
In Guan et al. (2013) [35], the authors defined the local network similarity (LNS) score for gene i that is defined as correlation of the Fisher’s z-transformed correlation coefficients between expression of gene i and all other genes between two conditions:
L N S i = corr ( arctanh ( c i j n ) , arctanh ( c i j c ) ) , (3)
where c i j s represents the correlation coefficient between expression levels of genes i and j in condition s = n for normal and s = c for cancer.
For synthetic data analysis, we have also introduced a D-score, computed as following (as used in Wang et al. (2009) [36]):
D i = ‖ d i j n - d i j c ‖ 1 , (4)
where d i j s is a normalized correlation (normalized to have zero mean and unit variance across genes) between genes i and j in condition s, also known as Glass’ d score [97].
We generated 100 pairs of datasets, each representing disease and normal conditions. Each pair of datasets contains 100 variables drawn from the multivariate normal distribution with zero mean and covariance matrices Σ1 and Σ2. Each dataset contains n1 and n2 samples, respectively, where n1 is randomly selected from uniform distribution between 100 and 110, and n2 is from uniform distribution between 16 and 26. This difference in n1 and n2 reflects the ratio of the cancer samples and normal samples in the gene expression data (Table 1).
For each of the 100 pairs of datasets, we divided 100 variables into the following three categories: 1) variables that have different sets of edge weights with other variables across two conditions, 2) variables that have exactly the same sets of edge weights with each other across the conditions, and 3) variables not connected with any other variables in the categories 2) and 3) in both conditions. For example, in Fig 1A, ‘1’ is in category 1) (i.e., perturbed genes). ‘2’, ‘4’, ‘6’, and ‘7’ are in category 2), and ‘3’ and ‘5’ is in category 3). In each of the 100 pairs of datasets, the number of genes in category #1 (perturbed genes), p, is randomly selected from uniform distribution between 5 and 15. The number of genes in each of the other two categories #2 and #3 is determined as (100 − p)/2.
We describe below how we generated the network edge weights (i.e., elements of Σ 1 - 1 and Σ 2 - 1) among the 100 variables. To ensure that only the genes in #1 have differing edge weights between two conditions, we generated two p × p matrices, X1 and X2, with elements randomly drawn from a uniform distribution between -1 and 1. Then, we generated symmetric matrices, X 1 ⊺ X 1 and X 2 ⊺ X 2, and added positive values to the diagonal elements to these symmetric matrices, if its minimum eigenvalue is negative—a commonly used method to generate positive definite matrices [98]. They become submatrices of Σ 1 - 1 and Σ 2 - 1 for these p variables. Similarly, we generate a common submatrix for the variables in category #2—variables that have the same edge weights with other variables across conditions. Variables in category #3 have identity matrix as the inverse covariance matrix among the variables in that categories. Finally, we added mean zero Gaussian noise to each element of Σ 1 - 1 and Σ 2 - 1, where the standard deviation of the Gaussian noise is randomly selected between 0.5 and 5.
This procedure allows having datasets of varying levels of difficulty in terms of high-dimensionality and network perturbation, which provides an opportunity to compare the average performances of the methods in various settings.
To generate a conservative null distribution, we performed permutation tests by randomly reassigning cancer/normal labels to each sample, preserving the total numbers of cancer/normal samples. The correlation structure among genes would be preserved, because every gene is assigned the same permuted label in each permutation test. We then computed the DISCERN score for a random subset of 300 genes. We repeated this process to get over one million DISCERN scores to form a stable null distribution, which was used to compute empirical p-values.
For the survival-associated genes enrichment analysis, we first computed the association between survival time and each gene expression level. Genes that had a p-value from the Cox proportional hazards model (computed using survival R package) smaller than 0.01 were considered significantly associated with survival. These include 1,280 genes (AML), 1,891 genes (BRC) and 1,273 genes (LUAD) (S3 Table). Statistical significance of the overlap with top N DISCERN, LNS, D-score and ANOVA -scoring genes was computed by using the Fisher’s exact test based on the hypergeometric distribution function from scipy.stats Python package [99].
We presented the results on the comparison with three sets of genes that are known to be important in cancer (S2 Fig). Here, we describe how we obtained these gene sets. First, Malacards genesets were constructed based on the data from malacards.org website accessed in September 2012. Second, we used a set of 488 genes we downloaded from Catalogue of Somatic Mutations in Cancer website (CGC) [82]. For each cancer type, we considered the intersection between this list and the genes that are present in the expression data. Finally, a set of genes likely to contain driver mutations selected by MutSig was defined as those that pass q < 0.5 threshold based on 20141017 MutSig2.0 report from Broad Firehose.
To evaluate the performance of the DISCERN score on identifying genes to be used in a prognosis prediction model, we trained the survival prediction model using one dataset and tested the model on an independent dataset (Fig 4). To train the survival prediction model, we used the elastic net regression (α = 0.5) using glmnet CRAN package (version 1.9-8). Available clinical covariates—age for AML, and age, grade and subtype for BRC—were added as unpenalized covariates. Regularization parameter λ was chosen by using the built-in cross-validation function. Testing was always performed in the independent dataset with held-out samples from the dataset that was not used for training. For comparison, we trained the prediction model using 22 LSC genes [54] with age in AML, and 67 genes from the 70-gene-signature [83] (3 genes from the signature were missing in the dataset we were using) with clinical covariates (age, stage, and subtype) in BRC, as shown in Fig 4B and 4D, respectively.
The Encyclopedia of DNA Elements (ENCODE) is an international collaboration providing transcription factor binding and histone modification data in hundreds of different cell lines [100]. Data for ENCODE analysis were accessed through the UCSC Genome Browser data matrix [101] and processed using the BedTools and pybedtools packages [102, 103]. Two of the ENCODE cell lines—NB4 (an AML subtype [84]) and CD34+ (mobilized CD34 positive hematopoietic progenitor cells)—are closest to AML and normal conditions, and the DNase-seq data from these cell lines are available.
For each cell line, we used the DNase-seq data and the position weight matrices (PWMs) of 57 transcription factors (TFs) available in the JASPAR database [85] to find the locations of the PWM motifs that are on the hypersensitive regions. We identified the locations of these PWM motifs on the hg38 assembly by using FIMO [86] (p-value ≤10−5). We then intersected these motif locations with hypersensitive regions identified by the DNase-seq data for each TF. We repeated this process to identify active binding motifs of the 57 TFs in each of the cell lines, NB4 and CD34+.
For each TF, we identified the genes the TF differentialy binds to between cancer and normal cell lines. We assumed that a certain TF is bound near a gene if the center of the peak is in the active enhancer regions (marked by H3K27Ac) within 15kbs of the transcription start site (TSS) of the gene or the 5kb around the gene’s transcription start site. We show that for most of the TFs, differentially bound genes have significantly high DISCERN scores than those not (Fig 5A–5C).
The differential regulator score for each gene was computed by taking the number of differentially bound TFs and dividing it by the total number of TFs bound to the gene in any condition. We show that the differential regulator score is highly correlated with the DISCERN score (Fig 5D). For DNase-based analysis (Fig 5E), we defined a gene to be differentially regulated if hypersensitive sites detected by DNase-seq are within 150bp upstream of the gene in one condition and not in another.
A set of 605 Reactome pathways was downloaded through Broad Molecular Signature Database (MSigDB) [59]. We postulate that hypersensitive sites identified by DNase-seq in a particular cell line indicate the regions where important regulatory events occur, such as transcription factor binding. We constructed the list of differentially regulated genes by comparing the hypersensitive sites identified by DNase-seq data between cancer and normal cell lines within 150bp upstream from TSS of each gene. For each pathway, we computed the fold enrichment (= number of genes in the intersection of two groups of genes number of genes in the intersection by random chance) that measures the significance of the overlap between genes in the pathway and the identified differentially regulated genes. We compared the fold enrichment with when the genes in the intersection of differentially regulated genes and 1,351 significantly perturbed genes identified by DISCERN were used (S6 Fig). To reduce the noise, we only considered the pathways that had ≥5 genes in the overlap before filtering. The p-values were then FDR corrected for multiple hypothesis testing. Although p-values would measure the significance of the overlap between a gene set with a pathway, we used the enrichment fold as a measure of the significance of the overlap because we compared a set of genes with another set much smaller size.
|
10.1371/journal.pntd.0004870 | Insights into an Optimization of Plasmodium vivax Sal-1 In Vitro Culture: The Aotus Primate Model | Malaria is one of the most significant tropical diseases, and of the Plasmodium species that cause human malaria, P. vivax is the most geographically widespread. However, P. vivax remains a relatively neglected human parasite since research is typically limited to laboratories with direct access to parasite isolates from endemic field settings or from non-human primate models. This restricted research capacity is in large part due to the lack of a continuous P. vivax in vitro culture system, which has hampered the ability for experimental research needed to gain biological knowledge and develop new therapies. Consequently, efforts to establish a long-term P. vivax culture system are confounded by our poor knowledge of the preferred host cell and essential nutrients needed for in vitro propagation. Reliance on very heterogeneous P. vivax field isolates makes it difficult to benchmark parasite characteristics and further complicates development of a robust and reliable culture method. In an effort to eliminate parasite variability as a complication, we used a well-defined Aotus-adapted P. vivax Sal-1 strain to empirically evaluate different short-term in vitro culture conditions and compare them with previous reported attempts at P. vivax in vitro culture Most importantly, we suggest that reticulocyte enrichment methods affect invasion efficiency and we identify stabilized forms of nutrients that appear beneficial for parasite growth, indicating that P. vivax may be extremely sensitive to waste products. Leuko-depletion methods did not significantly affect parasite development. Formatting changes such as shaking and static cultures did not seem to have a major impact while; in contrast, the starting haematocrit affected both parasite invasion and growth. These results support the continued use of Aotus-adapted Sal-1 for development of P. vivax laboratory methods; however, further experiments are needed to optimize culture conditions to support long-term parasite development.
| Plasmodium vivax has a tremendous impact on public health; accounting for 13.8 million cases of clinical illness estimated in 2015, causing a wide spectrum of symptoms including severe disease. Development of new therapies requires a better understanding of the parasite’s biology, however, understanding the fundamental biological properties of P. vivax is challenging as there currently is no robust in vitro blood-stage culture system available. Unfortunately, this lack of understanding of the parasite’s basic biology, especially understanding the main variables that control blood-stage development such as nutrient-dependence, preferred host cell age, and successful in vitro format, is a major hurdle to the establishment of long-term in vitro P. vivax culture that is needed for basic and clinical research. In our present study, we used the P. vivax primate-adapted strain, Sal-1, obtained from Aotus lemurinus lemurinus, and investigated a set of the variables that control blood-stage development described in the literature. We specifically focused on addressing important parameters: host cell invasion, maturation within the invaded host cell, and egress and re-invasion into new host cells. All of which are critical for robust blood-stage growth and development.
| Plasmodium vivax is the most geographically widespread human malaria parasite causing 13.8 million clinical cases every year [1]. While P. vivax was once considered the benign malaria, it can cause severe disease and death [2,3]. P. vivax presents unique challenges compared to other human malaria parasites. P. vivax forms a dormant and relapsing stage in the liver (hypnozoites) and the transmissible stages (gametocytes) are found in peripheral circulation prior to the appearance of clinical symptoms [4], making it especially difficult to interrupt transmission. Despite the large burden of disease, P. vivax has long been neglected largely due to the lack of an in vitro culture system for the parasite. Recent calls for worldwide malaria eradication have placed new emphasis on the importance of addressing P. vivax as a major public health problem [5].
Long-term in vitro culture of P. vivax has been undermined by its inability to invade mature red blood cells (RBCs), as it is restricted to immature RBCs (reticulocytes) [6], which represent as little as 0.5%-1.5% in normal blood. Reticulocytes develop in the bone marrow and are released into the peripheral circulation where they continue to mature into RBCs. Enrichment of peripheral blood have failed to support long-term in vitro propagation, most probably due to donor variability and age of the reticulocyte [7]. Additionally, reticulocytes rapidly mature to RBCs in culture conditions [8,9], thus requiring continual replenishment, which leads to dilution of the culture. Furthermore, besides the host cell, variability exists in the parasite as well [10]. Thus, the heterogeneity in human P. vivax patient isolates has made it difficult to establish benchmark parasite characteristics such as specific host cell age, media and in vitro culture formatting preferences.
Parasite growth is a product of three main variables: the ability to invade the host cell, the ability to mature within the invaded host cell and the ability to egress and re-invade new host cells. An in depth understanding the ways to successfully support each of these three parameters is necessary to address the parasites’ exponential decrease observed in in vitro culture conditions that has represented the major hurdle to the establishment of long-term in vitro P. vivax culture [11].
Therefore, to overcome parasite variability and establish benchmark conditions to improve the development of a robust and reproducible in vitro P. vivax culture, we set out to perform head-to-head experimental comparisons for a selected number of culture conditions using the P. vivax primate-adapted strain Sal-1 obtained from Aotus lemurinus lemurinus. Aware of the ethical restrictions (number of animals (3Rs principle of reduction); bleeding volume) and of the fact that Aotus is a scarce resource, we designed the study in such a way as to have the minimum amount of animals infected and bleedings required to obtain statistical significance in terms of biological replicas for the most important variables tested. Technical replicates (duplicates or triplicates depending on the amount of blood available) were plated for all experiments and variability was generally low between both biological and technical replicates. Due to the restrictions reported above, we had to be selective on the number of variables to include in the present study, we reviewed literature on short-term in vitro culture [11, 12, 13, 14 and references therein] to assess which variables appeared most critical to the development of a successful in vitro culture. One of the most successful attempts using an Aotus-adapted P. vivax strain (Chesson) was that of Golenda, et al in 1997 [14], which reported successful doubling at each generation in vitro. Thus, we used the apparently successful methodology described in [14] as a framework and starting point for our study.
To understand host cell requirements [11, 12, 13, 14], we tested different sources of reticulocytes and procedures for their enrichment to assess their effect on invasion efficiency. We then compared different methods of leuko-depletion to determine if there was an effect on the parasite. Furthermore, we utilized stabilized forms of nutrients to see if they could support growth. Finally, we tested different culture formatting conditions such as shaking versus static formats and altering the starting haematocrit.
The experimental protocol was approved by the ICGES Institutional Laboratory Animal Care and Use Committee (CIUCAL) in accordance with procedures described in the “Guide for the Care and Use of Laboratory Animals,” 1996, the International Guiding Principles for Biomedical Organizations of Medical Sciences (CIOMS) and the laws of the Republic of Panama; protocol approval number 2013/06. Animals were housed at Gorgas Memorial Institute of Health Studies (ICGES) in Panama. The animals were kept in climate control rooms with 12 air changes per hour (min/max temperature set to 21/25°C and 70–80% humidity) and a red/white fluorescent 12-hour cycle starting at 03:00 pm. They were kept in pairs (male and female), in stainless steel 4 unit quads cages (Lab Products Inc., Seaford, DE, USA) with dimensions of 27 x 23.5 x 29.5 inches. Each cage was fitted with a 3⁄4-inch-diameter PVC pipe perch placed across 2/3 of the length of the cage and a 6-inch-diameter x 14.5 inches long PVC T pipe nest-box. Cages were routinely cleaned and sterilized at 180° F weekly. Diet consisted of fresh fruit, cooked potatoes, cooked meat rice, pudding (to stimulate foraging behavior and well being), and a vitamin supplement diluted in an orange juice concentrate and mixed with wheat germ bran and sugar. 3x/week animals received Monkey Chow (New World Primate Diet 5040, LabDiet; PMI Nutrition International, LLC, Brenhwood, MO, USA) and fresh fruit. Water was administered daily in plastic bottles fitted with a zip tube (Girton; Millville, PA, USA). During routine bleeding procedures the animals were sedated with Ketamine at a dosage of 10 mg/Kg IM. No euthanasia was carried out during these experiments and once the animals were radically cured of malaria with mefloquine at 20 mg/kg orally once, they were transferred to the reproductive colony. All animals received daily veterinary care.
For the donor inoculum, female laboratory-bred and spleen intact Aotus lemurinus lemurinus monkeys, weighing between 789–850 g were infected with Aotus-adapted P. vivax Sal-1 and parasites were obtained after bleeding at a majority of ring stages. Parasites were cryopreserved with Glycerolyte (Baxter) following the Methods in Malaria Research [15]. The animals were housed at Gorgas Memorial Institute of Health Studies in Panama City, Republic of Panama and cared and maintained as described [16], in accordance with the reviewed and approved protocol entitled “Production of Aotus Plasmodium vivax SAL-1 and AMRU-1 infected blood for continuous in vitro culture attempts” submitted in 2013, registered in the ICGES-Institutional Animal Care and use Committee (CIUCAL-ICGES) under the accession number 2013/06 and following the criteria set forth in the International Guiding Principles for Biomedical Organizations of Medical Sciences (CIOMS) and the laws of the Republic of Panama.
Experimental monkeys were infected with parasites through the saphenous vein. Each parasite inoculum contained 5.4x106 P. vivax Sal-1-parasitized erythrocytes in a volume of 1 mL. Parasitemia was monitored daily following the Earle and Perez method [17] with Giemsa (Sigma) stained thick blood smears using 5 μL of blood collected by gently pricking the outer ear vein of the animals. Once the parasitemia had reached a peak, the different animals were bled. A total of 3mL (1–1.5 mL packed cells) was collected. To maximize the number of biological replicates while still limiting the number of animals we infected, we bled some monkeys twice (MN23026 and MN23009) but we took less blood at each draw so the total amount of blood collected was 3 mL. Immediately, following bleeding, the animals were radically cured with mefloquine orally at 20 mg/kg once as described [16]. At each bleeding thin Giemsa blood smears were prepared to obtain precise parasite counts and perform asexual (rings, trophozoites and schizonts) and sexual (gametocytes) staging before parasites were put in culture to test different culture conditions as described throughout the manuscript. The schematic describing the passaging is in Fig 1.
To evaluate the age of the reticulocytes that contained parasites, we used a double staining method, which simultaneously stains the reticulocyte and the parasite. 5 μL of packed infected cells were washed with PBS, mixed with an equal volume of new methylene blue (NMB) (Sigma) and allowed to stain for 15 minutes. Following the staining, the cells were smeared on a glass slide (Falcon) and air-dried. Slides were then methanol fixed and stained with Giemsa. In the resulting slides both the reticulum of the reticulocytes and the nuclear and cytoplasmic material of the parasites were stained (see S1 Fig, Fig 2).
Fresh reticulocyte preparations were obtained from Duffy antigen positive hemochromatosis blood (Brigham and Women’s Hospital and The German Red Cross Blood Bank) and Buffy packs (The Interstate Blood Bank INC. Memphis, USA). Three reticulocyte enrichment procedures were used: (i) a modified differential centrifugation method used in [14] (ii) a Percoll gradient [18] and (iii) a CD71+ immuno-magnetic purification method (Miltenyi, manufacturer’s specification).
Frozen enriched reticulocytes preparations were obtained from Duffy antigen positive hemochromatosis blood collected by the German Red Cross Blood bank using either (i) differential centrifugation or (ii) a Percoll gradient as described above. After enrichment reticulocytes were frozen with homologous plasma and 10% DMSO in cryotubes. Except for the frozen cells, reticulocytes were enriched and stored at 4°C and used within one week of preparation.
Two different leuko-depletion methods were compared: (i) CF11 columns and (ii) Plasmodipur filters. For all experiments, Giemsa stained smears were made before and after leuko-depletion to assess parasitemia and stage and health of the parasites. Parasitemia was determined by counting the number of parasites per 10,000–20,000 RBCs per slide as presented in Fig 4.
Throughout the study, we used two methods to standardize and normalize our measurements so we could compare different biological replicates. At each time-point, slides were made and Giemsa-smeared to evaluate parasitemia, stage and health of the parasites. As parasites were drawn at roughly synchronous stages (e.g. mostly rings or mostly trophozoites) and progressed in vitro synchronously, we were able to calculate conversions between different stages. Conversion was determined by staging the parasites at each time-point and multiplying the percent of each stage by the parasitemia (initial rings or trophozoites). The numbers of parasites were counted and percentages were calculated in the same way (resulting trophozoites or schizonts). The percent conversion was the calculated by taking the resulting parasitemia divided by the initial parasitemia. For invasion, we used a normalized measurement of the final ring parasitemia called parasitized erythrocyte multiplication rate (PEMR) [18]. Here, the final number of rings is divided by the input number of schizonts. Our PEMR calculations include only healthy rings with intact nuclei and ring cytoplasm on smears and do not include pyknotic cells.
As Golenda et al, 1997 was the only study to see doubling over multiple generations, we followed the formatting and media conditions that were used [14]. Therefore, unless specified, for all experiments, we cultured parasites in 4-well plates from Nunclon (Sigma D6789-1CS, Nunc176740) at 6% haematocrit (30μL of iRBCs in 500 μL of media). Unless otherwise stated, media for all experiments was prepared as follows: McCoy’s 5A Modified medium (Gibco) was supplemented with 25 mM HEPES (Sigma), 2 g/mL sodium bicarbonate (Sigma), 2 g/L D-glucose (Sigma) and 40 μg/mL gentamycin (Sigma). Media was supplemented always with 20% AB+ heat inactivated serum (The Interstate Blood Bank, Memphis, TN, USA). A candle jar was used to maintain the gas conditions of the 4-well plates (~18% O2). Cultures were maintained at 37°C. For invasion, once the majority of parasites reached the schizont stage, uninfected, human density-enriched reticulocytes were added at a 1:1 ratio doubling the haematocrit to 12% and after re-invasion cultures were diluted back to 6% haematocrit [14].
To determine the effect of stabilized nutrient components, 100X GlutaMAX (Gibco) was supplemented to a final 1X (Fig 5).
For the cultures in which we wanted to compare shaking versus static culturing formats, we had glass flasks specially made to the specifications of the T flasks that were used in 1997 [14]. Customized glass Erlenmeyer 5 mL and 10 mL funnels and type T funnel fuzzier cock and sintered Type T culture cylinder were manufactured by Ritmester B.V (Fig 6A). Flasks were gassed using standard parasite gas mixture (5%O2, 5%CO2, 90%N2) by attaching a specialized tube to the fuzzier cock. Shaking was done at 100 rpm on an orbital shaker (Thermo). Although with different volumes of media, haematocrit was respected to a 6%. These conditions mostly closely replicate [14].
When culture outcomes where studied regarding the starting hct, this was varied from standard 6% to 12% and 15%. Cultures were initiated at the indicated hcts using the blood directly from the draw and diluted to the appropriate hct with complete media (iRBC). We also tested a culture at 12% hct that was mixed 6% from the draw and doubled with uninfected blood (iRBC + RBC) (Fig 7).
Statistical analyses were performed using Graph Pad Prism (5.1).
Four adult Aotus lemurinus lemurinus monkeys (MN21014, MN28016, MN26032 and MN23009) served as donors of P. vivax-infected reticulocytes for this experiment after being sub-passaged from MN28010, MN29010 and MN31007 (Fig 1). To maximize the number of biological replicates for the media and formatting experiments, we bled MN26032 and MN23009 twice. While this limited the amount of starting blood for each condition, it maximized the replicates without having to infect additional animals. Their parasitemia was monitored daily by Giemsa stained thick blood smear. For the reticulocyte preference and invasion assays, we needed larger amounts of blood so we only bled MN21014 and MN28016 once (3mL) to maximize the blood volume. Here we are using the same batch of frozen parasite stocks for all monkey infections, as conventionally done in monkey ex-vivo drug studies, where generally infected blood from one single animal is used. When planning these experiments, we attempted to be as thorough as possible while keeping sight of the important ethical boundaries established for the work with non-human primates (3Rs principle of reduction) and of the fact that Aotus is a scarce resource. We have designed the study in such a way as to use the minimum amount of animals/bleedings required to obtain statistical significance in terms of biological replicas for the most important variables tested. Technical replicates (duplicates or triplicates depending on the amount of blood available) were plated for all experiments; variability was generally low.
Careful parasite staging was carried out on all smears as a way to strictly characterize the starting population of P. vivax-infected reticulocytes. A careful staging clearly demonstrates that the synchronicity of the parasite populations differs between monkeys and between blood draws and that it is therefore important to refer each dataset to its original population when invasion or maturation and growth are considered. A table is provided detailing the experiments each bleed (biological replicate) was used for (Fig 1) and the results from experiments where we had biological replicates are reported in the main text. Due to the fact that Aotus is scarce resource and widely accepted ethical constraints apply to this research, the number of animals was restricted to the minimum required to have enough biological replicates to test a selected number of variables. As we sometimes had small amounts of leftover infected blood, we considered it unethical to discard any material and we therefore used the excess to look into other variables of interest identified in the literature [11, 12, 13, 14], the results of which are shown in Supplemental Materials.
To identify the age of reticulocytes that best supports P. vivax invasion and growth, we surveyed the parasites coming directly from the bleed (ex vivo). Heilmeyer classification [19] based on New Methylene Blue (NMB) staining can be used to determine the age of the reticulocyte [10]. In four biological replicates, we found the youngest reticulocytes contain large amounts of RNA (reticulum) and stain most intensely (Heilmeyer I) while old reticulocytes will have very week staining (Heilmeyer IV) (Fig 2A and S1 Fig). Mature RBCs do not contain RNA and will not stain with NMB (no Heilmeyer classification). Double-staining with NMB and Giemsa revealed that ring stage samples isolated directly from Aotus showed 30% of the parasites being hosted inside strongly NMB stained cells with a very dense reticulum network (Heilmeyer I), 55% inside weakly NMB stained cells with a mild reticulum network (Heilmeyer II-III) and only 15% of the host cells presented either a weak reticulum (Heilmeyer IV) (Fig 2B and S1 Fig). Although further experiments would be required to demonstrate this finding, it suggests that the parasites preferentially invade younger reticulocytes in vivo in the Aotus model. When more mature samples predominantly containing trophozoites were obtained directly from the Aotus, the 30% of the parasites were still found in the strongly stained Heilmeyer II-III while the rest of the parasites were found in Heilmeyer IV reticulocytes (62%) and only 8% of the parasites were found in cells lacking NMB (mature RBCs) (Fig 2B and S1 Fig), indicating that if merozoites preferentially invade young cells, reticulocyte maturation of P. vivax infected cells progresses in vivo. Interestingly, in vitro data followed a somewhat different pattern where second-generation newly invaded rings (20 hr in vitro) were found 25% inside Heilmeyer I, 25% in Heilmeyer II-III and 50% in Heilmeyer IV (Fig 2B). When second generation trophozoites (20 hr in vitro from a more mature sample) were studied: 50% were inside Heilmeyer II-III and 50% in Heilmeyer IV. Whether this is due to reticulocytes maturing in vitro prior to parasite invasion or the parasite driving accelerated maturation of the reticulocyte in vitro is unclear.
Given the findings of the ex vivo staining and a recent report indicating reticulocyte preparation method can affect the success of invasion [20], we sought to understand if the reticulocyte source (blood from routine phlebotomies or Buffy packs) and enrichment method affected parasite invasion. In this experiment we performed a head-to-head comparison of invasion rates with reticulocytes from different sources enriched following the different procedures presented in the Material and Methods (density centrifugation, Percoll enrichment, immunomagnetic separation (Fig 3A) and frozen versus fresh reticulocytes (S2)). In two independent biological replicates (MN21014 and MN28016) and two separate experiments with technical replicates, we calculated the normalized invasion by PEMR (Materials and Methods, Fig 3B). Dunnett’s multiple comparison showed that P. vivax demonstrated a significant preference for fresh hemochromatosis blood derived reticulocytes enriched by differential centrifugation (Fig 3C and 3D), despite the fact that the reticulocyte numbers were lower than other methods of preparation (Fig 3A). In order to verify that human reticulocytes that were invaded as opposed to back invasion into the endogenous Aotus reticulocytes present from the original inoculum remaining in the culture, we performed flow cytometry analysis to determine the amount of P. vivax (PyBIP Alexa647) inside human cells (Glycophorin A/CD235A+) versus Aotus cells (Glycophorin A/CD235A-) (S2A and S2B Fig). BIP is an endoplasmic reticulum (ER)-targeted protein binding immunoglobulin protein whose P. falciparum version (PfBIP) reactivity against P. vivax has already been demonstrated [21].
We also tried reticulocytes from different hemochromatosis donors in Europe that had been enriched by the same methods as the fresh hemochromatosis blood but had been cryopreserved to preserve freshness in transit. Limited numbers of cells were recovered post-thaw so only one parasite biological replicate was tested (S2 Fig). Again, fresh, density enriched blood proved to provide a higher PEMR and resulting ring parasitemia (S2C and S2D Fig).
These results indicate that cryopreservation and Percoll may damage or somehow intrinsically alter reticulocytes during enrichment and preservation, thereby preventing successful invasion of the parasite. Therefore, in subsequent experiments, we decided to restrict ourselves to fresh hemochromatosis reticulocytes enriched by differential centrifugation to monitor invasion, growth and egress.
[14] used CF11 to deplete leukocytes. However, CF11 is no longer manufactured and most laboratories now use commercially purchased Plasmodipur filters or cellulose powder. Therefore, we set out to compare the affect that these two leuko-depletion methods have on parasitemia, staging and health of the parasite. In a few cases we found a correlation between the stages observed in the blood draw, the filtration procedure used and the course of parasitemia. In cases of mixed stage blood draws involving majority mature (trophozoite and schizont stage) parasites, the parasitemias post-Plasmodipur filtering was lower (sometimes, markedly so) when compared to post-CF11 parasitemias (Fig 4A). In four biological replicates, no major differences were observed in parasitemia over time, but the Plasmdipur-filtered parasites did persist for longer in one replicate (Fig 4B). Interestingly, in the four biological replicates, we found that the CF11-filtered parasites appeared compromised during invasion as indicated by a lower PEMR (Fig 4C). Although the trend was not significantly significant, as most of the observed difference was mostly contributed by a single biological replicate, MN23009a. Together these findings indicate that CF11 filtration retains less of the mature stages compared to Plasmodipur, but the mature stages present in CF11 elution may be less fit. This mixed outcome may be due to the contact with the CF-11 powder or due to the fact that CF11 allows ex vivo late stages to pass through, which may be less amenable to adapt to in vitro culture conditions than the early ex vivo stages.
We tested various preparations of media and supplements. While the parasitemias were not statically different (Fig 5A), a stabilized form of L-glutamine, GlutaMAX (Gibco), consistently supported parasite maturation along the whole parasite cycle in four independent biological replicates (Fig 5B). Greater benefits were observed in the transition from ring to trophozoite, with better conversion rates compared with McCoy’s 5A modified medium (control) alone (Fig 5C). Conversion from trophozoites to schizonts were somewhat negligible (Fig 5C), but overall, both parasites and host cells appeared healthier in the media supplemented with Glutamax as evidenced by reinvasion of schizonts in the supplemented cultures compared to the control media alone (Fig 5B). GlutaMAX is a dipeptide of L-alanyl-L-glutamine, which does not break down to form toxic byproducts such as ammonia like traditional L-glutamine. These results indicate that P. vivax could be acutely sensitive to build-up of waste or toxic products in in vitro conditions.
An important contradictory finding in the literature is the effect of shaking versus static conditions on in vitro P. vivax development. Golenda et al. [14] reported that shaking was necessary to facilitate egress and reinvasion while Mons et al. [12] reported that this actually harmed the parasites. To determine the effect of shaking, we compared both invasion and development of the parasites in shaking or static cultures. To exactly recapitulate the findings of Golenda et al. [14], we had special T type glass flasks constructed (Fig 6A). We found that the parasitemia remained higher and parasites persisted in the culture for longer when the cultures were maintained in static conditions (Fig 6B), similar to what Mons et al. reported [12]. In fact, while shaking and static cultures had similar conversion rates from rings to trophozoites (Fig 6C), shaking appeared to be somewhat only marginally beneficial to conversion from trophozoites to schizonts (Fig 6C). Despite Golenda’s findings, we determined that shaking had a negligible effect on invasion; PEMR 0.05 compared to static PEMR of 0.04 (Fig 6D). None of these results were statically significant.
In one biological replicate, MN26032a, we initiated cultures at 6% hct and 15% hct to determine if there was any effect on parasitemia (in static cultures). Interestingly, we observed a higher parasitemia that persisted for longer in vitro (S3 Fig). To examine this further, we decided to test two biological replicates, MN23009a and MN23009b, at different hcts at the initiation of the culture. We started the culture at either 6% hct or 12% hct (iRBC). To determine if the findings from S3 Fig (MN26032) were due to a higher overall parasite biomass, we also set up a culture at 12% hct but this time, we set up the culture at 6% hct and doubled the hct with uninfected blood (12% iRBC + RBC). Overall, we found that the 12% hct culture had a higher parasitemia overall, although no statically significant (Fig 7A). Interestingly, the parasites persisted for longer in vitro in the 12% hct condition (Fig 7A). Interestingly, the 12% hct iRBC + RBC provided the best conversion from rings to trophozoites but the 12% hct iRBC had the best conversion from trophozoites to schizonts (Fig 7B). Additionally, the 12% hct iRBC + RBC had a higher PEMR overall (Fig 7C). While these results are not statically significant, the trends indicate that differences in parasite biomass and formatting changes to the hct could influence the outcome of the culture.
Since Bass and Johnson’s first attempt to culture P. vivax in 1912 [22], many more unsuccessful attempts have been made to develop a continuous, long-term in vitro culture of P. vivax. To date, the most successful of such attempts has been the one by Golenda et al. [14] in 1997 in which they were able to maintain P. vivax in culture for 8 cycles, doubling the number of parasites at almost every cycle and thus maintaining a reasonably high parasitemia compared to other reports [23, 24]. Equally important the Golenda study [14] confirmed that non-human primate-adapted P. vivax (Chesson) derived from infections carried out in Aotus nancymai and Aotus lemirinus griseimenbra could be successfully adapted to in vitro culture conditions. The scope of our experiments was to use this one successful attempt and to optimize culture conditions for the Aotus lemurinus lemurinus adapted P. vivax Sal-1. These studies were designed to inform and improve standardized conditions for further testing of an in vitro culture system. To this end, we selected a wide range of conditions for testing, in part derived from literature and in part from previous findings obtained independently in our laboratories.
Two recent reports have demonstrated a preference of the parasite for young reticulocyte [25, 26]. The reticulocyte-prone rodent parasite P. yoelii, which mimics several biological features of P. vivax, has a strong tropism for the immature reticulocytes present in the bone marrow and spleen of BALB/c mice [25]. A second study found that clinical isolates of P. vivax from Thailand exhibited a strong preference for very young reticulocytes (CD71high and CD71med) in vitro [26]. Therefore, to confirm these two studies, we looked at the parasites coming directly out of the Aotus monkey (ex vivo). Like the rodent study, we observed that the majority of rings were found in young reticulocytes Heilmeyer I and II-III (Fig 2). Interestingly, most of the ex-vivo trophozoites were also found inside reticulum-positive reticulocytes (Heilmeyer II-III and Heilmeyer IV Fig 2). This is in contrast to Malleret et al. who have reported that invasion would trigger an accelerated maturation of the reticulocyte such that reticulum would be expected to be absent by the trophozoite stage [26]. We did observe, however, that the in vitro rings and trophozoites (second generation) were found in older reticulocytes and mature RBCs, which is consistent with [26]. Whether the ex vivo and in vitro discrepancies are linked either to an intrinsic difference between the non-human primate adapted strain P. vivax Sal-1 and the Thai isolates, or to an ex vivo versus in vitro effect on the infected host cells needs to be further explored. Alternatively, the reticulocytes may mature in vitro prior to invasion, which could explain why rings are found in older cells. More detailed studies are needed to fully confirm these differences but the overall conclusion that we can draw from our findings is that P. vivax rings ex vivo are found in young reticulocytes indicating a preference for invasion of the youngest reticulocytes.
As reticulocytes represent only a minor portion of whole blood samples [27], enrichment is necessary to increase the numbers of susceptible host cells for invasion in vitro. We wondered if reticulocyte preparation methods could affect the parasite invasion as specific enrichment procedures (e.g. Percoll) may alter the make up of surface cell markers thereby negatively influencing parasite invasion [13]. Therefore, we opted for a head-to-head comparison of three different enrichment methodologies: differential centrifugation, Percoll enrichment (density gradient) and a CD71+ immuno-magnetic purification method Interestingly, we found that the reticulocytes enriched by density centrifugation alone could support better invasion of the parasites suggesting that at least Percoll density gradients may negatively impact the cells overall (Fig 3). Another density gradient method [20] using Nycodenz was published after the conclusion of this study and was therefore not tested during the studies presented in this manuscript. While Percoll is made up of colloidal silica coated with polyvinylpyrrolidone, which can be toxic to cells, Nycodenz is a nonionic iodinated gradient medium, which is nontoxic. It has been shown that Percoll gradients can damage the membranes of sperm [28]. However, no side-by-side comparisons were made in [20], which would be important to consider. Taken together, however, these experiments and the report published by Roobsoong et al. [20], demonstrate that the host cell health plays an important role in the outcome of in vitro cultures.
In the context of differential centrifugation, it is interesting to note that Golenda et al. [14] utilized a laborious method in which they achieved a five-fold increase in the number of reticulocytes (from 3–5% to 15–20%) after a subset of centrifugation and incubation steps with homologous serum. We used both the protocol described in Golenda et al. [14] as well as a simplified version with centrifugations at a maximum of 3,000 xg instead of the 35,000 xg ultra-centrifugations. With both procedures, we could obtain a maximum 3-fold increase in reticulocytes compared to the starting hemochromatosis blood reticulocytemia. A contributing factor in the difference in enrichment outcomes (5-fold reported by Golenda et al. [14] and 3-fold obtained in our laboratories) may be due to different treatment procedures for patients suffering from hemochromatosis today compared to 1997 [29]. In the past, repeated bleeds were performed prior to phlebotomy, resulting in an induction of reticulocytosis, while today phlebotomies are performed directly. Thus, the reticulocyte content in blood collected in the past was higher than the reticulocytes obtained from hemochromatosis patients today. Our data suggest that a lower number of reticulocytes in the blood at the start of the differential centrifugation procedures resulted in lower enrichment rates. Additionally, the repeat phlebotomies may also have resulted in a larger pool of young reticulocytes (post-emergence from the bone marrow) which may have resulted in higher invasion efficiencies and allowed for the doubling of the parasites at each generation in the Golenda et al. study [14]. This is also consistent with the preference we observed ex vivo with rings being found in the youngest reticulocytes.
Different factors may explain the difference in replication rates and parasitemia observed in our experiments compared to Golenda et al. [14]: (a) as explained above, reduced numbers of reticulocytes in the blood used to culture P. vivax, suggesting that a reticulocytes source other than the standardly prepared hemochromatosis blood is required if the 15–20% reticulocyte enrichment levels obtained by Golenda et al [14] through differential centrifugation, and reported to be instrumental to maintenance of good P. vivax multiplication rates and parasitemia rates, are to be achieved; (b) the different P. vivax parasite strain used (Chesson [14] vs Sal-1), suggesting that multiplication rates and parasitemia may be at least in part an intrinsic characteristics of the specific P. vivax strain, (c) the karyotype of Aotus monkey used (nancymai and lemurinus griseimembra [14] vs lemurinus lemurinus) and adaptations related to these parasite-host interactions may influence the adaptability of specific P. vivax strains to culture or the longer amount of time in which the parasite was hosted in the donor monkey, potentially allowing for P. vivax to stabilize and synchronize (17–21 days [14] vs 11–14 days).
Although the results of the Golenda study could not be replicated [14], we set out to compare a number of related variables: leukocyte depletion and its relation to the synchronicity of the starting parasite population; media conditions and supplements; and static versus dynamic culture conditions to improve culture conditions.
The choice of a leukocyte depletion method is important when setting up an in vitro Plasmodium culture as the presence of leukocytes is detrimental to parasite survival and the method chosen can influence the parasite population used to start the culture (parasitemia, stage, viability). In general, no consistent difference was observed in maturation rates when P. vivax-infected reticulocytes were purified from Aotus blood draws using either CF11-powder columns or Plasmodipur filters (Fig 4). However, the use of CF11 columns for leukocyte depletion from Aotus blood samples frequently resulted in the isolation of a mix-stage parasite population including young and mature parasite forms. In contrast, the use of Plasmodipur mostly resulted in the isolation of synchronous ring-stage parasite populations and much lower parasitemias. However, the CF11 parasites appeared to have reduced invasion efficiencies (Fig 4). Different hypotheses can be put forward to explain this finding: it is possible that a greater proportion of late stages are allowed through the CF11 column and these might be potentially less fit for adaptation to in vitro conditions. Alternatively, the shear forces involved in CF11 filtration may have a detrimental effect on the fitness of the late stages.
Along this same lines, starting parasite cultures using well-synchronized ring-stage populations [14] could improve progression of the culture as theoretically less metabolically active ring stages could better adapt to an in vitro environment while more mature stages, already exponentially increasing its metabolism, would suffer a “shock” with the sudden changes on the evolving environment. Conversely, the majority of P. vivax in vitro drug assays published to date rely on setting up cultures with samples rich in mature trophozoites [30, 31], which are allowed to develop for 20-24h to schizonts and then enriched by gradient techniques to overcome the reduction in parasitemia commonly observed during in vitro maturation. In our experiments we could not demonstrate the benefit from a 24 hr in vitro adaptation of rings stages to culture conditions as decreasing parasitemias were the constant and we could observe invasion in P. vivax cultures starting with either rings and or mixed stages.
The choice of an appropriate medium composition for parasite growth cannot be underestimated as, to complete the cycle, the parasite needs to take in nutrients from the media and not be overexposed to toxic by-products. In general, all McCoy’s 5A media contain L-glutamine, an essential amino acid for energy production as well as protein and nucleic acid synthesis. However, L-glutamine is quite unstable and is known to be subject to spontaneous degradation accompanied by the generation of byproducts such as ammonia and pyrrolidone carboxylic acid. GlutaMAX is a commercially available alternative to L-glutamine (dipeptide, L-alanine-L-glutamine), which is more stable and does not spontaneously degrade. The mechanism of dipeptide utilization by the parasite involves the gradual release of peptidase during the life cycle in culture, which allows for the gradual hydrolysis of the dipeptide in the medium resulting is an efficient energy metabolism and a high-growth yield. In our experiments, we found that adding GlutaMAX to the medium was beneficial for P. vivax maturation (Fig 5), although the parasitemia was not statically significant, we found a longer persistence of parasites in vitro, better conversion and second round invasion compared to the normal base media containing L-glutamine. This indicates that P. vivax may be sensitive to ammonia and pyrrolidone carboxylic acid.
Literature [12, 14] suggests that P. vivax may benefit from the alternation of static and shaking conditions as well as from an increase in hct (from 6 to 12% in Golenda et al [14]) just prior to reinvasion. In this context, Bass and Johnson 1912, suggested that invasion of reticulocytes by P. vivax should physiologically occur through direct contact between a fully multinucleated schizont and its target cell. This theory is supported by more recent data [32] showing that some of the P. vivax biomass is able to sequester in hematopoietic organs with slow circulation such as bone marrow and spleen (therefore facilitating cell-to-cell encounter) where the P. vivax parasite has access to an environment more rich in reticulocytes (especially the more immature subpopulations) than in peripheral blood. While based on this evidence it appears important to mimic a physiological template of constant circulation conditions through shaking, that would increase the chances of merozoites encountering reticulocytes, the time-point chosen for shifting from static to dynamic conditions may be critical. In fact, there is concern that changing from static to shaking conditions when P. vivax schizonts are already mature, could be detrimental on the P. vivax in vitro culture due to the fragility of this parasite stage [12]. Therefore, the shift from static to shaking conditions was performed early, prior to complete maturation of schizonts, in order to profit also from the early release of merozoite from early schizonts. Our data suggest that despite the marginal benefit shaking provides in terms of maturation, egress and re-invasion, this effect is not consistent across multiple samples and seems to depend largely on the staging of the starting population (Fig 6). Our findings suggest that shaking at 100rpm does not harm either maturation or invasion. Furthermore, we found that doubling the hct when the culture is first set up is beneficial to P. vivax growth (Fig 6). Although, in principle this appears counter-intuitive as a higher hct leads to a decrease in the medium nutrients available to the parasite, the beneficial effect observed may be explained by the need for a certain parasite/cell density to promote active communication between parasites through an exosome-like mechanism and/or a more direct contact among cells [33, 34]. In fact, we found that increased hct supported better parasite development (Fig 7), but this was due to an increase in uninfected cells rather than total parasite biomass.
In conclusion, while Plasmodium falciparum is the most life threatening of the 5 malaria parasite species-affecting humans, P. vivax causes the highest morbidity and is the most geographically widespread of the human Plasmodium. Therefore, there is a need to target this parasite in order to achieve malaria eradication. In order to make advances in the understanding of P. vivax’s biology, an in vitro culture for this parasite is urgently needed. Currently, research on P. vivax relies either on the use of field isolates or on the use of non-human primates adapted strains such as Sal-1, AMRU and Chesson [32]. Owing to their reproducibility, non-human primate samples have been widely used for P. vivax drug and vaccine in vitro and in vivo assays, and in the most successful attempt to date to establish an in vitro culture [14]. As we look to establish conditions to improve existing short-term in vitro P. vivax cultures and move towards establishing long-term cultures, we have chosen to use the well-adapted NHP P. vivax strain Sal-1 obtained from Aotus lemurinus lemurinus to benchmark parasite characteristics (specific media preferences, bottlenecks in in vitro growth and/or invasion). Although more experiments are needed to be performed some of the trends that were observed, our preliminary findings suggest that P. vivax Sal-1 strain invasion efficiency in culture was dependent on the reticulocyte enrichment method chosen and that parasite growth was positively influenced by the supplementation of culture media with stabilized nutrients indicating that P. vivax may be extremely sensitive to waste products. While the leuko-depletion method used appeared to have little impact on parasite progress, preliminary evidence suggests that shaking conditions and haematocrit may affect both P. vivax invasion and growth. Although further experiments are warranted to further investigate the huge number of variables able to influence P. vivax invasion, growth and egress, many of them not attempted in this work, we believe that the systematic approach presented in this paper, represents a point of departure for investigators interested in establishing long-term P. vivax cultures both with isolates from people in endemic countries and non-human primate samples.
|
10.1371/journal.pgen.1007849 | Elevated pyrimidine dimer formation at distinct genomic bases underlies promoter mutation hotspots in UV-exposed cancers | Sequencing of whole cancer genomes has revealed an abundance of recurrent mutations in gene-regulatory promoter regions, in particular in melanoma where strong mutation hotspots are observed adjacent to ETS-family transcription factor (TF) binding sites. While sometimes interpreted as functional driver events, these mutations are commonly believed to be due to locally inhibited DNA repair. Here, we first show that low-dose UV light induces mutations preferably at a known ETS promoter hotspot in cultured cells even in the absence of global or transcription-coupled nucleotide excision repair (NER). Further, by genome-wide mapping of cyclobutane pyrimidine dimers (CPDs) shortly after UV exposure and thus before DNA repair, we find that ETS-related mutation hotspots exhibit strong increases in CPD formation efficacy in a manner consistent with tumor mutation data at the single-base level. Analysis of a large whole genome cohort illustrates the widespread contribution of this effect to recurrent mutations in melanoma. While inhibited NER underlies a general increase in somatic mutation burden in regulatory elements including ETS sites, our data supports that elevated DNA damage formation at specific genomic bases is at the core of the prominent promoter mutation hotspots seen in skin cancers, thus explaining a key phenomenon in whole-genome cancer analyses.
| Cancer is caused by somatic mutations that typically occur in protein-coding genes. However, the advent of whole genome sequencing has made it possible to venture beyond protein-coding DNA in search of non-coding mutations with putative cancer driver roles. Indeed, recent studies, in particular in skin cancers, describe individual positions in gene regulatory regions (promoters) that are recurrently mutated in many independent patients, suggestive of a contribution to carcinogenesis. In this paper, we show that recurrent promoter mutations arise at these sites due to an exceptional propensity to form UV-induced DNA damage lesions (pyrimidine dimers) at specific transcription factor binding sites. The effect is present in cellular DNA but not in naked acellular DNA, meaning that the sites need to be occupied by their transcription factor partners in order to induce favorable conditions for DNA damage formation. This explains an important confounding phenomenon in whole cancer genome analyses, and has implications for the interpretation of recurrent somatic mutation patterns in non-coding DNA.
| Whole genome analysis of cancer genomes has the potential to reveal non-coding somatic mutations that drive tumor development, but it remains a major challenge to separate these events from non-functional passengers. The main principle for identifying drivers is recurrence across independent tumors, suggestive of positive selection, which led to the recent identification of frequent oncogenic mutations in the promoter of telomere reverse transcriptase (TERT) that can activate its transcription [1, 2]. However, mutation rates vary across the genome, and local elevations may give rise to “false” recurrent events that can be misinterpreted as signals of positive selection. While known covariates of mutation rate, such as replication timing and chromatin organization [3, 4], transcriptional activity [5] and local trinucleotide context [6], can be accounted for to improve interpretation [7], the non-coding genome may be particularly challenging. Mutational fidelity may be generally reduced in this vast and relatively unexplored space, as indicated by the presence of mechanisms directing DNA repair specifically to gene regions [8, 9], and yet-unexplained mutational phenomena may be at play.
Indeed, recent studies have described a remarkable abundance of recurrent promoter mutations in melanoma and other skin cancers, often noted to overlap with sequences matching the recognition element of ETS family transcription factors (TFs) [10–16]. Strikingly, a large proportion of frequently recurring promoter mutations in melanoma occur at distinct cytosines one or two bases upstream of TTCCG elements bound by ETS factors as indicated by ChIP-seq, within a few hundred bases upstream of a transcription start site [17]. While often interpreted as driver events, we recently showed that these sites exhibit highly elevated vulnerability to UV mutagenesis, as evidenced by their rapid induction following low-dose UV light exposure in cultured cells [17]. The effect has often been attributed to locally impaired nucleotide excision repair (NER) caused by binding of ETS TFs [16, 18, 19]. However, our analysis of skin tumors lacking global NER (XPC -/-) contradicted this model [17] and the mechanism remains unclear. An understanding of this phenomenon, which may underlie a large part of all non-coding recurrent events in human tumors beyond TERT [10, 12, 16], would resolve a key question that continues to confound whole cancer genome analyses.
Here, through analysis of 221 whole tumor genomes, we first demonstrate the widespread impact of TTCCG-related mutagenesis on the mutational landscape of melanoma. Moreover, through UV exposure of a panel of repair-deficient human cell lines, we rule out inhibited DNA repair as core mechanism. Finally, we generate the highest resolution map of UV-induced cyclobutane pyrimidine dimers (CPDs) in the human genome to date, which provides clear evidence that ETS-related promoter hotspots are associated with strong local elevations in the efficacy of UV lesion formation at specific genomic bases.
To assess the impact of TTCCG-related mutagenesis on the landscape of recurrent mutations in melanoma in a more sensitive way than previously possible, we assembled a cohort of 221 melanomas characterized by whole genome sequencing by TCGA and ICGC [20, 21]. These heavily mutated tumors averaged 110k somatic single nucleotide variants (SNVs) per sample, expectedly dominated by C>T transitions and a mutational signature characteristic of mutagenesis by UV light through formation of pyrimidine dimers (S1 Fig).
Notably, despite the genome-wide scope, nearly all highly recurrent mutations were found near annotated transcription start sites (TSSs) (Fig 1a). For example, of the 22 most recurrent individual bases (mutated in ≥18 patients), four were known drivers (BRAF, NRAS or TERT promoter mutations) while the rest were at most 524 bp away from a known TSS. Further, the vast majority of highly recurrent promoter sites were found in conjunction with TTCCG sequences (Fig 1a and 1b), indicating a widespread influence from ETS elements to the mutational landscape of melanoma. Analysis of TTCCG-related recurrent mutations in relation to enhancers further supported that the effect is largely restricted to promoters (S2 Fig).
Of 51 recurrent promoter mutations (+/- 500 bp from TSS) mutated in ≥ 12 tumors, 42 (82%) had a TTCCG element in the immediate (+/- 10 bp) sequence context, rising to 86% after excluding the known TERT C228T and C250T promoter mutations (Fig 1b and S1 Table) [1, 2]. Most were within 200 bp upstream of a known TSS, as expected for functional ETS elements (Fig 1b and 1c) [22]. Among the few remaining sites, two (upstream of AP3D1 and TMEM102) were instead flanked by TTCCT sequences likewise matching the ETS recognition motif (Fig 1b) [22] and the numbers are thus conservative. The fraction TTCCG-related sites increased as a function of recurrence, from 291/550 promoter sites (53%) at n ≥ 5 to 7/8 (88%) at n ≥ 20, excluding the known TERT sites (Fig 1d). For comparison, only 0.60% of C>T mutations in the dataset exhibited TTCCG patterns, underscoring their massive enrichment in recurrent positions.
As noted previously [17], there was a strong correlation between the number of mutated TTCCG hotspot sites and the total mutational burden in each tumor, compatible with these sites being passive passengers (Spearman’s r = 0.94, P = 1.5e-106; Fig 1e). Also confirming earlier observations, the TTCCG-related promoter hotspots were found preferably near highly expressed genes, as expected under a model where interaction with an ETS TF rather than sequence-intrinsic properties are responsible for elevated mutation rates in these sites (Fig 1f). Taken together, these analyses clearly demonstrate that ETS-related mutations account for nearly all highly recurrent non-coding hotspots genome-wide in melanoma, as well as hundreds of less recurrent sites not detectable in previous analyses based on smaller cohorts.
Recent studies have shown that NER, the main DNA repair pathway for UV damage, is attenuated in TF binding sites, leading to elevated somatic mutation rates [18, 19]. While plausible as a mechanism for ETS-related mutation hotspots [16], we recently showed that TTCCG elements were associated with elevated mutation rates also in cutaneous squamous cell carcinomas (cSCCs) lacking global NER (XPC -/-) [17]. We also established that mutations can be easily induced in TTCCG hotspot sites in cell culture by UV light, thus recreating in vitro the process leading to recurrent mutations in tumors [17]. We decided to use the RPL13A -116 bp hotspot site, notably more frequently mutated (58/221 tumors) than both canonical TERT sites and on par with BRAF V600E at 60/221 (Fig 1a and 1b), as a model to further investigate a possible role for impaired NER.
To this end, we UV-exposed A375 cells with intact NER as well as fibroblasts with homozygous mutations in four key DNA repair components: XPC, required for global NER, ERCC8 (CSA) and ERCC6 (CSB), required for transcription coupled NER (TC-NER), and XPA which is required for lesion verification in both global and TC-NER (S3 Fig). Correct genetic identity and complete homozygosity for the mutant allele was confirmed by whole-genome sequencing of all four mutant cell lines (S2 Table). Even limited UV exposure led to high cell mortality in the mutant cell lines, forcing us to limit the exposure to a single low dose of UVB (20 J/m2) during approximately two seconds followed by three weeks of recovery, after which cells were assessed for RPL13A promoter mutations using error-corrected amplicon sequencing (Fig 2a) [23]. Between 7,332 and 13,774 error-corrected reads at ≥10x oversampling were obtained for each of 10 different libraries (Fig 2b).
Strikingly, even at this miniscule dose, subclonal somatic mutations appeared preferably at the known hotspot site in A375 cells (Fig 2c) as well as in all of the mutant cell lines (Fig 2e–2g), despite abundant possibilities for UV lesion formation in flanking assayed positions. As expected, absolute mutation frequencies were low, less than 0.5% in all samples, bringing us close to the detection limit in some samples as indicated by noise in the untreated controls (Fig 2d, 2e and 2g). In combination with earlier data from tumors lacking global NER [17] and the fact that the mutations are almost exclusively positioned upstream of TSSs where TC-NER should not be active (Fig 1b and 1c), these results argue against impaired global NER as well as TC-NER as the basic mechanism behind TTCCG hotspot formation.
It was established decades ago that DNA conformational changes induced by interactions with proteins can alter conditions for UV damage formation [24, 25], which prompted us to investigate whether ETS-related promoter hotspots may arise due to locally favorable conditions for UV lesion formation. For this, we adapted a protocol first established in yeast using IonTorrent sequencing [26] to the Illumina platform (Fig 3a), to generate a genome-wide map of CPDs in A375 human melanoma cells immediately following UV exposure, before DNA repair processes have had a chance to act.
CPDs were preferably detected at TT, TC, CT and CC dinucleotides as expected (Fig 3b). An elevation at AT dinucleotides was consistent with an earlier report where this was attributed primarily to AT[T/C] sites, suggesting a contribution from CPDs at flanking dipyrimidines [26]. By comparing with median detection frequencies at non-dipyrimidines, we estimated the false positive rate for CPDs at dipyrimidines to vary from 5.5% for TT up to 17.0% for CC. The number of detected CPDs after removal of PCR duplicates scaled nearly linearly with simulated sequencing depth, indicating favorable random representation of CPDs (Fig 3c). A total of 202.1 million CPDs were mapped to dipyrimidines throughout the cellular genome (Fig 3c), constituting the highest resolution CPD map to date to our knowledge. Additionally, 95.3 million CPDs were mapped in UV-treated naked (acellular) A375 DNA lacking interacting proteins, while a non-UV-treated control, which expectedly yielded limited material, produced 18.5 million CPDs (Fig 3c).
We next investigated CPD formation patterns at TTCCG mutation hotspot positions identified above in melanoma (Fig 1a and 1b). 291 recurrently mutated (n ≥ 5/221 melanomas) TTCCG promoter sites (+/-500 bp from TSS) were aligned centered on the mutated base such that CPD density in these regions could be determined. This revealed a striking peak in CPD formation that coincided with the hotspots, which was largely absent in naked DNA lacking bound proteins or in non-UV control DNA (Fig 4a). Additionally, more recurrently mutated sites showed a stronger CPD signal, compatible with increased CPD formation being the key mechanism (Fig 4b).
For a more detailed understanding, we subcategorized the 291 melanoma ETS hotspot sites into four main groups based on sequence and mutated position. The strongest mutation hotspots, such as RPL13A and DPH3, typically occurred at cytosines one or two bases upstream of the TTCCG element (Fig 1b and S1 Table), which notably is outside of the core motif and therefore not expected to disrupt binding [22]. In CCTTCCG sites (n = 82 unique loci), recurrent C>T transitions would typically appear at both 5ʹ cytosines (underscored) independently or, less frequently, as CC>TT double nucleotide substitutions. Aggregated CPD density over these sites, centered on the motif, revealed a strong peak bridging these two bases, which notably was absent in naked DNA (Fig 4c). Thus, when the TF site is occupied, CPDs form efficiently between the two pyrimidines, leading to C>T mutations at either base although with a preference for the second position, in agreement with established models for UV mutagenesis [27]. The same pattern of strongly elevated CPD formation in cellular, but not naked, DNA was observed between the same positions in TCTTCCG and CTTTCCG sites (n = 57 and 27, respectively), with C>T mutations expectedly forming only at the first or second pyrimidine depending on the position of the cytosine (Fig 4d and 4e).
Many of the less recurrent bases in melanoma were often found at the first middle cytosine of a TTCCG motif (S1 Table). Interestingly, a large fraction of these sites lacked a dipyrimidine at the two key positions identified above thus prohibiting CPD formation there, with ACTTCCG being the most common pattern (44/82 sites), which indeed matches the in vivo ETS consensus sequence [22]. Compatible with the mutation data, the strongest CPD peak was observed at the middle TC dinucleotide, and in agreement with the lower mutation recurrence, this signal was weaker compared to the other site types (Fig 4f). Of note, elevated CPD formation between these bases could also be clearly seen in the other site categories (Fig 4c–4e). Taken together, these analyses based on genome-wide CPD mapping provide strong evidence that locally elevated CPD formation efficacy shapes the formation of mutation hotspots at ETS binding sites.
Earlier studies have described a general increase in mutation rate in promoter regions, attributed to reduced NER activity at sites of TF binding including ETS sites [18, 19]. To investigate a possible contribution from increased CPD formation to this pattern, we first determined the overall mutation rate in melanoma near TSSs, which confirmed a sharp increase in upstream regions that coincided with reduced NER as determined by XR-Seq (Fig 5a and 5b) [28]. Further confirming earlier data [19], this increase was abrogated in XPC -/- cSCCs lacking global NER, arguing against a major contribution from increased CPD formation (Fig 5a). Consistent with this, CPDs were found to form at near-expected frequencies when aggregated over these regions (Fig 5c). Interestingly, subtraction of TTCCG-related mutations revealed that these constitute a large proportion of promoter mutations in melanomas, but not in XPC -/- cSCCs, supporting a notable contribution from inhibited NER in ETS sites to the overall burden increase in promoters (Fig 5a).
While elevated UV-induced DNA damage is important in the formation of ETS-related recurrent mutation hotspots, we conclude that this effect has negligible impact on the general increase in mutation burden in regulatory regions. This is instead explained by repair inhibition including a prominent contribution from impaired NER in ETS sites, which can likely further add to elevated mutation frequencies at recurrent hotspot positions.
Proper analysis of recurrent non-coding mutations requires an understanding of how mutations arise and distribute across the genome in the absence of selective pressures. Here, we provide a mechanistic explanation for the passive emergence of recurrent mutations at specific positions in TTCCG/ETS sites in tumors in response to UV light, and also demonstrate the massive impact of such mutations on the mutational landscape of melanoma using a large whole genome cohort.
Mutations at -116 bp in the RPL13A promoter were used here as a model to study mutation formation at ETS hotspot sites in vitro in repair-deficient cell lines, which ruled out inhibited repair as sole mechanism. Of note, this site is more recurrently mutated than the individual TERT C228T/C250T sites and nearly as frequent as chr7:140453136 mutations (hg19) pertaining to BRAF V600E, thus representing the second most common mutation in melanoma and likely other skin cancers. Notably, mutations were detectable at this site in cultured cells following a UVB dose of 20 J/m2 UVB, equivalent to about 1/200th of the monthly absorbed UVB dose in July in Northern Europe [29]. This underscores the extreme UV sensitivity of ETS hotspots and explains their high recurrence in tumors.
Genome-wide mapping of CPDs revealed that TTCCG-related mutation hotspots exhibit highly efficient CPD formation at the two bases immediately 5ʹ of the core TTCC ETS motif. The effect was lost in naked acellular DNA, showing that structural conditions for elevated CPD formation are induced when the TF binding site is in its protein-bound state. Interestingly, most functional ETS sites are expected to lack pyrimidines in the two key positions [22] thus prohibiting pyrimidine dimer formation, and conditions for forming a strong mutation hotspot are therefore only met in a subset of sites with CC, TC or CT preceding the TTCCG element. Additionally, CPDs form at lower but still elevated frequency at the middle TTCCG bases, consistent with weaker recurrence for mutations in these positions. CPD and cancer genomic data are thus in strong agreement, providing a credible mechanism for the formation of ETS-related mutations hotspots in UV-exposed cancers.
As demonstrated here, frequent mutations at ETS-site hotspots are expected for purely biochemical reasons in UV-exposed cancers. Consequently, several observations are compatible with passenger roles for these mutations: The most recurrent sites arise at cytosines outside of the core TTCC ETS recognition element [22] where they are not expected to disrupt ETS binding. While mutations in the middle of the motif, common among the less frequent hotspots, should disrupt binding, ETS factors tend to be oncogenes that are activated in cancer [30], and it can be noted that TERT promoter mutations instead enable ETS binding through formation of TTCC elements [1, 2]. The mutations tend to arise near highly expressed housekeeping genes rather than cancer-related genes, and the particular set of sites that are mutated varies inconsistently in-between tumors. Moreover, as would be expected in the absence of selection and in contrast to known driver mutations [17], the number of mutated ETS sites in a tumor is strongly determined by mutational burden.
Our results complement a recent study by Mao et al. [31], which was published during the preparation of this manuscript. This study likewise determined CPD formation patterns in ETS binding sites using whole genome CPD mapping obtaining results that are in full agreement with ours, and additionally proposed a structural basis based on available crystallography data for increased CPD formation at the center TC dinucleotide in the ETS-DNA complex, which was demonstrated to promote CPD formation also in vitro. Thus, data on CPD formation patterns from two independent studies, in combination with our data showing a sharply elevated mutation rate at the RPL13A TTCCG hotspot site in vitro in the absence of NER, support that base-specific elevations in CPD formation efficacy forms the foundation for prominent promoter mutation hotspots in skin cancers. At the same time, inhibited DNA repair explains a general increase in mutation burden in regulatory elements including ETS sites, which could act synergistically to further amplify elevated mutation rates at ETS-related hotspots. Future studies may want to better quantify the relative contributions of these effects, as well as define the exact subset of ETS factors or other proteins that interact with DNA at TTCCG-related mutation hotspot sites.
Whole genome somatic mutation calls from the Australian Melanoma Genome Project (AMGP) cohort [20] were downloaded from the International Cancer Genome Consortium’s (ICGC) database [32]. These samples were pooled with whole genome mutation calls from The Cancer Genome Atlas (TCGA) melanoma cohort [21] called as described previously [10]. Population variants (dbSNP v138) and duplicate samples from the same patient were removed, resulting in a total of 221 tumors. Whole genome sequencing data from 5 XPC -/- cSCCs and matching peritumoral skin was obtained from Zheng et al., 2014 [5], and aligned with bwa (v0.7.12) [33] followed by mutation calling using VarScan 2 (v2.3) [34] and subtraction of population variants.
Gene annotations from GENCODE [35] v19 were used to define TSS positions, encompassing 20,149 and 13,307 uniquely mapped coding genes and lncRNAs, respectively, considering the 5ʹ-most annotated transcripts while disregarding non-coding isoforms for coding genes. Processed RNA-seq data was derived from Ashouri et al., 2016 [36]. Enhancer annotations were derived from ChromHMM segmentation (Core 15-state model, E6 and E7 regions, representing enhancers and genic enhancers, respectively) of epigenomic data from foreskin melanocytes (Roadmap celltype E059) [37].
XP12, GM16094, GM16095 and GM15893 cells were a kind gift from Dr. Isabella Muyleart, University of Gothenburg. Cells were grown in DMEM + 10% FCS + Penicillin/streptomycin (GIBCO). Cells were subjected to a single low dose UVB (20 J/m2) and left to recover for three weeks. DNA was extracted with Blood Mini kit (Qiagen).
To detect and quantify mutations we applied SiMSen-Seq (Simple, Multiplexed, PCR-based barcoding of DNA for Sensitive mutation detection using Sequencing) as described previously [17]. Sequencing was performed on an Illumina MiniSeq instrument in 150 bp single-end mode. Raw FastQ files were subsequently processed as described using Debarcer Version 0.3.1 (https://github.com/oicr-gsi/debarcer/tree/master-old). For each amplicon, sequence reads containing the barcode were grouped into barcode families. Barcode families with at least 10 reads, where all of the reads were identical (or ≥ 90% for families with >20 reads), were required to compute consensus reads. FastQ files were deposited in the Sequence Read Archive under BioProject ID SRP158874.
A375 cells were grown in DMEM + 10% FCS + Penicillin/streptomycin (Gibco, Carlsbad, MA) and were treated with 1000 J/m2 UVC following DNA extraction and DNA from untreated cells was isolated as a control, both in duplicates. Additionally, naked DNA from untreated cells was irradiated with the same dose, to provide an acellular DNA control sample. DNA was extracted with the Blood mini kit (Qiagen, Hilden, Germany). Purified DNA (12 μg) was sheared to 400 bp with a Covaris S220 in microtubes using the standard 400 bp shearing protocol. CPD-seq was modified from Mao, Smerdon [26] to adapt it to Illumina sequencing methods using primers described previously in Clausen et al., 2015 [38] (S3 Table). Briefly, sheared DNA was size selected with SPRI select beads (1.2 vol) (Life Technologies, Carlsbad, CA) and the purified product (approx. 4 μg) subjected to NEBNext end repair and NEBNext dA-tailing modules (New England Biolabs (NEB), Ipswich, MA). ARC141/142 (8 μM) was then ligated to the sheared and repaired ends O/N with NEBNext Quick Ligation module. DNA was purified with 0.8 vol CleanPCR beads and treated with 40 units Terminal Transferase (TdT, NEB) and 0.1mM dideoxy ATP (Roche, Rotkreuz, Switzerland) for 2h at 37 °C. DNA was purified and incubated with 30 units T4 endonuclease V (NEB) at 37 °C for 2 h, followed by purification and treatment with 15 units APE1 (NEB) at 37 °C for 1.5 h. DNA was purified and treated with 1 unit rSAP (NEB) 37 °C 1 h followed by deactivation at 65 °C for 15 minutes. DNA was purified, denatured at 95 °C for 5 min, cooled on ice and ligated with the biotin-tagged ARC143/144 (0.25 μM) overnight at 16 °C with NEBNext quick ligation module. DNA fragments with the biotin tag were captured with 20 μl Streptavidin Dynabeads (Invitrogen, Waltham, MA) and the DNA strand without the biotin label was released with 2 x 40 μl 0.15 M NaOH and ethanol precipitated. This single-stranded DNA was resuspended in 14.9 μl H2O and used as the template to synthesize double-stranded products using ARC154 (0.25 μM) by incubating with Phusion High-Fidelity DNA Polymerase (Thermo Scientific, Waltham, MA) at 98 °C for 1 min, 58 °C for 30 s and 72 °C for 1 min. The now double-stranded library was purified and amplified for 15 cycles with ARC49 and ARC78-82 (0.3 μM each) to add Illumina barcodes and indexes. Two cellular UV-treated, two no-UV controls and one naked DNA control library were prepared, for a total of five libraries. The libraries were pooled with equal volumes of each of the libraries and sequenced using a NextSeq High Output kit (Illumina, San Diego, CA). The data has been deposited in GEO under accession GSE119249.
FastQ files were aligned pairwise with Bowtie 2 version 2.3.1 [39] to hg19, using standard parameters. For the -UV control and +UV cellular DNA samples, replicates were merged with Picard MergeSamFiles version 2.18.7 (http://broadinstitute.github.io/picard). Duplicate reads were marked with Picard MarkDuplicates version 2.18.7 [40] with the parameter VALIDATION_STRINGENCY = LENIENT. Further analysis was performed in R with Bioconductor [41], where CPD positions were extracted as the two bases upstream and on the opposite strand of the first mate in each read pair, removing those that mapped outside of the chromosome boundaries. Only biologically possible CPDs detected at dipyrimidines sites were considered in the CPD counts and downstream analyses. Data from duplicate libraries were pooled to achieve higher coverage, since downstream results were in close agreement when considering these libraries individually. To simulate lower coverage libraries, the bam files were subsampled with samtools view version 0.1.19-44428cd [42] with the parameter -s at 0.25, 0.5 or 0.75, and the subsequent bam files were reanalyzed as described above.
For analyses of CPD formation patterns, C>T mutations and repair activity around TSSs, these regions were divided into 20 bp bins in which CPD counts or overlapping XR-seq reads were determined. XR-seq data from wild-type NHF1 skin fibroblasts was obtained from Hu et al., 2015 [43], and consisted of normalized read counts in 25 bp strand-specific bins. Background frequencies of dinucleotides and trinucleotides in hg19 were counted with EMBOSS’s fuzznuc [44], using the parameters -auto T -complement T. Expected mutations were calculated by randomly introducing the same number of mutations as observed in the window based on observed probabilities for C>T mutations at different trinucleotides estimated from the complete mutation dataset. Expected CPDs were calculated in the same way, maintaining the number of CPDs in the observed data, but based instead on genomic dinucleotide counts.
|
10.1371/journal.pgen.1002010 | A Bow-Tie Genetic Architecture for Morphogenesis Suggested by a Genome-Wide RNAi Screen in Caenorhabditis elegans | During animal development, cellular morphogenesis plays a fundamental role in determining the shape and function of tissues and organs. Identifying the components that regulate and drive morphogenesis is thus a major goal of developmental biology. The four-celled tip of the Caenorhabditis elegans male tail is a simple but powerful model for studying the mechanism of morphogenesis and its spatiotemporal regulation. Here, through a genome-wide post-embryonic RNAi-feeding screen, we identified 212 components that regulate or participate in male tail tip morphogenesis. We constructed a working hypothesis for a gene regulatory network of tail tip morphogenesis. We found regulatory roles for the posterior Hox genes nob-1 and php-3, the TGF-β pathway, nuclear hormone receptors (e.g. nhr-25), the heterochronic gene blmp-1, and the GATA transcription factors egl-18 and elt-6. The majority of the pathways converge at dmd-3 and mab-3. In addition, nhr-25 and dmd-3/mab-3 regulate each others' expression, thus placing these three genes at the center of a complex regulatory network. We also show that dmd-3 and mab-3 negatively regulate other signaling pathways and affect downstream cellular processes such as vesicular trafficking (e.g. arl-1, rme-8) and rearrangement of the cytoskeleton (e.g. cdc-42, nmy-1, and nmy-2). Based on these data, we suggest that male tail tip morphogenesis is governed by a gene regulatory network with a bow-tie architecture.
| Morphogenesis is a process in which cells change their shape and position to give rise to mature structures. Elucidation of the molecular basis of morphogenesis and its regulation would be a major step towards understanding organ formation and functionality. We focus on a powerful model for morphogenesis, the four-celled tail tip of the C. elegans male, which undergoes morphogenesis during the last larval stage. To comprehensively determine the components that regulate and execute male tail tip morphogenesis, we performed a genome-wide RNAi screen. We identified 212 genes that encode proteins with roles in fundamental processes like endocytosis, vesicular trafficking, cell–cell communication, and cytoskeletal organization. We determined the interactions among several of these genes to reconstruct a first draft of the genetic network underlying tail tip morphogenesis. The structure of this network is consistent with the "bow-tie architecture" that has been proposed to be universal and confers evolvability and robustness to biological systems. Bow-tie networks have a conserved core which is linked to numerous input and output components. Many components of the network underlying tail tip morphogenesis in C. elegans are conserved all the way to humans. Thus, understanding tail tip morphogenesis will inform us about morphogenesis in other organisms.
| Morphogenesis involves the coordinated change in the shape of cells and tissues during development, eventually giving rise to functional structures in the adult animal. Such coordinated change must occur at the correct time and in the proper position. In the case of structures that differ between the sexes, this process must also be regulated sex-specifically. While many genes and pathways are known that regulate development, the identity of genes that link regulation to the execution of morphogenesis have been more difficult to ascertain [1]. The many different cues and signals that must be integrated to control morphogenesis, combined with the complexity of the molecular machinery associated with this process, suggest that a large number of genes and gene products are involved.
To elucidate the molecular mechanisms underlying morphogenesis, the first step is thus to determine what components are involved in its regulation and execution and to determine how they interact in a network. In the pursuit of such an aim, it is advantageous to use a simple model structure that still demonstrates all the properties of cellular morphogenesis. The model we use is the male tail tip of Caenorhabditis elegans. This structure is made up of four epithelial ("hypodermal") cells, hyp8–hyp11, which are born during embryogenesis. Embryonic morphogenesis of hyp8–hyp11 leads to the formation of a pointed, whip-like tail tip. The tail tip retains this shape throughout the lifespan of the hermaphrodite. However, during the last larval stage (L4) of males, these conical cells are dramatically remodeled to form the rounded tip of the adult [2]–[4]. Male tail tip morphogenesis begins when hyp8–11 fuse to form a syncytium; fusion is followed by detachment of the cells from the overlying cuticle. Towards the middle of the L4 stage, the syncytium changes its shape from conical to round and moves anteriorly; these morphogenetic events cease at the end of the L4 stage (Figure 1A). In hermaphrodites, the tail tip cells do not fuse and do not change shape. The tail tip model thus allows the study of a sexual dimorphism at the cellular level. Another advantage of this model is that male-specific mutations in C. elegans can be propagated through the self-fertile hermaphrodites, even if the mutations affect male fertility, mating ability, or viability.
A few mutations in genes involved in the regulation of tail tip morphogenesis have been described previously [5]–[7]. Some of these mutations impede or completely block tail tip morphogenesis, resulting in the retention of the pointed larval tail tip in the adult male, a phenotype that we call "Lep" (Figure 1C). This phenotypic designation is derived from the term "leptoderan," which—in the taxonomic literature—describes the unretracted, pointed male tail tip in some nematodes related to C. elegans [8], [9]. Other mutations cause precocious onset of (and thus an extended total period for) tail tip retraction, which results in over-retracted ("Ore") and thus abnormally shortened adult male tails [6] (Figure 1C). Studies of these mutants have revealed a few of the important regulatory components for tail tip morphogenesis. DMD-3 was suggested to be a central regulator of tail tip morphogenesis, as it is required for tail tip retraction in males and is sufficient for inducing ectopic morphogenesis in hermaphrodite tail tips [7]. The gene encoding this DM-domain transcription factor, dmd-3, is a homolog of dmrt1 in vertebrates and doublesex in Drosophila [10], [11]. DMD-3 functions cooperatively and partially redundantly with a closely related factor, MAB-3 [7], also involved in somatic sex determination [12]. TRA-1, the most downstream global regulator in the sex-determination pathway, inhibits the expression of these genes in hermaphrodites [7], [13], [14]. The initiation of dmd-3 expression at the proper developmental stage is controlled by the "heterochronic" pathway [7]. Finally, maintenance of dmd-3 expression levels is regulated by Wnt signaling and a feedback loop involving both MAB-3 and DMD-3 [7].
Only one effector of a cellular process is known to be downstream of DMD-3 and MAB-3, namely the fusogen-encoding gene eff-1, which is important for the many fusions that occur between epithelial cells in C. elegans [15], [16]. DMD-3 and MAB-3 induce expression of eff-1 through an unknown post-transcriptional mechanism [7].
To find additional genes with roles in tail tip morphogenesis, we carried out a genome-wide, post-embryonic RNAi-feeding screen. This screen identified 212 candidates. Starting with these candidates, we used network model-building [17], transgenic reporter lines and expression epistasis analysis to construct a first draft of the gene regulatory network for tail tip morphogenesis. We found that dmd-3 expression is regulated by a nuclear hormone receptor (NHR-25), a new heterochronic gene, Hox anteroposterior patterning factors, and GATA transcription factors. NHR-25 is in turn negatively regulated by dmd-3 and mab-3. In addition, dmd-3 and mab-3 negatively regulate other signaling modules, including the TGF-β pathway. We also found that DMD-3 and MAB-3 regulate the localization or expression of genes involved in vesicular trafficking/endocytosis, cell polarity and cytoskeletal organization. Our data thus strongly support the hypothesis that DMD-3, MAB-3 and NHR-25 are central regulators of tail tip morphogenesis.
The genetic architecture for tail tip morphogenesis which emerges from this analysis closely resembles the "bow-tie" architecture, a possibly universal characteristic of robust, evolvable systems [18]. A bow-tie network has many inputs and outputs that are connected through a conserved core. Versatile weak linkages form the interface between the core and the input and output. In such a network, there is not a simple flow of information from input to output through the core; instead, there is extensive global and local feedback regulation found at every level [19].
A bow-tie architecture has been found to underlie a variety of biological networks: metabolic networks [20], [21], the Toll-like receptor signaling network [22], the epidermal growth factor receptor signaling network [23], the osmolarity glycerol signal-transduction pathway in yeast [24], stress response networks [25] and the immune system [26], [27]. In fact, it has been proposed that the bow-tie architecture of regulatory networks is ubiquitous because this structure ensures not only robustness but also evolvability [18], [19].
We found evidence for the existence of each aspect of a bow-tie architecture in the gene network governing tail tip morphogenesis. To our knowledge, this is the first time that this kind of network architecture has been explicitly identified in the context of development and morphogenesis. However, we believe that other developing systems are also governed by bow-tie genetic networks, supporting the proposition that this network architecture is universal.
To identify tail tip morphogenesis genes, we cultured animals on dsRNA-expressing bacteria from L1 to adult and looked for evidence of defective morphogenesis. Using the Ahringer RNAi-feeding library [28], we screened 16,131 genes, approximately 83% of the genome. Genes that gave a positive Lep or Ore phenotype were screened again; only repeatable positives were kept as candidates. We identified 212 genes, of which 190 produced Lep phenotypes, 14 produced Ore phenotypes, and 8 produced both Lep and Ore phenotypes in a single experiment (Figure 1D). Positives in each category were analyzed for GO-attribute enrichment using FuncAssociate [29]. Relative to the genome, Lep positives showed enrichment of GO-attributes associated with components or processes involved in morphogenesis, such as anchoring junctions and cell migration. However, Ore and Ore/Lep positives were enriched for genes involved in cell division/cytokinesis, and nuclear/chromosome organization (Figure 1D). The 212 candidates are listed and categorized by developmental pathway or annotated cellular process (if known) in Table S1. Raw RNAi data are publicly available along with representative images via the "Male Tail Tip Database" (MTTdb, at http://wormtails.bio.nyu.edu).
We identified components of conserved and widely studied developmental regulatory pathways (e.g. Hox anteroposterior patterning factors, TGF-β signaling module, heterochronic pathway, and GATA transcription factors). Importantly, the screen also identified components of fundamental cellular structures and processes that are likely to be important for the execution of morphogenesis. These include vesicular trafficking/endocytosis, cellular polarity, cytoskeleton, cell junctions, and nuclear export/import. Genes with the most severe and/or most penetrant RNAi phenotypes were selected for further study (Table 1).
Hox genes play central roles in shaping animal body plans by distinguishing different fields of cells along the anteroposterior axis [30]. Hox proteins regulate not only the expression of other transcription factors, but also of genes involved in specific processes such as morphogenesis [31]. Our screen has shown that posterior Hox patterning is crucial for tail tip morphogenesis. RNAi knockdowns of the Abd-B homologs php-3 and nob-1 postembryonically cause Lep phenotypes (Figure S1A, Tables S1 and S2). Likewise, a php-3 null allele, ok919, results in a 100%-penetrant Lep phenotype of moderate severity (Figure S1B, Table S2). The severity of the php-3(ok919) phenotype is increased with nob-1 RNAi treatment (Table S2). A nob-1::gfp translational fusion construct (containing a genomic fragment with the nob-1 gene and 9 Kb of the 5′-regulatory region) is expressed in the tail tip cells hyp8–11 throughout larval development in both sexes (Figure 2A and data not shown). A php-3::gfp translational fusion driven only by the short 500-bp intergenic region between nob-1 and php-3 shows variable expression in the nuclei of hyp8–11 during tail tip morphogenesis (data not shown). This latter transgene only partially rescues the tail tip phenotypes of php-3 mutants (Table S2), suggesting that regulatory elements upstream of both php-3 and nob-1 are required for appropriate expression of php-3. Using a tail tip-specific promoter (from the gene lin-44 [32]), we observed that expression of the PHP-3::GFP fusion protein product in hyp8–11 is sufficient to rescue the php-3(ok919) Lep phenotype (Figure S2A and Table S2). Mosaic animals that only express PHP-3::GFP in a subset of tail tip cells do not show rescue (Figure S2B). These data suggest that Hox-mediated patterning by PHP-3 and NOB-1 is carried out cell-autonomously and is required in all tail tip cells (hyp8–11) for proper morphogenesis.
GATA transcription factors play important regulatory roles during the differentiation of multiple cell types in normal development [33], [34] and during tumorigenesis [35]. RNAi knockdown of the gene for the GATA factor egl-18 resulted in Lep phenotypes (Figure S1C, Tables S1 and S2). An egl-18 null allele, ok290, causes Lep phenotypes of varying severity with 45% penetrance (Figure S1D, Table S2). Another GATA transcription factor that was missed in our RNAi screen, elt-6, shares an operon with egl-18 [36]. We repeated the RNAi treatment against elt-6 and observed low-penetrance, low-severity tail tip defects (Table S2). RNAi treatment for elt-6 in the egl-18(ok290) mutant strain, however, dramatically increased the penetrance and severity of the tail tip defects, such that tail tip morphogenesis failed altogether in some individuals (Table S2). Interestingly, an egl-18::gfp translational fusion is expressed in the nuclei of the main body epidermal syncytium hyp7, but appears to be excluded from the tail tip cells hyp8–11 (Figure 2B). This observation suggests that EGL-18 and ELT-6 function in cells adjacent to the tail tip to regulate tail tip morphogenesis cell-nonautonomously. Furthermore, transforming wild-type animals with the egl-18/elt-6 operon regulated by the lin-44 promoter disrupted morphogenesis (data not shown), suggesting that morphogenesis requires exclusion of these GATA factors from hyp8–11. In other epidermal cells, it has been observed that EGL-18 and ELT-6 repress cell fusion [37]. Thus, their exclusion from the tail tip cells may be required to allow tail tip cell fusion and subsequent morphogenesis.
TGF-β signaling is a major conserved cell-signaling module which regulates multiple processes during the development of all animals and also during the progression of cancer [38]. The screen identified two components of this pathway. RNAi treatment against sma-3, which encodes a Smad protein, and daf-4, encoding the TGF-β receptor, resulted in Lep phenotypes (Figure S1E, Tables S1 and S2). A sma-3 null allele, e491, resulted in a 57%-penetrant low-severity Lep phenotype (Figure S1F, Table S2). Of the other known components of the TGF-β pathway, only sma-2 showed any RNAi phenotype: RNAi against sma-2 significantly enhanced the penetrance of the Lep phenotype of sma-3(e491) (Table S2). A transgenic strain expressing a sma-3::mCherry translational reporter shows expression at low levels in the cytoplasm of the tail tip of both sexes at the beginning of the L4 stage. In males, the SMA-3::mCherry fusion protein enters the nuclei of hyp8–11 prior to tail tip morphogenesis and remains in both the cytoplasm and nuclei during morphogenesis (Figure 2C). In hermaphrodites, SMA-3::mCherry remains cytoplasmic throughout L4 and never enters the nuclei (data not shown). The dynamic localization pattern of SMA-3::mCherry suggests that TGF-β-mediated gene expression occurs concurrently with tail tip morphogenesis. Consistent with this hypothesis, DAF-4::YFP fusion protein localizes to the plasma membranes of hyp8–11 during tail tip morphogenesis (Figure 2D). Taken together, the RNAi results, loss-of-function mutant phenotypes and expression patterns of sma-3 and daf-4 suggest that TGF-β signaling is required during tail tip morphogenesis.
Previous work has shown that Wnt signaling plays a crucial role in the regulation of tail tip morphogenesis [5]. Consistent with those findings, the RNAi screen identified additional genes that are in or interact with the Wnt pathway. RNAi treatments against sys-1 (β-catenin) and lit-1 (Nemo-like kinase) each resulted in Lep phenotypes (Table S1, Figure S3C, S3D). A transgenic strain expressing a SYS-1::GFP fusion protein shows faint cytoplasmic expression in hyp8 and hyp11 but not in hyp9 or hyp10 (Figure S3C). A strain expressing a LIT-1::GFP fusion protein shows nuclear expression in hyp9 and hyp10 but not in hyp8 or hyp11 prior to morphogenesis (Figure S3D).
We identified multiple nuclear hormone receptor genes in our screen: nhr-9, nhr-23, nhr-25 and nhr-165. NHR-25 is the C. elegans homolog of FTZ-F1 [39] which is a highly conserved protein with diverse functions regulating embryonic patterning [40], [41], and ecdysone-mediated molting in Drosophila [42]. Both nhr-25 RNAi and a hypomorphic allele, ku217, showed Lep phenotypes (Figure S1G and S1H, Tables S1 and S2). Furthermore, nhr-25 RNAi treatments on the ku217 strain showed a complete lack of male tail morphogenesis in 33% of the animals (N = 24, Table S2), a phenotype reminiscent of the mab-3(e1240);dmd-3(ok1327) double mutant [7]. We were not able to produce transgenic lines expressing NHR-25 fusion proteins due to lethality, as previously reported [43]. Instead, we made a transgenic strain expressing a transcriptional reporter containing the 5′- and 3′-regulatory regions for nhr-25. This reporter shows a dynamic expression pattern in which expression in hyp8–11 is intense prior to and at the beginning of morphogenesis, rapidly shuts off in late L4, and is never on in adult animals (N = 34 adults) (Figure 2E). RNAi against another nuclear hormone receptor gene, nhr-165, showed a low-penetrant, mild, yet reproducible Lep phenotype upon RNAi treatment (Table S1, and Figure S3A). An nhr-165::gfp translational reporter shows nuclear expression in the lateral hypodermis (hyp7, but not hyp8–11) near the time of larval molts. This expression is highest and seen in the largest number of cells in the tail prior to the L3–L4 molt (Figure S3A).
The heterochronic pathway ensures that tail tip morphogenesis is initiated precisely at the beginning of the L4 stage [6]. We identified genes in this pathway: the zinc-finger transcription factor gene blmp-1, and known suppressors of the miRNA let-7 [44]. BLIMP-1 is a transcriptional repressor that regulates ecdysone-mediated molts in Drosophila [45] and differentiation of multiple cell types in humans and mice, such as lymphocytes [46] and primordial germ cells [47], [48]. Intriguingly, BLIMP-1 is a target of let-7-mediated degradation in Reed-Sternberg cells, a Hodgkin-lymphoma cell line, suggesting a possibly conserved interaction with the heterochronic pathway, of which let-7 is a member [49], [50]. RNAi treatments directed against blmp-1 produced Ore phenotypes (Table S1). A deletion allele, tm548, produces an Ore phenotype with 100% penetrance (Table S2) due to precocious initiation of tail tip morphogenesis during the L3 stage (N = 26, data not shown). A transgenic line expressing a BLMP-1::GFP fusion protein shows both nuclear and cytoplasmic expression in hyp8–11 throughout development, although cytoplasmic expression is most intense during tail tip retraction (Figure 2F).
Of the 41 suppressors of let-7 lethality identified by Ding et al. [44], six were positives in our screen. RNAi knockdown of two (pri-2, npp-6) resulted in the Ore phenotype, of two others (spg-7, smo-1) in the Lep phenotype and of two further genes (cdt-1, xpo-2) in both phenotypes in a single experiment. Four of these genes—pri-2, npp-6, cdt-1, and xpo-2—are predicted by N-Browse [17] to interact in a subnetwork that includes other genes for which RNAi knockdown also produced Ore phenotypes (Figure 3). It is thus possible that these additional genes also influence the timing of tail tip morphogenesis. It is still unclear why both Ore and Lep phenotypes are observed in a single experiment. One possible explanation is that precision in timing of tail tip morphogenesis is lost when certain gene-products are removed. In this case, morphogenesis might begin too early in one animal and too late in another, leading to Ore tails and Lep tails, respectively.
Post-transcriptional regulation appears to play an important role during the coordination of tail tip morphogenesis, as our screen has identified multiple genes that encode RNA splicing factors, kinases, phosphatases and ubiquitinating enzymes. One such gene is ubc-12, which is a part of the NED-8 conjugating system and has been shown to be important in epidermal differentiation during embryogenesis in C. elegans [51]. RNAi-knockdown of ubc-12 resulted in larval lethality in the RNAi hyper-sensitive rrf-3(-) background (Table S4). However, we identified ubc-12 in a pilot screen which was carried out in the wild-type background; ubc-12 RNAi treatments produced a highly penetrant and severe tail tip defect (Figure S3B). UBC-12::GFP expresses intensely in the cytoplasm of hyp10 just prior to and during tail tip retraction (Figure S3B).
We identified many genes known to play central roles during the execution of morphogenesis. Such genes encode proteins involved in vesicular trafficking, endocytosis, cell-cell communication, cytoskeletal rearrangement, establishment of cellular polarity and cell-cell transport.
Adjacent to the tail tip cells lie the dendritic projections of the PHC neurons. Two genes that produced Lep phenotypes upon RNAi treatment are expressed in these neurons but not in the tail tip cells: the calcipressin-encoding gene rcn-1 [54] and ptl-1, a gene encoding a tau-like microtubule-associated protein [55], [56]. A null allele of ptl-1, ok621, produced Lep phenotypes with 23% penetrance (Table S2). PTL-1::GFP and RCN-1::GFP fusion proteins are expressed in the PHC neurons prior to and during hypodermal morphogenesis. In adults, PTL-1::GFP is expressed in most neurons of the tail (data not shown). RCN-1::GFP is expressed in most tail neurons and in the support cells of the phasmid neurons (socket cells, Figure 4H). This pattern is consistent with the previously described adult expression pattern of RCN-1 [54].
With a list of genes required for tail tip morphogenesis, we next sought to characterize the interactions between these genes. We constructed a working hypothesis for these interactions using N-Browse, a publicly available database which integrates the information from numerous genome-wide studies to build gene networks [17]. We manually entered our candidate genes into N-Browse, excluding those that did not have predicted functions or known interactors. N-Browse produced a genetic network that included not only our candidate tail tip morphogenesis genes, but additional genes predicted to be nearest-neighbor interactors. To the resulting N-Browse network, we manually added gene interactions (edges) based on published work not represented in N-Browse [6], [7], [44], [51], [57]–[61] (Figure 5). This analysis predicts the involvement of genes not identified in our screen. We tested two of these predictions by repeating RNAi knockdown with different methods and/or by scoring larger numbers of males. We could thus validate roles in morphogenesis for elt-6 and vav-1. elt-6 has genetic interactions with egl-18 (positive in our screen) in other contexts. It showed a low-penetrance Lep phenotype when the RNAi experiment was repeated and more males were scored. In addition, elt-6(RNAi) enhances the Lep phenotype of egl-18(ok290) mutants (Table S2). Also, the network model predicts that vav-1 has interactions with php-3, cdc-42 and inx-12 (all positives in our screen, Figure 5). Although vav-1 treatment by RNAi via feeding did not result in detectable phenotypes, administering RNAi against vav-1 by soaking did cause tail tip defects (data not shown).
We next asked whether the network model—developed from information in other systems—has biological relevance for tail tip morphogenesis. To this end, we tested a selection of the predicted interactions by genetic and expression epistasis analyses. The results are detailed below.
Here, we used systemic RNAi to identify components that are involved in male-specific morphogenesis of the tail tip of C. elegans. RNAi via feeding in C. elegans provides a simple yet powerful means for identifying the regulatory and structural components and pathways of developmental processes [28], [62]. The methodology we employed (Materials and Methods and Figure 1B) allowed us to quickly score for subtle defects in morphogenesis at high magnification. Many of the genes we identified have roles in embryonic processes and are lethal when knocked down (e.g. nob-1, cdc-42). This justifies our approach to treat worms with RNAi postembryonically and it underscores the power of RNAi as a tool for identifying postembryonic roles of genes that have essential embryonic functions. To minimize the number of false negatives, we performed the screen on the RNAi hypersensitive strain, rrf-3 [63]. For a number of reasons, however, we believe that there are still other tail tip genes to be identified. First, our screen did not cover the entire genome (approx. 83%). Second, because of complete larval lethality, about 2% of the genes in our screen were not scored, including the known tail tip regulator lin-41 [6] (Table S4). Third, of the previously known tail tip genes, only lin-44, which encodes for the Wnt ligand, was found in the screen. Other known tail tip genes, tlp-1 and dmd-3, which have representative clones in the library, were missed. Finally, two genes (elt-6 and vav-1) not found in the screen, but predicted by the N-Browse network analysis, turned out to have RNAi-induced tail tip phenotypes when treated in a different genetic background (i.e. elt-6 RNAi in the egl-18(ok290) background) or by soaking instead of feeding (vav-1). The number of false positives is likely to be very small since only genes which were positive in the primary and secondary screen were considered. The 212 candidate genes identified in this process were significantly enriched with morphogenesis-related GO attributes relative to the genome at large, consistent with what would be predicted if our screen were successful. Some candidates were studied further to elucidate their roles in regulating or effecting tail tip morphogenesis.
Developmental decisions, such as the initiation of morphogenesis, require the input of multiple signaling pathways and result in a coordinated response by many different components of the cell. The response must be robust against perturbations from the internal and external environment. Robustness and precision of biological processes are thought to be facilitated by a bow-tie (or hourglass) architecture of the gene regulatory network [18], [19]. Characteristics of bow-tie networks include the following. (1) Many inputs and outputs are connected to a conserved core. (2) Versatile weak linkages form the interface between input and core and between core and output. (3) Systems control is facilitated by positive and negative feedback at every level. (4) Modularity and partial redundancy or degeneracy are two other properties of the bow-tie network architecture that contribute to the robustness of biological systems [19], [64].
We used the data about gene interactions described here in combination with published information to delineate a first draft of the gene regulatory network underlying tail tip morphogenesis in C. elegans males. Although the reconstruction of this network has only just begun, we already find many features that are consistent with bow-tie architecture.
A network of interactions is called modular if it can be subdivided into relatively autonomous components (modules) that are built of highly connected parts but are more loosely connected to other modules [65]. Modularity is a major contributor to the robustness and evolvability of a system, since perturbations and mutations can occur within a module with minimal effects on the whole system [19]. It has been proposed that modularity facilitates evolutionary change by allowing new connections to be made between modules without disrupting the core function of the modules [66]. Modularity has been observed in many networks [67], [68]. The modules of metabolic networks have bow-tie structure, just like the networks themselves [21]; that is, bow-tie architecture can be nested.
In the gene regulatory network of male tail tip morphogenesis, we find evidence for many conserved regulatory and effector modules. Regulatory modules include Hox patterning, the sex-determination and heterochronic pathways and TGF-β signaling. We also identified tail tip roles for additional Wnt pathway components, i.e. the SYS-1 beta-catenin, the MIG-1 Frizzled-like receptor, and the LIT-1 Nemo-like MapK. Effector modules consist of conserved components controlling vesicular trafficking and endocytosis, establishment of cellular polarity, cytoskeletal rearrangement, and cellular fusion.
Degeneracy describes the coexistence of structurally or mechanistically different components that can perform similar roles or are interchangeable under certain conditions [64]. Degeneracy confers robustness because, in a system composed of partially redundant elements, one element can compensate for the failure of another. One mechanism that generates degeneracy is gene duplication. In the male tail tip morphogenesis system, we find several examples for degeneracy due to gene paralogy. We think that MAB-3 and DMD-3 function partially redundantly because the phenotype of the mab-3(e1240);dmd-3(ok1327) double mutant is more severe than that of dmd-3(ok1327) or mab-3(e1240) alone [7]. Similarly, two other pairs of paralogs function semi-redundantly: the Hox genes php-3 and nob-1 and the GATA transcription factors egl-18 and elt-6, which form an operon. In both cases, removal of both genes results in a much more severe disruption of morphogenesis than removal of only one gene.
Both modularity and degeneracy contribute in a major way to the robustness of a system [19]. Indeed, the majority of RNAi knockdown phenotypes suggest that male tail tip morphogenesis is very robust against genetic perturbations. In most cases, the effects of RNAi (as well as some of the mutations tested) were subtle and the penetrance was low, suggesting that there is extensive buffering of the system against partial depletion of individual transcripts (or reduced functionality of genes).
The architecture of the genetic network regulating male tail tip morphogenesis in C. elegans is congruent with the bow-tie model, since we find evidence for all the characteristics of bow-tie networks. We find modularity, degeneracy, a conserved core, weak linkage and positive and negative feedback loops connecting spatial, sexual and temporal inputs to the cellular responses required for morphogenesis. To our knowledge, this is the first time that a genetic network regulating a morphogenetic process has been specifically investigated for bow-tie architecture. However, it is likely that other morphogenetic events—e.g., the development of the eye in flies and possibly mammals and the formation of the pharynx in C. elegans—are also controlled by bow-tie regulatory networks. In both examples, components have been identified which are likely to be part of the conserved core. Drosophila eye development is in part controlled by the products of eight eye specification genes. Deletion of either one of these genes leads to a drastic reduction or loss of the adult eye, whereas ectopic expression of all but one results in retinal development outside of normal eye tissue [88], [89]. The fly eye specification genes are conserved with orthologs in mammals. Expression of one of them, dachshund, is regulated by at least 36 upstream factors, including the TGF-β signaling pathway, the transcription factor Zerknüllt and several other patterning genes (e.g. krüppel, snail and dorsal) [88], suggesting the existence of an extensive input fan in this system. The FoxA transcription factor PHA-4 is a central regulator of pharynx development in C. elegans. pha-4 is the only zygotic gene that deletes the entire pharynx when mutated [90]. FoxA transcription factors are conserved from Cnidaria to mammals and are always associated with the digestive tract [90]. PHA-4 has hundreds of targets, many of which are directly regulated at the transcriptional level [91], [92]. In this system, feedback loops have been identified as well [90]. Thus, there is evidence for a conserved core, an output fan and system control as elements of a bow-tie network architecture for pharynx morphogenesis. Finding bow-tie networks in multiple developmental systems supports the notion that this architecture is a universal feature of evolved gene regulatory networks and is favored by selection due to its robustness. The male tail tip thus provides a simple model for investigating not only morphogenetic mechanisms, but also the properties of a universally important genetic regulatory architecture.
Genetic manipulations and culturing of C. elegans were performed as previously described [98]. We use the following nomenclature for transgenes (similar to that used by Ziel et al. [99]). Transcriptional reporters are designated by the name of the gene associated with the promoter, followed by a “>” and the reporter gene to which it is fused (e.g., dmd-3>yfp). Translational reporters are designated by the gene, followed by “::” and the reporter to which it is fused (e.g., sma-3::mCherry). Unless otherwise stated, the endogenous promoter is used to drive expression of translational reporters. If a different promoter is used, we use both designations (e.g., lin-44>php-3::gfp represents the php-3 gene fused to the gfp gene, with expression driven by the promoter of the lin-44 gene). Unless otherwise stated, the unc-54 3′UTR is used for all constructs. The protein product of a construct is designated with capital letters (e.g., PHP-3::GFP). No construct employed cDNA; all introns were included.
Strains with transgenes generated for this paper are listed in Table S3. Other strains used for this study are listed below.
CB4088 = him-5(e1490)V. This is the otherwise wild-type, male-producing strain used as the background genotype in this study (from Caenorhabditis Genetics Center, CGC).
BW2020 = ctIs57[nob-1::gfp + rol-6] (a gift from Zhongying Zhao, University of Washington, Washington). Hermaphrodites of this strain were crossed with CB4088 males to obtain a him-59(e1490)V; ctIs57 strain.
DF125 = php-3(ok919)III; him-5(e1490)V. Made by crossing CB4088 males with RB998 hermaphrodites. RB998 = php-3(ok919)III (from CGC).
DF159 = rme-8(b1023)I; him-5(e1490)V. Made by crossing CB4088 males with DH1206 hermaphrodites. DH1206 = rme-8(b1023)I (from CGC).
DF160 = blmp-1(tm548)I; him-5(e1490)V; fsIs3[dmd-3>yfp + unc-122>gfp].
DF161 = blmp-1(tm548)I; him-5(e1490)V. Made by crossing CB4088 males with hermaphrodites carrying the blmp-1(tm548) allele (from Shohei Mitani, National BioResource Project, Tokyo Women's Medical University School of Medicine, Tokyo, Japan).
DF163 = sma-3(e491)III; him-5(e1490)V. Made by crossing CB4088 males with CB491 hermaphrodites. CB491 = sma-3(e491)III (from CGC).
DF164 = egl-18(ok290)IV; him-5(e1490)V. Made by crossing CB4088 males with JR2370 hermaphrodites. JR2370 = egl-18(ok290)IV (from CGC).
DF167 = him-5(e1490)V; nhr-25(ku217)X. Made by crossing CB4088 males with MH1955 hermaphrodites. MH1955 = nhr-25(ku217)X (from CGC).
DF171 = him-5(e1490)V; bIs34[rme-8::gfp + rol-6]. Made by crossing CB4088 males with DH1336 hermaphrodites. DH1336 = bls34[rme-8::gfp + rol-6] (from CGC).
DF177 = him-5(e1490)V; nhr-25(ku217)X; fsIs3[dmd-3>yfp + unc-122>gfp].
DF178 = mab-3(e1240)II; dmd-3(ok1327) him-5(e1490)V; bIs34[rme-8::gfp + rol-6].
DF196 = him-5(e1490)V; xnIs8[pJN343: nmy-2::mCherry + unc-119(+)]. This strain was made by crossing CB4088 males to hermaphrodites carrying the transgene xnls8 which were a generous gift from Jeremy Nance (NYU Skirball Institute, New York, New York).
DF197 = mab-3(e1240)II; dmd-3(ok1327) him-5(e1490)V; xnIs8[pJN343: nmy-2::mCherry + unc-119(+)].
DF199 = ptl-1(ok621)III; him-5(e1490)V. Made by crossing CB4088 males to RB808 hermaphrodites. RB808 = ptl-1(ok621)III (from CGC).
JJ1473 = unc-119(ed3)III; zuIs45[nmy-2::gfp + unc-119(+)] (from CGC).
KC447 = rrf-3(pk1426)II; him-5(e1490)V. A generous gift from King L. Chow (Hong Kong University of Science and Technology, Hong Kong, China).
KC529 = eri-1(mg366)IV; him-5(e1490)V. (from K. L. Chow).
UR157 = fsIs2[dmd-3>yfp + unc-122>gfp]I?; him-5(e1490)V. A generous gift from Douglas Portman (Rochester University, New York).
UR279 = mab-3(e1240)II; dmd-3(ok1327) him-5(e1490)V (from D. Portman).
WM79 = rol-6(n1270)II; neEx[lit-1::GFP + rol-6(su1006)] (from CGC). Hermaphrodites of this strain were crossed with CB4088 males to obtain a rol-6(n1270)II; him-5(e1490)V; neEx[lit-1::GFP + rol-6(su1006)] strain.
The genome-wide RNAi-feeding screen was carried out in the RNAi-hypersensitive background rrf-3; him-5 (strain KC447). RNAi effects on embryogenesis were bypassed by feeding siRNA-expressing bacteria to synchronized L1 larvae. Following a recommendation by K. Chow (pers. comm.), L1 larvae were plated onto a thin film of agar (1.5 ml per 60 mm plate) containing 2 mM IPTG and 100 µg/ml ampicillin. Worms were cultured to adulthood (3 days) at 20°C at which time a square of agar with worms was removed and mounted directly onto a glass slide and covered with a coverslip (Figure 1A). All scoring was carried out at 400x with a Zeiss Axioskop equipped with Nomarski differential interference contrast. Images were recorded with a C4742-95 “Orca” Hamamatsu digital camera and Openlab software, ver. 3.0.9 (Improvision). The secondary screen was carried out in the same way but in a different RNAi hypersensitive background, eri-1 (strain KC529). RNAi clones consistently conferring a Lep or Ore tail tip phenotype were sequenced to confirm the targeted genes. All scoring data and images are available via our male tail tip database, MTTdb, at http://wormtails.bio.nyu.edu.
Translational fusions and transcriptional reporters were constructed by overlap-extension PCR as previously described [100]. The 5′ upstream sequence and coding sequences for php-3 (-500 bp to the stop codon), egl-18 (-3691 to stop), rcn-1 (-4797 to stop), ptl-1 (-2105 to stop), arl-1 (-550 to stop), abcx-1 (-437 to stop), cdc-42 (-1814 to stop), sys-1 (-3411 to stop), inx-12 (-3975 to stop), blmp-1 (-4916 To stop), nhr-165 (-1559 to stop), and ubc-12 (-363 to stop), were PCR-amplified from genomic DNA and fused to gfp and the unc-54 3′-UTR amplified from pPD95.75 (Addgene). The 5′-upstream sequence and coding region of sma-3 (from -1169 bp), amplified from genomic DNA, was fused to mCherry [101] and to cfp, which were PCR-amplified from pGC326 (a gift from E. J. Hubbard) and pPD136.61 (Addgene), respectively. To make the DAF-4 reporters, the upstream sequence and coding region (from -5091) was fused to yfp that was amplified from pPD136.64 (Addgene). A transcriptional reporter of daf-4 fused the upstream sequence (-4865 to -1) to the NLS and gfp amplified from pPD122.13 (Addgene). The transcriptional reporter for nhr-25 fused the upstream sequence (-9100 to -1) to the NLS and gfp from pPD122.13 (Addgene) followed by the 3′-UTR for nhr-25 (stop to +760 bp). Transcriptional reporters for inx-12 (-3975 to -1) and inx-13 (-1557 to -1) were fused to yfp (pPD136.64 (Addgene)) and cfp (pPD136.61(Addgene)).
Transgenes were microinjected at concentrations ranging from 5–20 ng/µl along with 100 ng/µl pRF4 (rol-6(su1006)) as injection marker. Multiple lines were analyzed for each construct using epifluorescence (Axioskop with mercury lamp, 400 or 1000x). Representative images of fluorescence expression patterns are available via the MTTdb database at http://wormtails.bio.nyu.edu. Strain names and primer sequences are provided in Table S3.
Genes identified in our screen were manually entered into N-Browse2 (http://Aquila.bio.nyu.edu/NBrowse2/nbrowsetest.jnlp) [17]. Only a subset of these genes showed annotated interactions, and only those with an interaction one or two edges away from another candidate gene were added to our network (Figure 5). Information from other studies [6], [7], [44], [51], [57]–[60] which show genetic or direct interactions with known or newly identified tail tip genes were also included (Figure 5).
|
10.1371/journal.pcbi.1005475 | Mechanisms underlying different onset patterns of focal seizures | Focal seizures are episodes of pathological brain activity that appear to arise from a localised area of the brain. The onset patterns of focal seizure activity have been studied intensively, and they have largely been distinguished into two types—low amplitude fast oscillations (LAF), or high amplitude spikes (HAS). Here we explore whether these two patterns arise from fundamentally different mechanisms. Here, we use a previously established computational model of neocortical tissue, and validate it as an adequate model using clinical recordings of focal seizures. We then reproduce the two onset patterns in their most defining properties and investigate the possible mechanisms underlying the different focal seizure onset patterns in the model. We show that the two patterns are associated with different mechanisms at the spatial scale of a single ECoG electrode. The LAF onset is initiated by independent patches of localised activity, which slowly invade the surrounding tissue and coalesce over time. In contrast, the HAS onset is a global, systemic transition to a coexisting seizure state triggered by a local event. We find that such a global transition is enabled by an increase in the excitability of the “healthy” surrounding tissue, which by itself does not generate seizures, but can support seizure activity when incited. In our simulations, the difference in surrounding tissue excitability also offers a simple explanation of the clinically reported difference in surgical outcomes. Finally, we demonstrate in the model how changes in tissue excitability could be elucidated, in principle, using active stimulation. Taken together, our modelling results suggest that the excitability of the tissue surrounding the seizure core may play a determining role in the seizure onset pattern, as well as in the surgical outcome.
| Much attention has been devoted to the mechanisms underlying epileptic seizures. However, so far, the morphology of how seizures start on electrographic recordings (i.e. the seizure onset pattern) has been neglected as a potential indicator of the underlying dynamic mechanism. In this work, we take a spatio-temporal modelling approach to reproduce and understand two major seizure onset patterns. We find that it is not necessarily the initiation in the seizure core that determines onset pattern, but that the excitability of the surrounding “healthy” tissue plays a pivotal role in how the seizure onset appears. In agreement with previous computational modelling work, we hypothesise that indeed the two patterns could arise due to different dynamic onset mechanisms, where the surround excitability differs fundamentally in their dynamic properties. Our hypothesis indeed also offers a simple explanation for the clinically reported difference in surgical outcome for the two onset patterns. As an outlook, we also propose a possible way to track such changes in the surround excitability in a proof-of-principle computational demonstration.
| Focal seizures are episodes of highly disruptive brain activity, which are considered to arise from local sites of pathological abnormality in the brain. Identification of the precise focal origin in a given patient is crucial for the clinical management of their epilepsy. Various clinical evidence can point to the nature and site of the origin, but critical amongst these are the data recorded by electroencephalography (EEG) or by the invasive alternative of electrocorticography (ECoG). Previous clinical studies have suggested that focal seizures could have different onset mechanisms [1–4]. Related to this, there is a clear heterogeneity in clinical outcomes: only a subset of patients responds to medications [5], or surgical removal of the presumed epileptic focus [6–8].
In terms of the appearance and morphology of focal seizure activity, generally two types of common onset patterns are reported: The low amplitude fast (LAF) activity, which is characterized by oscillations in the beta to gamma range of initially low amplitude that slowly increases as the seizure progresses; and the high amplitude slow (HAS) activity, which is generally described as a slower oscillation below the alpha range with a high amplitude at onset [9–18]. Depending on the study, and subtype of epilepsy, finer distinctions in the onset pattern have been characterized, and different quantitative measures have been used to categorize the onset patterns. For example, many studies use a cut-off frequency in the alpha band to distinguish between LAF other onset patterns [10, 11, 16, 17], but the choice generally varied between 8 Hz and 20 Hz.
The majority of studies find that the LAF pattern is most frequently occurring pattern [10, 14, 16–18]. Intriguingly, despite different categorization criteria, and subtypes of epilepsy included, it appears that the LAF pattern is associated with a good surgical outcome when removed [9, 10, 12, 14, 16, 18]. Depending on the study, and subtype of epilepsy considered, a potentially worse surgical outcome is implicated for the HAS compared to the LAF pattern [10, 12, 14, 16]. However, the literature is not entirely consistent; such a state usually reflects difficulties arising from limited clinical information, perhaps requiring more precision either in the classification of the diagnosis, or of the nature of the onset pattern [19, 20].
Furthermore, the LAF pattern is associated with a larger seizure onset zone (in both temporal and extratemporal lobe seizures), which is measured by the number of contacts first showing epileptic activity [11, 17]. Additionally, the cortical excitability as measured by cortico-cortical evoked potentials (CCEPs) is lower in the seizure onset zone of the LAF pattern compared to the HAS pattern [15]. The evidence presented here indicates distinct mechanisms generating the different onset patterns. However, to our knowledge, no spatio-temporal interpretation exists so far explaining the basis of these distinct mechanisms.
To provide a mechanistic explanation, we turn to our previously suggested theoretical framework, which demonstrated the possible heterogeneity of onset mechanisms in focal seizures [21]. Our model utilised established concepts of increased excitability [2, 22], impaired inhibitory restraints [23–25], bistability and triggers [26–28] to provide a classification of dynamic mechanisms of focal seizure onset. We now show that two of the distinct classes of dynamical mechanisms described in our earlier study appear to correspond to the clinical LAF and HAS patterns.
To begin the investigation of seizure onset patterns in the model, we first show that the model is capable of reproducing the clinical observations of low amplitude fast (LAF), and high amplitude slow (HAS) patterns. As the clinical categorisation of the onset patterns differ considerably across different studies, we focused on the amplitude of the pattern in our model as a main aspect to reproduce. Most studies agree that the LAF pattern is a very low amplitude activity compared to other patterns, which gradually grows over time. Interestingly, as we shall see, the LAF pattern is intrinsically associated with a higher frequency of oscillation than the HAS pattern in our model.
The LAF onset is shown in Fig 1A. In agreement with the clinical recording, the seizure onset on the simulated ECoG electrode (which is placed over the entire simulated cortical sheet) in the model arises out of the background state with low amplitude fast oscillations (~14 HZ). Spatio-temporally, the LAF activity on the simulated ECoG is caused by small patches of microdomains displaying localised oscillatory microseizure activity. As the simulated seizure progresses, more microdomains become involved, and the surrounding tissue is recruited slowly. In other words, the microdomains coalesce slowly and form larger contiguous and coherent patches of seizure activity. The amplitude of the simulated ECoG average activity also grows. Eventually, the entire simulated tissue is recruited into the seizure activity. The slowing down in frequency of the simulated ECoG oscillations as the seizure progresses is a direct consequence of the recruitment process. The oscillatory frequency of seizure activity in smaller patches is intrinsically faster than the frequency of a large contiguous piece of tissue.
The high amplitude onset is shown in Fig 1B. Also in agreement with the clinical recording, the seizure starts with a HAS oscillation on the simulated ECoG. The oscillation starts relatively slow (~9 Hz) compared to the LAF pattern. On the spatio-temporal view, it becomes clear that the initiation and subsequent evolution of this HAS pattern is very different from that of the LAF onset in the model. Upon activation of one patch of localised seizure activity, this activity invades the surrounding tissue rapidly as a propagating wavefront. Some long-range activation also appears subsequently through the long-range connections in the model cortical sheet. The entire simulated cortical sheet is fully recruited within 0.8 s. In contrast to the LAF onset patterns, the low frequency at onset of HAS onset patterns remains fairly constant, even throughout the initial evolution phase (first 1-2 seconds), similar to the clinical observation.
We also want to highlight here that the model reproduces not only the initial onset pattern, but to some extent also the evolution of the seizure. For example, the growing amplitude in the LAF pattern is shown in our model, which after several seconds develops into a high-amplitude spiking activity. However, our model does not simulate the termination of the seizure. Hence, we are only proposing to investigate the mechanism of onset and initial evolution of the seizure. The mechanism of the complete seizure development and eventual termination may be driven by processes on longer time scales [29, 30] which we leave as an investigation for a future study.
After confirming that the model is indeed capable of reproducing both onset patterns qualitatively, we turn to a systematic study of the conditions that lead to either pattern.
In our previous work [21], we demonstrated that focal seizure onset can be understood in terms of the spatial regions of seizure activity, and surrounding non-seizing tissue. In particular, we showed the different classes of seizure onset were associated with a different excitability setting of the surrounding territory. Hence, we choose to investigate surround excitability in terms of onset patterns. Here, the term excitability refers to the proximity to the monostable seizure state in the parameter space (see Methods for details). Changes in the excitability levels of the surround can be caused by a range of parameters (e.g. reductions in the inhibitory restraint). In the following, we choose to investigate the parameter P (constant input to the excitatory population). Equivalent results hold for other parameters that induce bistability (e.g. parameter Q, see S1 Fig. compared to Fig 2).
We also demonstrated the importance of the spatial organisation of the seizure core in our previous work [21]. In other words, the number and size of microdomains initiating the microseizure activity also influences the overall dynamics at seizure initiation. For example, 10% of minicolumns displaying microseizure activity can either all be spatially arranged into one large patch, or ten separate smaller patches. This example is illustrated in Fig 2C in the top two panels. Therefore, we shall also study the effect of size and density of the microdomains on seizure onset patterns.
Fig 2A shows the resulting scan of seizure amplitude (illustrated in the size of the dots) with respect to the three parameters (i) surround excitability (Psurround), (ii) total percentage of minicolumns displaying microseizure activity, and (iii) number of microdomains. The high amplitude onset patterns (big dot sizes) are all found for high values of excitability of the surrounding tissue (Psurround). To study this further, we plotted the distribution of seizure onset amplitudes for each value of excitability of the surrounding tissue (Psurround) in Fig 2B. High amplitude onset patterns (with amplitudes higher than 0.5) are generally found in areas of high surround excitability (Psurround > −2). When the surround excitability drops, the onset amplitude is generally below 0.5. Indeed, the parameter region supporting high amplitude onset patterns overlaps very well with the parameter region supporting bistable dynamics (between −2.5 < Psurround < −1.1), i.e. the seizure and background states coexist in this parameter region (Fig 2B). The total percentages of minicolumns with microseizure activity, and the number of patches they form, introduce some variation in the onset amplitude (Fig 2B). Thus we decided to study their effect in two scenarios of the surround excitability: one where the surround is monostable in the background state, and one where the surround is bistable.
For Psurround = −1.5 (Fig 2C, top, green frame), the surrounding is in the bistable state. An increase in the total percentage of minicolumns with microseizure activity would lead to a faster recruitment of the surrounding. Therefore, we expect the increase in the total percentage of minicolumns with microseizure activity to generally lead to an increase in onset amplitude, which is exactly the case (onset amplitude increases from about 1 to 2.4). By the same argument, if the minicolumns with microseizure activity are arranged into more patches, this essentially means multiple initiation points for recruitment. Hence we expect the total recruitment to be faster, and the onset amplitude to be higher, which again is confirmed in the scan (onset amplitude increases by about 0.6 on average with increasing number of patches). For Psurround = −4 (Fig 2C, bottom, blue frame), the surrounding is in the monostable background state. In this scenario, recruitment of the surrounding is not induced easily. At the initial phase of the seizure, merely the microdomain oscillatory activity is detected on the ECoG. Therefore, the increase in total percentage of minicolumns with microseizure activity leads to a slight increase in onset amplitude (by about 1.5 when only one patch is present). Conversely, if the minicolumns with microseizure activity are arranged into several patches, their oscillatory activity is not coordinated and on average would appear with smaller amplitude than a single patch. Note that both effects are smaller in amplitude than the described effects for the bistable scenario, i.e. Fig 2C top and bottom figures have different colour axis ranges. Finally, as already shown in Fig 1, the low amplitude onset is generally associated with a slightly higher frequency than the high amplitude onset patterns. This can be seen again in Fig 2B. In the parameter region of increased excitability, the frequency of onset is significantly decreased. Again, this is due to the intrinsic properties of oscillatory domains. Large contiguous domains of seizure activity show a lower oscillation frequency in the model than smaller domains.
We conclude that overall the high amplitude slow (HAS) onset pattern is mostly strongly associated with an increased excitability of the surround leading to bistability of background and seizure state, and the low amplitude fast (LAF) onset pattern occurs in the monostable background setting. The triggers of the seizure onset in both cases are localised microdomains displaying increases in activity. The amount of microdomains and the spatial arrangement of these can influence the onset amplitude to some degree. However, the dominant parameter for the seizure onset amplitude is the excitability level of the surrounding tissue.
To investigate the consequence of the insight that altered surround excitability might underlie different onset patterns; we simulate the impact of surgical resections in the model for both onset patterns. This investigation is motivated by the clinical indication that the low amplitude onset pattern might be associated with a better post-operative outcome in focal seizures [9, 10, 12, 14, 16, 18]. In Fig 3 we show the evolution of both onset patterns before and after a simulated surgery. The in silico surgery in this case is performed by resecting 20% of the cortical sheet, which includes the area from where the onset starts. In this case it is an area of particularly high microseizure density. In order to ensure a fair comparison, we have chosen the same location and dynamics for the microseizure domains in both onset patterns. The only difference between the two scenarios is the surround excitability. In the case of the high amplitude onset, the surrounding is chosen to be bistable. In the scenario of low amplitude onset, the surrounding is chosen to be monostable in the background state.
Fig 3A and 3B show both onset patterns as we would expect before the simulated resection. After the simulated surgery, in the case of the LAF onset pattern (Fig 3C), seizure activity is still seen in some microdomains, but these no longer recruit the surrounding. Only a very low amplitude fast oscillation remains visible on the ECoG. In case of complete resection of all microdomains, the tissue would not show any remaining seizure activity. However, for the HAS patterns (Fig 3D), the simulated surgery removed the original trigger of the seizure, but does not stop full recruitment of the remaining tissue, as other microdomains are now equally able to trigger recruitment. We conclude that in our model, the high amplitude onset pattern is also associated with a worse surgical outcome. This is due to the increased excitability of the surround leading to bistability, which after surgery enables recruitment from alternative triggers. In the low amplitude onset pattern, the increased excitability in the surround is lacking. Therefore, the removal of the seizure onset zone is sufficient to prevent full recruitment. Our result is to be seen as supportive evidence that LFA might be associated with better surgical outcome. However, as described in the introduction, the clinical picture is still unclear due to the differing definitions of onset patterns, and subtypes of epilepsy investigated.
Finally, we suggest a perspective for possible alternative treatment options, especially in the case of seizures involving a high excitability level in the surround. As we have demonstrated, surgical removal might not be sufficient as a treatment for patients suffering from this type of seizure onset. The main challenge is the “global” bistability, where the seizure state coexists with the background state. This might not be easily treated by removal of tissue, but rather requires a “global” treatment at the correct time. (See Discussion for more details on the possible spatial extent of such bistabilities). Future studies have to elucidate the exact extent of such areas of increased excitability. If the spatio-temporal extent of such areas is known, this knowledge can be used to either surgically treat the affected tissue (e.g. multiple subpial transections), or for example to inform therapeutic brain stimulation devices when and where to stimulate.
One inherent problem of inferring where and when these increased cortical excitabilities occur is that these changes might not be observable by passive recordings (of ECoG, LFP, or even neuronal firing). This is because the increased cortical excitability can create bistable states, but the pathological abnormality is latent and invisible electrophysiologically. Hence, in order to infer when and where these bistabilities might exist, it is necessary to track changes in excitability spatio-temporally. One way to achieve this is by active probing. Fig 4 shows a proof of principle simulation, where we track the changing global excitability level over time using stimulation responses.
We simulated an entire cortical sheet undergoing changes in excitability levels over time (Fig 4A). The resulting ECoG recording is shown in Fig 4B. No obvious observable difference can be found for different levels of excitability. We therefore analysed the stimulation response amplitudes, as it is an established clinical way of measuring excitability levels [15, 31]. In 9 locations on the sheet we simulated stimulation (affecting one local minicolumn) using a transient increased input to the excitatory populations (parameter P) for 6 ms. The result of one of the 9 stimulation sites is shown in Fig 4C. We then measured the maximal stimulation response, and averaged this over the 9 stimulation sites, giving us an average stimulation amplitude (Fig 4D). This average stimulation amplitude was found to track the underlying changes in the excitability levels (Pearson correlation coefficient: 0.8246, p < 0.001).
In summary, we conclude that the two focal seizure onset patterns are distinguished by differing tissue excitability in our model. The high amplitude onset pattern is associated with an increased surround excitability that creates the coexistence (bistability) of a seizure state in the surrounding tissue. Conversely, the low amplitude onset pattern requires domains displaying localised seizure activity, which are embedded in a surrounding that is not bistable. As a direct consequence of this, the model shows that local removal of (epileptogenic) tissue can prevent recruitment and a full-blown seizure in low amplitude onset patterns, but not necessarily in high amplitude patterns. As an outlook, and to address the question of what alternative to removal of tissue one can pursue to prevent seizures in the case of a bistable surround, we propose the idea of tracking surround excitability levels to enable the operation of targeted intervention devices.
In our model we were able to characterize two main types of onset pattern, the low amplitude fast pattern, and the high amplitude slow pattern. These two patterns agree qualitatively with seizure onset patterns described in the clinical literature in their initiation and evolution of the activity amplitude, and also to some extent in frequency of the seizure activity. However, the clinical literature highlights a range of different onset pattern morphologies with a high amplitude at onset, e.g. high amplitude spikes, sharp waves, or high amplitude spike-and-wave activity, etc. [17, 18]. In our model, we did not investigate the possibility of generating the different subtypes of a high-amplitude onset pattern, as we were focusing on the difference in onset pattern introduced by a bistable surround. Nevertheless, it is worth mentioning that this is not an intrinsic limitation of the model, as for example we and others have shown in earlier studies that more complex seizure waveforms can be generated even in “simple” neural population models such as the one used here (e.g. [28, 29, 32, 33]). Hence it is not inconceivable that more complex onset patterns could be found in the model with the right parameter (time scale) settings. We leave this investigation for a future study.
In terms of seizure onset, the diffuse electrodecrement or attenuation onset pattern deserves attention, as it has also been shown to correlate with a worse surgical outcome [16, 18]. Currently, to our knowledge, neither our model, nor any other model has provided a mechanistic insight regarding this pattern. Particularly, as it may not only be a seizure onset pattern, but can occur in the middle of a seizure [34]. The spatial scale of this pattern is also generally diffuse, often involving most of the recording electrodes [18], making it a difficult event to study for our model. However, we speculate that the diffuse nature may indicate a wide-ranging abnormal surround. If that is the case, the abnormal surround may be the real driver behind the seizures, and local removal of the apparent seizure onset zone may not be indicated. To extend our understanding of this pattern further, a model of different spatial scale (whole brain scale, e.g. [35–37], or at the scale of an ECoG grid, e.g. [38–40]) is perhaps required, potentially also requiring subcortical structures (e.g. [41, 42]).
We noted that the faster frequency for the low amplitude onset pattern is an intrinsic property of independently oscillating patches of tissue in the model. This is not only an intriguing phenomenon in the model, which deserves further investigations, but might also suggest an alternative mechanism by which frequency changes during a seizure can be explained. The exact frequency range for low amplitude fast activity at seizure onset is so far inconsistently defined between studies. For example some studies call all frequencies above 8 Hz as fast activity [16], while other studies use a cutoff frequency of 13 Hz [10], or even higher [14]. However, it appears that all studies refer to oscillations in the alpha/beta band. In that regard it might be interesting to clarify clinically 1) the amplitude and frequency range of these fast onset activities, 2) their relationship to the background activity, and 3) how they evolve over time, to enable future modelling efforts in this direction.
It is also an open question how these fast activities in the beta band relate to high-frequency oscillations (HFOs) at seizure onset (gamma range or beyond, see e.g. [32]). Multiple suggestions for the cellular mechanisms of HFOs exist, and their role in epilepsy is also being discussed [43]. High frequency oscillations phase-locked to the lower frequency band (1-25 Hz) at seizure onset have been proposed recently to be related to the multi-unit activity of neural populations [44, 45]. In an adequate translation of our model (converting the relative firing rates to multi unit activity and their HFO signatures), these phase-locked HFOs could also be observed in our model output. This is because the firing rates increase at seizure start, which is directly translated to the LFPs we show in our figures. In our model an increase in HFO would thus be observed in both onset patterns, perhaps with a lower power in the low amplitude onset pattern. Perucca et al. indeed observe that the HFO rate increases in general at seizure onset, independent of the onset pattern [17]. However, this will deserve a closer examination and direct simulation in order to confirm that our model may be able to show HFOs at seizure onset. Furthermore, when distinguishing between ripples (80-200 Hz) and fast ripples (200-500 Hz) in the HFOs, Levesque et al. show that the distribution of the two types of HFOs differs between different types of seizure onset in a rat model of temporal lobe epilepsy [46]. The authors find a higher ripple rate associated with the LAF pattern, and a higher fast ripple rate associated with the HAS pattern. Future studies (possibly with more detailed neuronal models) will show if these observations can be related to our hypothesis of differences in surround excitability and onset mechanism.
The choice of our spatio-temporal model for this study is crucial, as we needed to investigate the spatial interaction of epileptogenic patches with their surround. This was possible, as our previous study [21] showed a natural way of understanding surround excitability/bistability and the incorporation of epileptogenic patches. Furthermore, the model is computationally efficient given the volume of tissue simulated, and the level of abstraction of the model has proven sufficient in capturing the onset patterns in their most obvious qualitative features. However, we acknowledge that other details may prove important to capture more details and mechanisms of onset patterns in future studies. There is still much work to do in future to achieve a full understanding of what generates different onset patterns spatio-temporally, and our work can serve as a starting point for future in silico approaches, as it naturally links key concepts (tissue excitability, recruitment, and core/surround) with observations of onset patterns.
So far, we have only commented briefly on the possible spatial extent of the surround with increased cortical excitability (which we termed “global” increase in excitability). Badawy et al. measured increases in cortical excitability by the motor cortex threshold, and showed that such increases can be detected even in focal seizures with the seizure onset zone (SOZ) not in the motor cortex [2]. Hence, this suggests that the spatial extent of areas with increased excitability may involve widespread networks, perhaps even an entire hemisphere. However, [15] showed that the CCEP response amplitude was significantly lower in electrodes far away from the SOZ, suggesting that the area of increased excitability could be limited to the coverage of an ECoG grid. Furthermore, [47] showed that delayed responses to single pulse electrical stimulation are mainly seen in the SOZ, suggesting that this area has altered excitability levels. Further studies are required to explicitly map the spatio-temporal evolutions of the (abnormal) shifts in excitability on the cortex.
Interestingly, Enatsu et al. also found that in both onset patterns, the excitability of the SOZ is increased compared to a location of the ECoG that is far away from the SOZ [15]. This suggests that an increase in the excitability of the surround might occur in both onset patterns. This is also in agreement with findings that motor thresholds are consistently altered in focal seizure patients and not just a subset of patients [2]. In the model this observation is also not surprising. Tissue with localised patches of microseizure activity would also display a slight increase in stimulation response, but not as much as compared to the bistable surround. However, this also needs to be accompanied by improvements in clinical recordings to provide a higher resolution spatio-temporal clinical picture, over more extended areas. Additionally, computational models that span the spatial scales from minicolumns to an entire ECoG grid may be beneficial to tease apart the different spatial scales of excitation that could occur.
The source and mechanism causing the increased surround excitability is also unclear. In both patients with acquired structural epilepsies and their siblings, an increase in motor cortex excitability could be demonstrated [48]. This suggests that the increased cortical excitability is not necessarily a pathological feature, but rather to be understood as an enabling, pro-ictal condition. The study also suggests that genetic factors might underlie the increased excitability.
In cases of lesional epilepsy, tumors, or congenital dysplastic lesions, the surrounding tissue is generally also considered abnormal, and is therefore also resected. The increased surgical success rates in such cases [49] indicate that resection of the surround is a feasible and working strategy. However, it is also conceivable that the tissue around a tumor are in fact the seizure core, and the surround as we understood it in our model is the wide brain network the tumor and surrounding tissue are embedded in. Again, this highlights that the spatial scale at which to consider ‘surround’ and ‘core’, and the meaning of ‘increased excitability’ need more precise clinical and theoretical definitions. In this context, we wish to highlight that our results do not necessarily indicate a bigger resection area (i.e. also remove the increased surround excitability). Rather, our results suggest that some patients may have an enabling surround (of unknown spatial extent), predisposing them to generate seizures more easily. The increased surround excitability is to be understood as a pro-ictal condition. Such a condition on its own would not generate seizures, but needs provocation or driving from an epileptic generator. In patients who experience a rapid recurrence of their seizures soon after surgery, such epileptic provocations or generators may already exist (e.g. pre-existing micro-seizure domains). In other cases, there may be an early post-surgical improvement, but relapse over a longer period of time. This latter case may occur if a new epileptic generator forms slowly over time, perhaps by mechanisms of plasticity [50].
To further validate our model hypotheses, we derived two predictions to compare to clinical literature. Firstly, the model predicts an increased excitability in and around the seizure onset zone (SOZ) for high amplitude onset pattern. Indeed Enatsu et al. showed that the stimulation response amplitude (after CCEP) is higher in the seizure onset zone (SOZ) of high amplitude onset seizures compared to low amplitude onset seizures [15]. Our model further predicts that in the low amplitude fast onset pattern, recruitment is only possible given a distributed network of patches of microseizure activity. This means that for seizure initiation, (micro-)seizure activity at multiple sites must be visible. We restricted our simulation to the spatial extent of a single ECoG electrode, but it is conceivable that such networks of patches can extend to multiple electrodes. Hence we would expect the observation of a wider seizure onset zone (i.e. more electrodes to show seizure activity at onset). This is indeed also the case, as reported by [17].
Further support for our model is provided by a recent study, which used microwires in conjunction with depth electrodes in mesial temporal lobe structures to record temporal lobe seizures [51]. Their findings confirm the spatio-temporal patterns observed in our model: low amplitude fast onset patterns are shown to begin with small patches of seizure activity, which slowly grow in size and coalesce in the initial phase of the seizure. However, in order to fully test the validity of the model finding, high-resolution spatio-temporal recordings with a significant spatial coverage (e.g. high density electrode arrays [52]) would be required to compare the evolution of the seizure in patients to our model simulations. Additionally, to test for increased excitability levels, spatially, and temporally, an active stimulation paradigm might be required. Future studies should systematically test the excitability levels at all spatial locations covered by ECoG, possibly at different time points, (i) to elucidate if the cortical excitability around the SOZ is increased in high amplitude compared to low amplitude onset seizure, and (ii) to determine the spatial extent of such an increased excitability.
Finally, we proposed the perspective of a spatio-temporal monitoring system for cortical excitability. Such a system would offer the possibility of alternative treatment strategies to surgical removal of tissue. In particular, a closed-loop control system can use the information on cortical excitability to deliver spatio-temporally precise control stimulation. Various control frameworks exist already, including electric [53–57], magnetic [58], and optogenetic [59] control. We also summarised several theoretical investigations in our recent review [60], and also see [61]. The key contribution of our work towards development of such closed-loop control systems is that we highlight the need for a spatio-temporal controller. Much of the recent attention has been focused on temporal control, but the spatial aspect has been largely neglected. We demonstrated here that the distinction of spatial territories into seizure core and surround is crucial to understand seizure onset. Furthermore, we demonstrated that only controlling the core might not be sufficient in some types of seizures, but controlling the surround could be as important.
Previous studies have demonstrated that the spatial distinction of a seizure core and a surround is a useful concept in the theoretical [21] and the clinical [25, 44, 45, 62] context. Using this concept allowed us to demonstrate that the excitability level of the surround is a crucial factor in determining the onset pattern, and the pathological abnormality in the model. We also suggested further studies to be performed to test our model hypotheses; and the validity of the concepts proposed here. If validated, our model essentially suggests that patients with different onset patterns should be treated differently. In the case of low amplitude fast onset patterns, a traditional resection of the SOZ should lead to seizure freedom. In the case of high amplitude onset patterns, we suggest the control of the excitability in the surround to be crucial.
The model used here has been introduced in detail in our previous publication [21], and we use the same basic settings and parameter values as in the previous publication (specific changes in parameters are shown in Table 1). The MATLAB code for the model is also published on ModelDB (ID: 155565) alongside the previous publication.
The basic unit of our model is a cortical minicolumn, which is modelled by a Wilson-Cowan unit (Wilson and Cowan, 1973) consisting of an excitatory and an inhibitory neural population (E and I). It assumes that the E and I populations interact with each other (Fig 5A), and the this interaction influences the firing activity of the target population. This established model has been shown to be able to capture some coarse-grained dynamics of neural populations [63–66], and is at the same time quick to simulate. Importantly, we demonstrated in our previous work [21] that this model offers a good trade-off between level of abstraction while still capturing key seizure dynamics to study seizure onset mechanisms. It allows for the simulation of spatio-temporal patterns of seizure activity, and the analysis of spatial tissue heterogeneities and mesoscopic connectivity in that context.
The equations to simulate a single minicolumn is, exactly as in our previous study [21], a simplified Wilson-Cowan unit given in Eq 1:
τ E · d E d t = - E + S i g m ( C E → E · E + C I → E · I + P + A s · S ( t ) ) τ I · d I d t = - I + S i g m ( C E → I · E + C I → I · I + Q ) , (1)
where E is the fractional firing activity in the excitatory population; I is the fractional firing activity in the inhibitory population. P and Q model the baseline activation of the E and I populations (or sometimes also referred to as the basal input level). S(t) is a noise input to E to reflect noise, or input from other brain areas that is uncorrelated to the dynamics of local interest. As is the coupling strength of the noise input. The connectivity constants Ci → j (with i,j = E or I) determine the coupling strength between the populations, which essentially indicate how input from the source population is interpreted by its target population (see Fig 5). E.g. CE→I controls how input from the E population influences the I population.
Sigm(…) is a sigmoid function, which is derived from a distribution of firing thresholds in the underlying neural population [67]. It is defined as S i g m ( x ) = 1 1 + e x p ( - a ( x - θ ) ), where a is the steepness of the sigmoid and θ is the offset (in x) of the sigmoid. We fix the sigmoid parameters (a = 1,θ = 4) following previous work [33], as variations in the other parameters P,Q,C effectively result in a change of the sigmoid shape.
To form a cortical sheet, we concatenate 150 by 150 minicolumns (Fig 5A), and couple them to each other. Note that the full system is also described by Eq 1, if we understand E,I,P,Q and τ as vectors, and the connectivity C as matrices. The spatial scale of the cortical sheet is derived from the assumption that each minicolumn is about 50μm × 50μm in size [68]. A macrocolumn is then formed by 10 × 10 minicolumns, which agrees with the size suggested in [69]. The lateral connections between the minicolumns include local feed-forward excitation and inhibition, as well as remote feed-forward excitation (see Fig 5B for a schematic). The projection targets of local feed-forward excitation and inhibition are chosen with a certain probability that falls off with distance, as is shown in slice studies (e.g. [70]), and as is adopted by most modelling studies (e.g. [71, 72]). We chose parameters for cortical connectivity following the suggestions in [73], which are derived from tract tracing experiments in human cortical tissue. Specifically, [73] also suggest to include the so-called remote feed-forward excitation. This is a mesoscopic type of connection established by principle neurons, which target clusters of cells which can be several millimetres away (see Fig 5B, right column for an example). The exact algorithm for finding the connectivity is described in the supplementary information of our previous work [21], following [73], and we also provided these connectivity matrices with our code.
Finally, the noise input is structured at the cortical sheet level, such that a macrocolumn (10 by 10 minicolumns) receives the same S(t) [68, 74]. We used noise values drawn from a standard normal distribution. The effective noise coupling strength is set to As = 1. In this setting the system is not entirely dominated by the noise input but the noise influences the deterministic dynamics. Simulations of the system used a Euler-Maruyama fixed step solver, with a stepsize of 2 ms. Qualitatively equivalent results are found for smaller stepsizes. Fig 5 schematically summarises the model. Supplementary S1 Text additionally shows some simulation results for the deterministic system (i.e. without noise input) for comparison and completeness.
In our previous analysis [21] we demonstrated that the dynamics of the model cortical sheet can be characterised by distinct states. The sheet can exist in a background state, with little activity, where low amplitude noise dominates (region I in Fig 6). The second state is an oscillatory activity state (region III in Fig 6). We identify this oscillatory state as the seizure state (see [21]). Interestingly, a parameter region where both the background and seizure state coexist can also be identified (region II in Fig 6). This coexistence of states is termed “bistability”. (We call the case of only one state existing “monostability”). Here we only show the parameter space for the two input parameters (P and Q, which essentially model the level of baseline activation of the E and I populations, respectively). Equivalent states can be found for other parameters. For the majority of the manuscript, we operate in the parameter region outlined in red in Fig 6.
Importantly, we demonstrated previously that transitions to seizures can be triggered by a short transient input/stimulus when the simulated cortical sheet is in the bistable regime [21]. Such a bistable mechanism of seizure onset has also been proposed by other theoretical studies [26, 75]. However, we also show in our previous study that even in the monostable background state, it is possible to induce a transition to the seizure state in the whole sheet, when microdomains with autonomous seizure activity are embedded in the cortical sheet [21] (inspired by the clinical observations of microseizures [3]). In other words, seizures cannot only be provoked by transient stimuli in the bistable regime, but can also slowly penetrate into “healthy” monostable background tissue. We also presented the behaviour of the system in a dynamical systems context in S1 Text, to aid the understanding of our work in the theoretical domain.
We also use the term “excitability” throughout the manuscript. We will use this term to describe the proximity to the monostable seizure state in parameter space. In other words, how close the current parameter setting is to the dark blue region (III) in parameter space in Fig 6. This definition is intuitive in the context of our paper, because we showed in our previous work [21] that the proximity to the monostable seizure state directly relates to how “easy” it is to provoke seizure activity from perturbation stimulation. In our case, “easy” refers to how big the stimulus has to be spatially, and how strongly it activates the populations.
Increasing excitability throughout the manuscript usually refers to increasing P if not mentioned otherwise. However, P is not the only parameter that can lead to a change of excitability. For instance a decrease in Q would have the same effect as it moves the system closer to the seizure state. Indeed our results still hold when using a decrease in Q instead of an increase in P (see for example S1 Fig). In addition, we tested our results in this manuscript with regards to robustness to changes in conduction delays and boundary effect, we found no qualitative difference, and all conclusions still hold.
In order to simulate seizure onset, we focus on mechanisms that are triggered, or caused by small patches of localised seizure activity (termed class II and class III onset in our previous work [21]). These small patches will also be referred to as microdomains, and seizure activity restricted to such small patches of tissue will be termed microseizures [3]. A patch refers to a group of spatially contiguous minicolumns. To model microseizures, we put a patch in a monostable seizure state (in our case, we gradually changed Ppatch over 3 seconds to Ppatch = 1). P for the surrounding cortical sheet is termed Psurround, which is usually either set to −2.5 to make the surrounding monostable, or −1.5 to make the surrounding bistable.
To model different onset patterns, we scanned the number and spatial arrangement of the patches of microseizure activity. The spatial arrangement of patches is chosen randomly. Hence we scan 10 different configurations with the same algorithm, and report average results of seizure onset amplitude in our scans.
In the case of multiple patches in space, we assume that they do not all show seizure activity at once, but rather transition to the seizure state with a 50 millisecond gap between each patch. This is to ensure that we are not merely detecting our effects due to synchronisation effects following simultaneous activation. We also scanned the ‘onset time gap’ (between 6 and 100 millisecond, and also random time gaps) and found no qualitative difference in our results.
To ensure reproducibility, we shall also make the code to generate low and high amplitude onset patterns available on ModelDB (ID: 226074). The code includes the details described here.
Table 1 shows all the parameter we used for the figures in the Result section.
As in our previous work, we model the local field potential of a minicolumn by a weighted sum of its excitatory and inhibitory activity, and its noise input. This measure is an approximation of the post-synaptic potential in the pyramidal (excitatory) population. To model the recording of an ECoG that would lie over the simulated sheet, we used the average of all minicolumn LFPs, as an approximation. This average signal is then high pass filtered at 1 Hz, as clinical recordings do not usually include the DC component. (Our model actually shows a DC shift at seizure onset, but the details are not shown here. More related information can be found in a recent publication [29]).
We used a series of different algorithms (amplitude deviation detection based on the stable background amplitude; DC shift detection in the unfiltered time series; and deviation of steepness of time series from the background) to detect seizure onset from the simulated ECoG signal. All algorithms showed qualitatively similar results.
We simulated surgical removal of cortical tissue by removing all connections to and from the resected area. The resected area in this case is a strip of the sheet, 150 × 30 minicolumns in size.
To simulate stimulations in one minicolumn, we raised their input level P by 5 parameter units for 6 milliseconds. I.e. the target minicolumn receives a transient input via a short pulse to the parameter P. In order to gauge the spatial variance, we simulate the stimulation to different positions on the cortical sheet, we target 9 minicolumns in total. Their locations are arranged on a 3 by 3 grid on the cortical sheet, in positions 25, 75, and 125 in horizontal and vertical dimensions, respectively. The stimuli are delivered successively at spatially distant sites to minimise interference. This means that only one location is stimulated at any given time. The stimuli between sites are 6 milliseconds apart from each other. This is to avoid joint activation of all sites.
The stimulation responses are measured at each stimulation location. For this, we simply read out the LFP of the stimulated minicolumn. A moving window average is applied (window length of 20 milliseconds) to smooth signal, and the maximum value is taken as the response amplitude from the smoothed signal at each stimulation location. The average response amplitude at any time is simply the mean of the response amplitudes of all 9 locations.
This stimulation procedure is then repeated every 400 milliseconds, meaning that every 400 milliseconds, we derive a measure of average response amplitude over the whole simulated cortical sheet. This protocol has not been developed to be optimal, but is intended to demonstrate the principle of how excitability levels could be tracked using small stimuli.
|
10.1371/journal.pbio.1002330 | ELF5 Drives Lung Metastasis in Luminal Breast Cancer through Recruitment of Gr1+ CD11b+ Myeloid-Derived Suppressor Cells | During pregnancy, the ETS transcription factor ELF5 establishes the milk-secreting alveolar cell lineage by driving a cell fate decision of the mammary luminal progenitor cell. In breast cancer, ELF5 is a key transcriptional determinant of tumor subtype and has been implicated in the development of insensitivity to anti-estrogen therapy. In the mouse mammary tumor virus-Polyoma Middle T (MMTV-PyMT) model of luminal breast cancer, induction of ELF5 levels increased leukocyte infiltration, angiogenesis, and blood vessel permeability in primary tumors and greatly increased the size and number of lung metastasis. Myeloid-derived suppressor cells, a group of immature neutrophils recently identified as mediators of vasculogenesis and metastasis, were recruited to the tumor in response to ELF5. Depletion of these cells using specific Ly6G antibodies prevented ELF5 from driving vasculogenesis and metastasis. Expression signatures in luminal A breast cancers indicated that increased myeloid cell invasion and inflammation were correlated with ELF5 expression, and increased ELF5 immunohistochemical staining predicted much shorter metastasis–free and overall survival of luminal A patients, defining a group who experienced unexpectedly early disease progression. Thus, in the MMTV-PyMT mouse mammary model, increased ELF5 levels drive metastasis by co-opting the innate immune system. As ELF5 has been previously implicated in the development of antiestrogen resistance, this finding implicates ELF5 as a defining factor in the acquisition of the key aspects of the lethal phenotype in luminal A breast cancer.
| The transcription factor Elf5 defines hormone-insensitive and endocrine-therapy–resistant breast cancer. In this study, we have discovered that ELF5 drives the spread of tumor cells to the lungs. We demonstrate that the underlying mechanism for this metastatic spread is via recruitment of the innate immune system. Interestingly, this effect is able to overcome the other tumor-suppressive effects of ELF5 on cancer cells, such as reduced proliferation, motility, and invasion. This important finding challenges the more conventional view that the most potent determinant of metastatic activity lies within the cancer cell. We clearly demonstrate that the innate immune system strongly influences the metastatic activity of cancer cells despite their cell-intrinsic spread potential. Our previous work demonstrated that in luminal breast cancer, ELF5 is a key determinant of antiestrogen therapy resistance. Here, we show that the metastatic mechanism driven by ELF5 is most important in luminal breast cancer patients, in whom higher ELF5 expression is associated with low presence of cytotoxic T lymphocytes, an immune cell population responsible for tumor rejection. Thus, we now see that ELF5 may be behind the two most important processes that cause luminal breast cancers to progress towards the lethal phenotype; resistance to antiestrogen therapy and the development of metastatic activity. This understanding could pave the way for new therapeutic strategies to be devised and new predictive tests to be developed.
| Breast cancer is a heterogeneous disease in which subtypes predicting differential clinical outcome are recognized based on shared patterns of gene expression and mutation, indicating a shared path to cancer [1]. The most striking subtype distinction in breast cancer is provided by expression of ESR1, the estrogen receptor (ER). This divides breast cancer into two very different diseases, recognizable by more than their response to hormones and antiestrogen therapies. For example, the risk of recurrence remains constant for more than 20 y for ER+ disease, but drops dramatically after 5 y for ER- disease [2,3]. ER+ cancers are also more insensitive to chemotherapy than those that are ER- [4–6]. The basis for this phenotypic dichotomy probably includes the characteristics of the cancer’s cell of origin, which for the basal ER- and luminal ER+ breast cancer subtypes are thought to be the members of the mammary progenitor cell pool [7].
A key transcriptional determinant of cell fate decisions made by the progenitor cells is the ETS transcription factor ELF5 [8], which is first expressed as mammary stem cells differentiate to become progenitor cells, coincident with promoter demethylation [9]. In progenitor cells ELF5 levels fall under hormonal control. The systemic hormones of pregnancy prompt local mammary paracrine signals involving RANKL [10–12] to induce ELF5 [13,14], and force a progenitor cell fate decision that establishes the ER- secretory cell lineage responsible for milk production. An alternative progenitor cells fate, that of an ER+ hormone sensing cell, may result if ELF5 levels remain in check due to the dominance of the estrogen-driven phenotype [15].
In luminal breast cancer cells, a mutual negative-regulatory loop between ER and ELF5 occurs, which is dominated by ER and so keeps ELF5 levels low [16]. Conversely, ER- basal breast cancers are characterized by high ELF5 levels, while the stem-cell–like claudin-low subgroup does not express ELF5 [16]. Knockdown of ELF5 levels in luminal breast cancer cells has a small effect on proliferation, but a much greater effect is seen in ER- basal cell lines [16]. Importantly, ELF5 levels rise when MCF7 luminal breast cancer cells acquire antiestrogen resistance, and resistant cells become dependent on ELF5 for their proliferation [16]. Thus, increased ELF5 levels provide an escape pathway from inhibition of proliferation by antiestrogen therapy, facilitating disease progression. Whether ELF5 is involved in other key aspects of disease progression, such as metastasis, is unknown.
Like primary tumor formation, the acquisition of the metastatic phenotype involves events that alter both intrinsic cell behavior and the extrinsic responses of the host environment. An example of an intrinsic event is the gain of phenotypic plasticity, which regulates the acquisition of invasive and motile characteristics to cancer cells [17]. ELF5 influences phenotypic plasticity by driving the expression of epithelial characteristics, as shown by the fact that knockout of Elf5 in mice, or knockdown of breast cancer cells, caused the loss of epithelial patterns of gene expression, while forced Elf5 expression caused their gain [16,18].
An example of an extrinsic event is the interaction of the tumor with the host immune system. For example, in the mouse mammary tumor virus–Polyoma Middle T (MMTV-PyMT) model of breast cancer, knockout of CSF-1 depleted macrophages and delayed the development of lung metastases, while over expression caused the migration of macrophages into the tumor and accelerated metastasis [19,20]. Another important innate immune cell subset active in metastasis of mammary and breast cancer are myeloid-derived suppressor cells (MDSC) [21]. Their circulating numbers are increased by the presence of a tumor [22,23]. They invade primary tumors, where they promote angiogenesis, via Matrix Metaloproteinases (MMP) secretion and Vascular Endothelial Growth Factor (VEGF) production [24]. These cells inhibit and kill natural killer cells [25] and T-cytotoxic lymphocytes [26], while promoting the proliferation of the T-regulatory cell population and inhibiting dendritic cell maturation; all mechanisms that allow tumors to evade immune control [27]. In some contexts MDSC can also promote type II macrophage development and macrophage-assisted metastasis. In the MMTV-PyMT model of mammary metastasis, increased TGF beta signaling caused their recruitment to primary tumors. Depletion of their numbers reduced the number of lung metastases while tumor cell co-inoculation with MDSC increased the number of lung metastases [28,29].
We have used our inducible mouse model of mammary-specific ELF5 expression, in the context of luminal mammary tumors induced by PyMT expression, to investigate the roles played by ELF5 during mammary carcinogenesis and progression to metastatic disease.
To investigate the effects of Elf5 expression in breast cancer progression, we crossed our mammary epithelial specific ELF5-inducible transgenic mouse [8] with the MMTV-PyMT mouse model of luminal mammary cancer [30–32]. Triple-transgenic animals were created carrying one copy of each of the alleles (S1A Fig) on an inbred FVB/N genetic background. Time course experiments showed that after 7 d of Doxycycline (DOX) in the feed the ELF5 protein was detectable by western blot in established mammary tumors and that expression was maintained for at least 8 wk (S1B Fig).
Induction of ELF5 was measured in whole tumors by imaging EGFP fluorescence. A heterogeneous pattern of expression was observed (Fig 1A), which may have resulted from a chimeric expression pattern of the rtTA transgene, a feature of older MTB mice [33]. We used Kaplan-Meier survival plots to analyze primary tumor growth. Only mice that showed a tumor burden of ~10% (7%–13%) of body weight at autopsy were included in the analysis (Fig 1B). Overall survival at ~10% tumor burden showed no significant difference (Fig 1C LHS), however, forced expression of Elf5 produced tumors that were detected earlier (Fig 1C middle), but which took longer to then reach the ethical endpoint (Fig 1C right-hand side [RHS]). To overcome the effects of heterogeneous ELF5 induction (Fig 1A), we performed intraductal allografts of Fluorescence-Activated Cell Sorting (FACS)-sorted (Lin- and CD24+) tumor cells that were either EGFP (ELF5) positive or negative. Purified cells were injected into the mammary ducts of FVB/N host animals pretreated with DOX and maintained on DOX. EGFP+ transplants resulted in longer overall survival, longer time to tumor detection and longer time to the ethical endpoint, than transplants originated from EGFP- cells (Fig 1D). To demonstrate that EGFP/ELF5 was not only expressed in a particular subset within the mammary epithelium, we performed a similar experiment including allografts made from cells that were sorted (Lin- and CD24+) from excised tumors not carrying the ELF5 transgene (PyMT/wild type [WT]) or cells that were purified from tumors (PyMT/ELF5) made fluorescent by a short 7 d pulse of DOX administration, to allow flow capture of EGFP+ cells as before, but then injected into the mammary ducts of hosts either pretreated and maintained on DOX, or not ever treated with DOX (S2A Fig). As before, EGFP+ allografts maintained on DOX produced slower growing tumors. The two control groups (WT and EGFP+ with no DOX after transplant) produced tumors that expanded at indistinguishable rates.
The effect of ELF5 on a variety of cell-autonomous endpoints was examined. Cell proliferation was analyzed using a BrdU pulse to label cells in S-phase and EGFP IF to detect Elf5-expressing areas. We observed that much higher rates of cell proliferation occurred in the areas of the tumor which expressed low levels of ELF5, marked by low or no EGFP. This was observed after 2 wk of Elf5 induction (S2B Fig) and was maintained for at least 8 wk of DOX treatment (Fig 1E), indicating long-term functional activity of the Elf5 transgene. We used these flow-sorted primary cells to examine other cell-autonomous aspects of ELF5 action in tumor formation. ELF5 reduced the motility of tumor cells through a permeable membrane in a Boyden chamber, using serum as the chemo-attractant (Fig 1F), and also reduced the ability of these cells to invade through a layer of matrigel using the same apparatus (Fig 1G). Injection of primary cells into the tail vein of wild-type hosts produced engraftment of WT tumor cells in the lungs, but rarely when the cells expressed ELF5 (Fig 1H). We compared these cell populations using Affymetrix MoGene transcript expression arrays and examined the expression of genes indicative of epithelial and mesenchymal characteristics. Long-term induction of ELF5 produced a detectable mesenchymal to epithelial transition while EGFP- cells showed no change and resembled WT PyMT cells (Fig 1I). Together these data show that forced Elf5 expression reduced cancer cell proliferation, motility, invasion and mesenchymal characteristics, corresponding with reduced primary tumor growth in the MMTV-PyMT mouse mammary cancer model.
Induction of ELF5 caused wide-spread tumor hemorrhage. This was apparent as small and discrete areas of hemorrhage after 2 wk of induction that rapidly developed to affect the entire tumor (Fig 2A). Haematoxylin and eosin (H&E) histology showed pools of erythrocytes within the affected area of the tumor and macrophages exhibiting hemosiderin (Fig 2B). Infiltrating CD45+ leukocytes were found associated as clusters or along basement membrane planes between lobular structures (Fig 2C). Quantification using flow cytometry (FC), revealed a 6-fold increase in Ter119+ tumor erythrocytes (Fig 2B RHS) and 2-fold increase in CD45+ leukocytes (Fig 2C RHS). Immunohistochemical staining for endothelium using antibodies recognizing CD31 revealed a higher vascular density with finer and more branched vessels in response to Elf5 (Fig 2D). Flow cytometry showed a 1.5-fold increase in CD31+ endothelial cell content of tumors. Quantification of endothelial area using CD31 immunofluorescence (IF) confirmed a statistically significant increase in the vasculature in response to ELF5 (Fig 2E).
We used in vivo real-time intra-vital microscopy to examine tumor vasculature reorganization and increased blood vessel permeability. Intravenous injection of blood tracer quantum dots revealed their accumulation in the interstitial space of PyMT/ELF5 mice treated with DOX for 8 wk (Fig 3A), but not in control animals. Live time course imaging at the times indicated in Fig 3A showed that quantum dots accumulated in the interstitial space within minutes of injection and reached a steady state after 1 h. Quantification showed that accumulation of quantum dots in the spaces beyond 5 um from the center of major vessels was mostly complete within 30 min (Fig 3B). Blood vessel permeability was found to be very consistent between individual mice of the same genotype and the increased permeability of ELF5high tumors was highly statistically significant (Fig 3C).
The ability of Elf5 to induce an angiogenic response in the PyMT tumors was analyzed using an independent experimental system. Two independent cell lines established from explanted PyMT tumors were stably infected with the pHUSH construct encoding a DOX inducible Elf5 (V5 tagged) expression cassette [16]. PyMT-ELF5-V5 cells robustly expressed ELF5-V5 upon DOX exposure (Fig 3D). PyMT-ELF5-V5 cells were maintained in culture with and without DOX for 2 wk, harvested, re-suspended in matrigel and placed subcutaneously in the flank of congenic FVB/n recipients. Hosts on DOX showed increased recruitment of vasculature around the implantation site (Fig 3E). Flow cytometric analysis of the cells captured within the matrigel revealed greater infiltration of CD31+ cells from DOX treated hosts (Fig 3F).
Overall these data demonstrate that ELF5 exerts a potent angiogenic force that produces an aberrant leaky vasculature.
We examined the effect of the induction of ELF5 on the metastatic behavior of the PyMT model. In control animals, constitutive PyMT expression produced no visible lung metastatic nodules by the time the primary tumors reached the ethical endpoint of 10% body weight (Fig 4A), but small metastases within the lungs were detectable by H&E histology (Fig 4B). DOX administration in control animals had no effect on metastasis (Fig 4C and 4D). Induction of ELF5 from 6 wk of age resulted in a dramatic increase in metastasis to the lungs, now visible as numerous nodules on the surface of the lung at the ethical endpoint (Fig 4E) and large and numerous metastases within the lungs by H&E histology (Fig 4F). Induction of ELF5 for 2 wk once tumors were palpable also increased the size and number of detectable lung metastases (Fig 4G and 4H) but with more variable penetrance between animals compared with longer DOX treatment. Most of these metastases expressed ELF5, observed by visualization of EGFP (Fig 4I) and by ELF5 IHC (Fig 4J). Quantification showed a positive correlation between the size of the metastatic lesion and the level of ELF5 protein (Fig 4K). Unlike the primary tumors the metastases showed no regions of hemorrhage. Quantification of H&E stained sections showed statistically significant increases in the number of lung metastases (Fig 4L and 4M). Measurement of metastatic area produced similar results (Fig 4N). Induction of ELF5 greatly increased the amount of PyMT-mRNA present in blood (Fig 4O), suggesting increased numbers of circulating tumor cells. Elf5 is a master regulator of the development and remodeling of the mammary epithelium during pregnancy. During this period Elf5 is intensively expressed. We found that the metastasis-promoting effect of Elf5 was comparable to that produced by pregnancy in this model (Fig 4P).
We purified Lin- CD24+ EGFP+ mammary epithelial cancer cells from the primary tumors and lung metastases of DOX treated PyMT/ELF5 mice, and Lin- CD24+ mammary epithelial cancer cells from the primary tumors of DOX treated PyMT/WT mice, and examined the differential patterns of gene expression using Affymetrix arrays analyzed by LIMMA. Functional gene networks were identified by Gene Set Enrichment Analysis (GSEA) and were visualized using the Enrichment Map plugin for Cytoscape software (Fig 5A) (for a PDF version that can be zoomed in on, see S3 Fig). EGFP+ cells were compared to WT cells from primary cancers to discover functions altered by ELF5 induction, shown by the inner node color, while the outer node color shows how these functions changed in EGFP+ primary compared to EGFP+ lung metastasis. Functions related to cell cycle control, DNA repair, transcription, and translation were suppressed by ELF5 during primary carcinogenesis and remained similarly suppressed in the metastases. Aspects of kinase-based cell signaling were increased by ELF5 during primary carcinogenesis but were then generally suppressed following metastases, although GPCR-mediated signaling increased during carcinogenesis and increased again following metastasis. These results are consistent with Elf5 action in human breast cancer cell lines MCF7 and T47D [16]. Strikingly, we identified functional clusters related to an inflammatory response that were activated in the ELF5-driven primary tumors, but reversed in the metastases. To investigate this further, we extended the GSEA to include molecular signatures of immunologic origin. Guided by an automated clustering approach, we identified gene-sets related to HGF and IL4, inflammation, immune system and interferon responses, and activated monocytes, which were all enriched in the primary tumors in response to ELF5 and suppressed in the metastases (Fig 5B). Fig 5C shows a heat map of the Normalized Enrichment Score (NES) for each individual gene-set included in the defined functional clusters (S1 Table).
We identified patients from the TCGA breast cancer cohort that were classified as having either a luminal A or luminal B PAM50 molecular subtype [34]. Each luminal subtype was stratified on ELF5 expression levels and ranked gene lists of differential expression were generated using LIMMA. These ranked lists were used as the input for GSEA, to allow comparison of the transcriptional response correlated with increased ELF5 expression in human luminal cancers. We found a positive correlation in luminal A tumors, whereas a negative correlation was found in luminal B patients (S4A and S4B Fig). Higher ELF5 expression in luminal A, but not B breast cancers, was broadly associated with the same five functional networks identified in the PyMT/Elf5 model: HGF and IL4, invasive phenotype, monocytes, immune system involvement, inflammation and the interferon response (S4C and S4D Fig). These observations suggest that ELF5 expression produces a more similar response in human luminal A breast cancer to that observed in the ELF5-driven mouse PyMT model.
Taken together, these findings confirm our observations made in human breast cancer cells regarding the function of ELF5, and indicate that, in vivo, these effects are coupled with the immune system, both in the PyMT model and in luminal A human breast cancer.
We sought to characterize the ELF5-driven inflammatory phenotype and its effect in metastasis. There is an extensive and persuasive literature regarding the pro-angiogenic and -metastatic roles of innate immune cells in the PyMT model. New drugs targeting the immune system are currently revolutionizing cancer treatment. We examined the recruitment and activation of tumor immune cell infiltrates in response to ELF5 using flow cytometry. We measured myeloid (Fig 6A) and lymphoid (Fig 6B) lineages as a percentage of the remaining total cells, or as a proportion of total CD45+ hematopoietic cells. S5 Fig shows the gating strategy and cell surface markers used to produce this analysis. Among the myeloid populations, MDSCs (defined as Gr-1+CD11b+) showed an increased proportion of either total cells or hematopoietic cells, however no significant changes were observed in the number of the other myeloid populations analyzed (Fig 6A). T- and B-cell lymphoid lineages increased as a proportion of total cells, indicative of increased inflammation (Fig 6B). Proportional with the total leukocyte population, B-cell increase was 1.5-fold higher in ELF5 tumors. Within the leukocyte T CD3+ population, T-CD8+ cell number was significantly decreased (2-fold) but no change was observed in the T-CD4+ population, increasing the T-CD4 to -CD8 cell ratio consistent with a MDSC-driven pro-tumorigenic immune suppressive microenvironment.
MDSC (Gr1+) can be subdivided in the granulocytic and the monocytic subset according to their expression of the antigen molecules Ly6G and Ly6C, (Mo-MDSC (CD11b+Ly6G-Ly6Chigh) and G-MDSC (CD11b+Ly6G+Ly6Clow) [35,36]. Flow cytometric analysis of these subsets in PyMT tumors determined that the main population was the Ly6G+ granulocytic subset (Fig 6C). Reactive Oxygen Species (ROS) play a major role in MDSC-mediated immune suppression though the impairment of T cell activation [26]. ROS production by MDSC was significantly increased in both infiltrated granulocytic and monocytic subsets in response to Elf5, consistent with a tumor permissive environment (Fig 6D). A large proportion of the infiltrated Ly6G+ population presented ROS production and this number was further increased to nearly 100% in response of ELF5. The intensity of ROS production was also increased in the MDSC populations in response to ELF5 (Fig 6D). Thus Elf5 increased the number and suppressive ability of tumor-infiltrated MDSC.
To determine if the increase in MDSC could account for the increase in metastases caused by ELF5, we used the specific Ly6G antibody to deplete the granulocytic MDSC population during induction of ELF5 in PyMT tumors. Two weeks of treatment with the rat Ly6G antibody resulted in a consistent and efficient depletion, no granulocytic MDSCs were observed in the blood of Ly6G-treated animals (Fig 7A), and a 98% depletion of tumor-infiltrated MDSC was observed (Fig 7B). As a result, only 1.5% of infiltrated Ly6G+ granulocytic MDSC cells were identified in both PyMT/WT and PyMT/ELF5 tumors in the CD11b+ compartment (Fig 7C). Ly6G depletion did not significantly affect the numbers of other infiltrated immune populations in PyMT tumors (S6 Fig). An analysis of the ROS production in the tumor infiltrated CD11b+ myeloid population showed a reduction of total ROS producing cells, consistent with a Ly6G granulocytic cell depletion and a less immune-permissive environment (Fig 7D). MDSC depletion reduced the number of lung metastases in both WT and ELF5 tumors (Fig 7E). We also observed that the antibody treatment reduced the number of red blood cells within the primary tumor (Fig 7F), establishing MDSCs as a key part of the mechanism responsible for both induction of metastases and the hemorrhagic tumor phenotype by ELF5.
To study the relevance of ELF5 in metastasis in luminal breast cancer patients, we analyzed a cohort of ER+ HER2- tumors staining for ELF5 protein levels using IHC (Figs 8 and S7). This cohort has more than 15 y of clinical follow-up [37]. All patients were treated with the antiestrogen Tamoxifen and none received chemotherapy. We observed nuclear and cytoplasmic patterns of ELF5 staining. Across all ER+ cancers, higher nuclear ELF5 staining predicted better overall survival (OS) after 10 and 15 y but not after 5 y (Fig 8A LHS). This prediction was relatively weak as the hazard ratio was 0.5 and the p-value 0.03. In contrast, higher cytoplasmic ELF5 staining predicted worse survival, and at 5 y this prediction was strong, with the hazard ratio greater than 3 at a p-value of 0.005. These same effects were evident for distant metastasis free survival (DMFS) where again cytoplasmic ELF5 level was a strong predictor of poor survival (Fig 8B LHS).
We used the St. Gallen definition of Ki67% to split these ER+ cancers into luminal A and B tumors [38]. We found that cytoplasmic ELF5 staining in luminal A patients predicted poorer overall survival with a large hazards ratio, especially at 5 y when it was 11 (Fig 8A). A similar effect was evident for distant metastasis free survival and a large hazard ratio was again evident at 5 y (Fig 8B). Nuclear staining in luminal A patients weakly correlated with poor prognosis in the 10 y follow up overall survival but this prediction was not maintained after 15 y follow up. In contrast, ELF5 levels either cytoplasmic or nuclear, had no predictive value for survival in the luminal B subtype. These results show that ELF5 predicts poorer survival and metastasis in the Luminal A subgroup and that it is a marker of early progression in this subtype.
An interesting observation, given that ELF5 is a nuclear transcription factor, is that cytoplasmic rather than nuclear staining provides this prediction in luminal A breast cancer patients. Although abrogation of ELF5 transcriptional action by restriction to the cytoplasm is suggested by this finding, alternative explanations exist. For example, the antibody epitope may be obscured when ELF5 is bound within a specific transcriptional complex so we caution against over interpretation of the nuclear/cytoplasmic dichotomy until it is better understood.
We studied the immunogenicity of ER+ luminal breast cancer tumors in relation to Elf5. In the absence of a reliable immunohistochemical technique that detects MDSC we instead correlated ELF5 IHC protein levels from this cohort with staining for lymphocytes. We used CD3 and the cytotoxic specific T CD8 maker, the T cell subset targeted by MDSC that was identified in the PyMT/ELF5 model (S2 Table). In this cohort of patients, it has been demonstrated that tumor infiltrated T CD8+ cells correlate with better patient prognosis, suggesting that presence of this cell type is associated with immune tumor rejection [39]. The presence of lymphocytes was analyzed according to their location, intratumoral (within the tumor nests), in the adjacent stroma and in distal stroma. Cytoplasmic ELF5 staining significantly correlated with increased intratumoral T CD3 cell numbers in the luminal ER+ cohort (Spearman’s rank p = 0.11, rs = 0.156), with no correlations seen with T cells adjacent or distant to the tumor. Despite this increase in total T lymphocytes in ELF5-high expressing tumors, the number of intratumoral T-CD8+ lymphocytes were significantly underrepresented (Spearman’s rank p = 0.04, rs = -0.203). Categorical Mann-Whitney analysis (cut off CD3 ≥ 2 cells; CD8 > 1 cell, based on X-tile analysis) confirmed the direct association (p = 0.075) between cytoplasmic ELF5 expression and intratumoral CD3 infiltration and the negative correlation with the T-CD8+ subset (p = 0.046). When the ER+ cohort was split into luminal A and B subtypes these effects were maintained, although the statistical power of the analysis was reduced due to the sample number (S2 Table). Interestingly, a strong inverse association of nuclear ELF5 staining and T-CD8+ cells was identified in the ER+ cohort and in the Luminal B subgroup.
These data indicate that luminal ER+ tumors with high Elf5 levels show higher intratumoral T lymphocytes, however the cytotoxic T-CD8+ population is selectively reduced. Our results in human breast cancer are consistent with our observations in mice suggesting the implication of ELF5 in a tumor permissive inflammatory environment. These data establish a strong case for further investigation of the role played by Elf5 in immunosupression and its relationship with survival in luminal A breast cancer.
B-cell lymphocytes analysis using the B20 marker in the Nottingham cohort revealed a high number of samples with absent staining [40]. Fifty-six percent (73/130) of the cases in this study were completely negative for B20 and 80% (105/130) lay below the statistical x-tile cutoff (B20 > 5 cells). In the positive cases, B-cells infiltrated in the tumor nests were rare, with the majority of B-cells localized at the distal stroma. No correlation with ELF5 expression was found using intratumoral or adjacent stromal B-cell numbers. Spearman and Mann Whitney analysis on total and distal B-cell number revealed inverse associations between ELF5 expression and the CD20 marker as indicated in S3 Table. B-cell number is directly associated with breast cancer specific survival and longer disease free interval in ER+ patients treated with anti-estrogen therapy [40]. In the MMTV-PyMT model, ELF5 contributes to tumor progression; this discrepancy might be as a result of the poor modeling of the distal stroma in the PyMT tumor FACS analysis, where the majority of the tissue analyzed corresponds to intratumoral and adjacent stroma. Taken together, these results indicate that B-cell analysis does not model ELF5 action in luminal breast cancer.
We show that induction of ELF5 in the PyMT model leads to an increase in lung metastasis because ELF5 recruits MDSCs to the tumor, which promotes leaky vasculature and causes an increase in lung metastasis. Interestingly this effect swamps the cell autonomous effects of ELF5, which predict a tumor suppressor action. Analysis of human breast tumor data suggests that these processes also operate in ER+ breast cancer, and analysis of survival data shows this is prognostic in Luminal A cancers, with ELF5 expression in the cytoplasm clearly identifying a group of luminal A patients with early disease progression. High cytoplasmic ELF5 expression in luminal patients also correlated with a pro-tumor inflammation characterized by decreased cytotoxic T-CD8 lymphocytes.
ELF5 has been proposed by Chakrabati and colleagues as a metastasis suppressor gene for all breast cancers [18], but our studies demonstrate that the luminal A subgroup shows the opposite response. Interestingly, we show that ELF5 produces a number of cell-autonomous phenotypic changes that are consistent with a tumor-suppressor role, such as reduced proliferation, invasion, motility, epithelialization, and colonization in a lung-seeding assay, some features of which have been previously reported by us [16] and by Chakrabati and colleagues [18] using different model systems. Our results point to the dominance of the immune system over cell autonomous characteristics in regulating the metastatic behavior of luminal A primary tumors, and so to the importance of pursuing immunoregulatory therapies for luminal A breast cancer.
Given the previously described role of ELF5 in the progression to antiestrogen insensitivity in luminal breast cancer, where ELF5 levels rise [16], our results now show that this escape pathway is likely to lead to metastasis via attraction of the innate immune system. This may represent a normal biological response, as macrophages and neutrophils are attracted to the mammary gland during periods of tissue remodeling, especially during weaning when the mammary alveoli are largely resorbed, returning the gland to a series of branched ducts. We observed enrichment of involution and lactation signatures in our transcriptional data in response to ELF5 in both the mouse model and the TCGA data sets. Higher ELF5 expression may result in the tumor being seen by the host as an involuting mammary gland, and the luminal A subgroup may possess a background phenotype which allows or best expresses this appearance. When we treated our mice with the anti MDSC antibody Ly6G we did not completely ablate metastasis, rather we returned metastasis to control levels. This shows other prometastatic pathways continue to operate. One key pathway demonstrated in the PyMT model is the role of macrophages [41], whose numbers were unaffected by ELF5 expression.
Hemorrhagic necrosis and intratumoral hemorrhage is observed in breast cancer [42], where it generates pain due to mastodynia in otherwise painless cancers. Short-term induction of ELF5 in the mouse provides a good representation of this human pathology, where isolated hemorrhagic regions are seen. Longer term induction produces a more severe effect than seen in the clinic. The basis for hemorrhage involves the recruitment of MDSC, as shown by its reduction following suppression of these cells with Ly6G antibody. We speculate that the earlier detection of in situ ELF5 tumors is due to the immune cell infiltration, making them larger than the WT controls, since further monitoring showed that they expanded more slowly. Unlike the primary tumor, our data show that colonies of cells growing in the lungs have found a supportive environment. Transcriptional signatures indicative of cellular stress are lost. Necrotic areas are not present and the hemorrhagic phenotype is lost. Interestingly innate immune system recruitment also appears to be absent in the metastases.
These results indicate that ELF5 is a major determinate of the lethal phenotype in luminal A breast cancer. Elf5 expression provides a marker that defines early disease progression in this otherwise slow to progress subtype, and may also define a group that should benefit from future immunomodulatory therapies.
Mice were maintained following the Australian code of practice for the care and use of animals for scientific purposes observed by the Garvan Institute of Medical Research/St. Vincent's Hospital Animal Ethics Committee (AEC), AEC#11/35 (previous) and AEC# 14/27 (current). Euthanasia was performed by asphyxiation with carbon dioxide gas, followed by cervical dislocation, in a separate area away from other animals. For all surgical procedures, animals were anesthetized with Isoflurane at a rate of 1L/minute oxygen 5% Isoflurane for induction and 1L/minute 2% Isoflurane for maintenance. Animals recovered from surgery at room temperature in a box “half on/half off” over a warm heat pad to prevent hypothermia. They received analgesia systemically and locally. Animals were closely monitored until they had regained the ability to right themselves, then placed individually in cages in a special purpose room. When required, animals were checked for blood on their coats that will be removed before they wake up from anesthesia. The next day animals are checked for general condition (e.g., alertness, weight loss, balance, and mobility).
The Elf5 inducible PyMT mammary tumor transgenic model has been generated by crossing the MMTV- Polyoma Middle T antigen (PyMT) mouse mammary tumor model [30] with the doxocyclin (DOX) inducible Elf5 Knock In mouse line [8]. The inducible promoter induces a bicistronic cassette codifying for the human version of Elf5 followed by EGFP using an IRES sequence. We used the rtTA locus under the MMTV promoter to control the expression of Elf5 in the mammary epithelial cells (MTB animals). All animals used in this study are heterozygous for Elf5, MTB, and PyMT. S1A Fig shows a schematic representation of the transgenic cassettes and genotypes used for the study. To induce the expression of the Elf5 and EGFP mice were exposed to a diet containing 700 mg/Kg of Doxocyclin (Gordon’s Specialty Stockfeeds). For the neutrophil depletion experiment, 100 μg of Ly6G antibody clone 1A8 (UCSF) was injected IP twice a week for 2 wk, a pretreatment injection was performed 2–3 d before DOX exposure. Syngenic FVB/n hosts were used for matrigel plug assays.
Elf5 was tagged at the 3′ end with V5 and incorporated into the pHUSH-ProEX vector (Genentech) [43] as descried before [16]. Elf5 expression was achieved using Doxycycline (Clontech) at 0.1 μg/ml. Luciferase/GFP [44] and pHUSH-ProEx plasmids were packed into retrovirus using PlatinumE cells (Cell Biolabs) using FuGene6 or X-Treme transfection reagent (Roche) following manufacturer instructions. PyMT cell lines were established in culture from enzymatically disaggregated PyMT tumors and double FACS-purification based on CD24 expression; and were maintained in DMEM medium containing 10%FBS, 1% L-Glutamine, 5 ug/ml Insulin, EGF 10 ng/ml, and 10 ng/ml cholera toxin. The line was considered to be established in culture after ten passages.
Flow cytometry was performed using FACS Canto II or LSR II (analysis) and FACS Aria III (analysis and sorting) from Becton Dickinson and exported to the FlowJo software (Tree Star Inc.) for data analysis. Reactive Oxygen Species was measured using the DCFDA reagent (Abcam). DAPI ([4′,6-diamidino-2-phenylindole dihydrochloride]) (Molecular Probes) or Propidium Iodide (Sigma) was used as death cell exclusion marker. Flow cytometry was performed using the following fluorophore conjugated antibodies: CD45, CD31, Ter119 from BD Pharmingen; CD3 (clone 17A2), F4/80 (clone BM8), Gr-1 (clone RB6-8C5), CD4 (clone GK1.5), CD8 (clone53-6.7), CD11c (clone N418), CD11b (clone M1/70), and B220 (clone RA3-6B2) from eBioscience; Ly6G (clone 1A8) and Ly6C (clone HK1.4) antibodies were purchased from BioLegend. For neutrophil depletion experiments Ly6G antibody (clone 1A8) was used (UCSF or Bio X Cell) and FACS performed using an anti-rat IgG secondary form BioLegend. A list of the defined populations using these antibodies is listed in S4A Fig. IF for CD31 was performed using OCT embedded tissue and the BD Pharmigen antibody clone MEC13.3.
Two established PyMT cell lines were stably transduced with a DOX-inducible pHUSH vector encoding Elf5 tagged with the V5 peptide [16]. PyMT pHUSH-Elf5-V5 cells were then exposed to 0.1 μg of DOX every other day for 10 d or remained untreated for control. 105 long term DOX and control PyMT pHUSH-Elf5-V5 cells were then harvested and mixed with 4C matrigel (1:9/vol:vol) and immediately injected subcutaneously in the flank of FVB/n recipients. Hosts were exposed to DOX containing food 24 h prior matrigel implantation and until collection or left untreated for control cells. Ten days after implantation matrigel plugs were extracted, cell suspensions prepared using collagenase digestion and processed for FACS analysis.
Normalization and probe set summarization was performed using the robust multichip average [45] implemented in the Affymetrix library [46] from R [47] as part of the NormalizeAffymetrixST module in GenePattern. Control probe-sets were removed from the arrays. Differential gene expression was then assessed for each microarray probe set using an empirical Bayes, moderated t-statistic implemented in Limma (Smyth, 2004) using the limmaGP tool in GenePattern. All pairwise experimental comparisons performed are described, where relevant, in the text.
Where indicated, the analysis tools utilizing GenePattern software [48] are available at the Garvan hosted GenePattern server http://pwbc.garvan.unsw.edu.au/gp/. Microarray data are available from GEO: GSE58729. Detailed information about mRNA extraction, purification, chip hybridization and processing can be also found in this link. All analysis results, additional GSEA gene-sets, and custom analysis scripts are available on request from the authors.
For the analysis of TCGA expression data, clinical and molecular annotation of samples was obtained from the Cancer Genome Atlas (TCGA) breast cancer publication [49]. Agilent mRNA expression microarray data (Level 3) was obtained from the TCGA data portal in January 2012. Missing expression values were imputed and replaced using the k-nearest neighbor (KNN) approach, with k = 10 (using the ImputeMissingValuesKNN module in GenePattern). The TCGA microarray data consisted of a total of 533 tumors. From this, we generated 2 subsets of patients based on their PAM50 classified molecular subtype [34], 231 with a Luminal A PAM50 sub-type and 127 with a Luminal B PAM50 subtype.
The samples in each of these luminal patient subsets, were each stratified on expression level of ELF5, and the top 25% (ELF5hi) and bottom 25% (ELF5lo) expressing samples were selected. For each of these ELF5 stratified groups, differential gene expression between ELF5hi and ELF5lo patient groups was assessed, for each gene, using an empirical Bayes, moderated t-statistic implemented in LIMMA [50] via the GP tool in GenePattern.
For all pair-wise experimental comparisons, Gene Set Enrichment Analysis (GSEA) [51] was run in pre-ranked mode using a ranked list of the LIMMA moderated t-statistics. One thousand gene-set permutations were performed using minimum and maximum gene-set sizes of 15 and 1,500, respectively. Gene-sets used in GSEA were extracted from version 3.1 and 4.0 of the Broad institute’s Molecular Signatures Database (MSigDB) [52] and extended with additional curated gene-sets from literature. All GSEA analysis was performed using a combined set of the c2, c6 (for Fig 5A), and extended with c7 gene-sets (for Fig 5B and 5C and S4 Fig) from MSigDB plus additional curated sets that we identified in the literature. This resulted in a total of 5,145 gene-sets (MSigDB v3.1 c2, c6 collections plus custom sets) used in the initial, exploratory analysis, shown in Fig 5A, and an expanded gene-set collection of 6,947 gene-sets (MSigDB v4.0 c2, c6, c7 collections plus custom sets) used in the analysis described in Fig 5B and 5C, and S4 Fig.
Network-based visualization and analysis of the GSEA results was carried out using the Cytoscape [53] Enrichment Map [54] plug-in, with permissive thresholds of: FDR (Q-value) = 0.25; p-value = 0.05 and overlap coefficient cutoff = 0.5. The functional networks definitions were based on the cytoscape pre-annotated clusters tool.
To identify functional clusters of gene-sets that were enriched in the PyMT/ELF5 tumors and the TCGA luminal A ELF5hi tumors an automated clustering approach was used. First, an EnrichmentMap network of the GSEA results of these two comparisons was carried out using conservative thresholds of: FDR (Q-value) = 0.05; p-value = 0.001, and overlap coefficient cutoff = 0.5. The “annotate clusters” feature in EnrichmentMap v2.1.0 (build 522) was then used, with default “clusterMaker” MCL cluster parameters, to generate a list of gene-set clusters with two or more members. Guided by these automated clusters and those identified in the exploratory analysis in Fig 5A, we defined five gene-set clusters of functional interest. These are listed in S1 Table along with the associated GSEA statistics.
RNA extraction was performed using the RNeasy extraction kit (Qiagen) following manufacturers procedure. For blood samples, Trizol (Ambion, Life technologies) lysis was performed before kit purification. High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems) was used for the cDNA preparation. Quantitative PCR was performed using the LightCycler480 (Roche) using SYTO9 as a dye and the 2-ΔCt method to analyze expression difference [55]. Q-PCR PyMT in blood was detected using the following primers: Fwd: tgtgcacagcgtgtataatcc and Rv: tcatcgtgtagtggactgtgg; and confirmed with Fwd: taagaaggctacatgcggatgggt and Rv: ggcacctggcatcacatttgtctt; and housekeeping gene GAPD using the following primers Fwd: agcttgtcatcaacgggaag; and Rv: tttgatgttagtggggtctcg. Q-PCR for Elf5 was detected using Taqman probe Mm00468732_m1 or Hs01063022_m1 (Applied Biosystems), and housekeeping gene GAPD, Mm99999915_g1 or Hs99999905_m1; using the 7900H Fast Real-Time PCR system (Applied Biosystems).
For ELF5 and GFP immunohistochemistry, slides were blocked with protein block after antigen retrieval using Dako buffers (pH 6.1 at 125°C for 2 min, or pH9 at 100C for 25 min), followed with 0.05%Tween in PBS or 0.2% TritonX100. Primary antibodies were incubated for 1 h, ELF5 1:500 (N20, sc-9645, Santa Cruz) or GFP 1:200 (A11122, Invitrogen), then followed by either Rabbit anti-Goat 1:100 (Invitrogen) and LSAB+ label (Dako) or Envision Rabbit 30 min (Dako), then detection with DAB+ (Dako). For ELF5 IHC in patient samples, following blocking of the 4 micron paraffin-embedded sections from breast cancer TMAs for endogenous peroxidases, antigen retrieval was performed using pressure cook-microwaving in EDTA buffer (pH 9) for 5 min. This was followed by 0.02% Tween in PBS blocking for 5 min. Primary antibodies were incubated overnight with ELF5 antibody 1:70 in 0.1% BSA.PBS (N20 sc-9645, Santa Cruz) at room temperature. Detection was performed using 1:1000 Rabbit anti-Goat in 0.1% BSA.PBS (Invitrogen A10537) for 20 min, followed by Envison+ system-HRP labelled polymer anti-rabbit for 20 min (Dako 4003). DAB chromogen solution (Dako) was applied for 6 min followed by methyl green counterstaining. ELF5 nuclear and cytoplasmic staining assessment was performed using H-Score analysis that encompasses both percentage positivity and staining intensity on a 0–300 scale.
GFP and BrdU co-immunofluorescence antigen retrieval was pH 9 and 100°C for 25 min, followed by 0.2% TritonX100 then 1:250 GFP (A11122, Invitrogen), and 1:200 BrdU (M0744, Dako) at 4°C overnight. This was followed by 30min incubation with AlexaFluor 488-tagged anti-rabbit antibody, AlexaFluor 555-tagged anti-mouse antibody (1:200; Invitrogen) and ToPro (1:2000; Invitrogen).
Protein analyses by Western Blot were done as previously described [16]. Primary Antibodies used were anti-β-actin (AC-15, Sigma), anti-ELF5 (N20, sc-9645, Santa Cruz) and anti-V5 (R960-25, Invitrogen).
Boyden Chamber assays (Bencton Dickinson) were performed by plating 1x105 cells (PyMT) in media containing 0.5% FBS, the chemotactic gradient was established by placing the insets into full media (10%FBS) containing wells. Invading cells were visualized with the Diff Quick Stain Kit (Lab Aids). Area measured with Image J 1.41 (Wayne Rasband, US National Institutes of Health).
Imaging was conducted on an inverted Leica SP8 confocal microscope and the excitation source used was a Ti:Sapphire femtosecond pulsed laser (Coherent Chameleon Ultra II), operating at 80 MHz and tuned to a wavelength of 920 nm. 10 ul of blood tracer quantum dots blood tracers (655nm Life Technologies) were injected through the tail vein of the animals. Images were acquired with a 25x NA0.95 water objective. A dichroic filter (560 nm) was used to separate the GFP signal from quantum dot emission, which were further selected with band pass filters (525/50 and 617/73, respectively). Intensity was recorded with external RLD HyD detectors. For z-stacks, images were acquired at a format of 1,080 × 1,080 and a z-step size of 2.52 μm.
Sample comparisons have been made by unpaired Student’s t test using the GraphPad Prism software, La Jolla California USA. All error bars showed in this paper correspond to standard error (SEM) unless otherwise stated. All analysis in clinical samples were performed using the SPSS software (SPSS Inc. Chicago USA), assessment of the correlation between IHC markers was performed using Spearman rank order correlation and Mann-Whitney U test. Kaplan-Meier curves and log-rank test were used for survival analyses.
The patient cohort is a subset of the Nottingham series [37] comprising Luminal ER+ patients treated with tamoxifen but no chemotherapy, the distinction of luminal A or luminal B subtype was made according to the St Gallen criteria: n = 126 versus survival (74 luminal A, 52 luminal B); n = 129 versus DMFS (76 luminal A, 53 luminal B). Optimal staining cutpoints for analysis were selected using Xtile.
|
10.1371/journal.pgen.1001251 | Thymus-Associated Parathyroid Hormone Has Two Cellular Origins with Distinct Endocrine and Immunological Functions | In mammals, parathyroid hormone (PTH) is a key regulator of extracellular calcium and inorganic phosphorus homeostasis. Although the parathyroid glands were thought to be the only source of PTH, extra-parathyroid PTH production in the thymus, which shares a common origin with parathyroids during organogenesis, has been proposed to provide an auxiliary source of PTH, resulting in a higher than expected survival rate for aparathyroid Gcm2−/− mutants. However, the developmental ontogeny and cellular identity of these “thymic” PTH–expressing cells is unknown. We found that the lethality of aparathyroid Gcm2−/− mutants was affected by genetic background without relation to serum PTH levels, suggesting a need to reconsider the physiological function of thymic PTH. We identified two sources of extra-parathyroid PTH in wild-type mice. Incomplete separation of the parathyroid and thymus organs during organogenesis resulted in misplaced, isolated parathyroid cells that were often attached to the thymus; this was the major source of thymic PTH in normal mice. Analysis of thymus and parathyroid organogenesis in human embryos showed a broadly similar result, indicating that these results may provide insight into human parathyroid development. In addition, medullary thymic epithelial cells (mTECs) express PTH in a Gcm2-independent manner that requires TEC differentiation and is consistent with expression as a self-antigen for negative selection. Genetic or surgical removal of the thymus indicated that thymus-derived PTH in Gcm2−/− mutants did not provide auxiliary endocrine function. Our data show conclusively that the thymus does not serve as an auxiliary source of either serum PTH or parathyroid function. We further show that the normal process of parathyroid organogenesis in both mice and humans leads to the generation of multiple small parathyroid clusters in addition to the main parathyroid glands, that are the likely source of physiologically relevant “thymic PTH.”
| Due to the important role of PTH in the regulation of physiological activities, disorders in PTH production can cause many diseases in humans. Thus it is very important to understand where PTH is produced and how it is regulated. Many people have been found to have ectopic and supernumerary parathyroid glands without clear ontogenesis. In addition, the thymus, which develops together with the parathyroid during embryogenesis, has been proposed to be an auxiliary source of PTH with endocrine function; however, PTH is also a tissue-restricted self-antigen expressed by the thymus. In this paper, we provide insights into the ontogeny and function of thymus-associated PTH. We found that ectopic and supernumerary parathyroid glands originate from the normal developmental process underlying the separation of parathyroid and thymus, resulting in misplaced parathyroids close or attached to thymus. In the thymus, thymic epithelial cells can produce a low level of PTH via a different mechanism than the parathyroid and provide functional data that TEC-derived PTH does not have endocrine function. In summary, our data show that the thymic source of PTH has no endocrine function and, instead, has an expression pattern in the thymus consistent with that of a self-antigen for negative selection.
| Mammals have evolved an integrated system consisting of the parathyroid glands, bone, kidney and the intestine, to regulate ionized calcium and inorganic phosphorus homeostasis in the extracellular environment [1]. Circulating ionized Ca2+ and inorganic phosphorus are required for a wide range of physiological activities, including neuromuscular excitability, muscle contraction, energy storage, bone mineralization, blood coagulation and cardiovascular functions. Parathyroid hormone (PTH) produced by the parathyroids acts as the key endocrine regulator to modulate the physiological actions in the bone, kidney and the intestine to maintain the homeostasis of ionized calcium and inorganic phosphorus concentrations in the extracellular environment. Failure of calcium and phosphorus homeostasis, which can result from PTH production disorders, causes serious physiological consequences in human [2].
The parathyroid glands were long thought to be the sole source of PTH production and secretion. However, analysis of the aparathyroid Gcm2 null mouse mutant phenotype identified the thymus, a primary lymphoid organ, as an auxiliary source of circulating PTH in addition to the parathyroids in mice [3], [4]. Thymic PTH was found to come from small clusters of unidentified cells under the thymic capsule in wild-type mice, although the ontogeny of these intrathymic PTH-expressing cells and the regulation of PTH expression in these cells are not clear. In humans, ectopic parathyroid cells have been found in a variety of different locations, most commonly in the thymus [5], which was thought to account for the origin of intrathymic parathyroid adenomas in some patients [6]. However, the significance of thymus-associated PTH for the survival of Gcm2 mouse mutants was called into question by the phenotype of Pth null mutants, which can survive in the complete absence of PTH [7], [8].
Despite their distinct primary functions, the parathyroid and thymus organs have a close relationship during organogenesis, initially developing from two shared parathyroid/thymus primordia originating from the bilateral 3rd pharyngeal pouches [9]. Analysis of mouse mutants has shown that initial formation and early patterning of the thymus and parathyroid domains are controlled by a common regulatory pathway, including Hoxa3, Pax1, Pax9 and Eya1 [10], [11]. Once the organ domains are specified, their differentiation is regulated by two different organ-specific transcription factors, Gcm2 (for parathyroid) and Foxn1 (for thymus) [9]. In humans the bilateral 3rd and 4th pharyngeal pouches are thought to give rise to four parathyroids [12]–[14]; the pair of inferior parathyroid glands develop together with the thymus from the 3rd pharyngeal pouches, while the pair of superior parathyroid glands (not present in mice) develop with the ultimobranchial bodies from the 4th pharyngeal pouches [5]. Accessory parathyroids have also been reported in animals and in humans; although their origins were difficult to determine by histology alone, these structures were proposed to originate either during organogenesis, or to be induced postnatally in response to experimentally or surgically induced hypoparathyroidism [12]–[15]. Furthermore, intrathymic parathyroid adenomas have been hypothesized to originate from “ectopically migrating parathyroid cells” [6].
The original analysis of the Gcm2 mutant mouse reported that these mice were aparathyroid from embryonic stages [4]. Our subsequent analysis of the role of Gcm2 in parathyroid organogenesis showed that Gcm2 controls the differentiation and survival of parathyroid precursor cells, but is not required to specify the parathyroid domain within the pouch endoderm [3]. Without Gcm2 function, parathyroid precursor cells fail to differentiate and then undergo apopotosis by embryonic day 12, resulting in an aparathyroid phenotype [3], [4]. Mutation of Gcm2 in humans has also been associated with hypoparathyroidism [16], [17]. However, the role of Gcm2 in the development of extra-parathyroid PTH-expressing cells is as yet unknown.
To clarify the ontogenesis, regulation of PTH expression, and physiological role of extra-parathyroid PTH-expressing cells, we studied parathyroid and thymus organogenesis in the mouse. We showed that clusters of ectopic parathyroid cells between the parathyroid and thymus or attached to the thymus resulting from incomplete separation of these two organs during normal organogenesis. Analysis of parathyroid organogenesis in human embryos showed a similar phenomenon. Absence of these misplaced parathyroid cells in the thymus in Gcm2−/− mice caused a significant decrease of thymic PTH expression but still left a low level of thymic PTH expression, which we identified as originating from mTECs expressing PTH in a Gcm2-independent but Foxn1-dependent manner. Our results indicate that mTEC-derived PTH is not secreted into the general circulation and does not function as a backup mechanism of parathyroid glands, but may function as a self-antigen for negative selection. We further show that the lethality associated with Gcm2 mutation is not related to the presence of thymic PTH or serum PTH levels. Our results also have implications for the molecular mechanism of promiscuous gene expression of tissue-restricted self-antigens in mTECs. Our data also provide an explanation for the origin of ectopic parathyroid adenomas that are often associated with human hyperparathyroidism.
We compared the phenotypes of Gcm2−/− mutants on the C57BL/6J and 129/C57BL/6J F1genetic backgrounds for survival and parathyroid function. We found that Gcm2−/− mutants on a C57BL/6J genetic background had a nearly 100% lethality rate (Figure 1A), compared to 56% on the 129/C57BL/6J F1genetic background (Figure 1B) and to about 30% with additional backcross generations onto 129S6 (Table 1), confirming the original report [4]. Analysis of fetal parathyroid organogenesis in mutants from both genetic backgrounds confirmed our earlier data showing a complete absence of Pth-positive parathyroid cells [3] (Figure 2B, 2D and 2E). These data show that the reduced lethality on the 129/C57BL/6J hybrid background is not due to incomplete deletion of the parathyroids.
To test whether 129/C57BL/6J hybrid Gcm2−/− mutants had a higher serum PTH concentration than the C57BL/6J Gcm2−/− mutants that failed to survive, we measured serum PTH levels in E18.5 fetal Gcm2−/− mutants with different genetic backgrounds. Most Gcm2−/− mutants on both genetic backgrounds had undetectable serum PTH levels, with only a few individuals of each genetic background showing variable levels above the detection limit (3/23 for 129/C57BL/6J; 3/13 for C57BL/6J; Figure 1C). This dramatic reduction of serum PTH levels in Gcm2−/− mutants is consistent with other reports on a variety of genetic backgrounds [18], [19]. These results show that serum PTH levels in the Gcm2−/− mutants are not affected by genetic background, and that the lethality phenotype observed in Gcm2−/− mutants is not related to serum PTH levels.
Heterozygotes on the 129/C57Bl6 hybrid genetic background also had low or undetectable serum PTH levels. This difference in steady-state PTH levels was not correlated with differences in maternal ionized calcium levels, which were similar in heterozygote and wild-type females from both strains, and parathyroid glands in heterozygotes from the hybrid background were histologically normal (data not shown). Variations in PTH levels have been reported between C3H/HeJ and C57BL/6 mice, including change in PTH levels in response to altering the calcium content of the diet, as well as differences between strains in BMD, calcium absorption, serum calcium, and calcitriol levels [20]. As serum chemistry was normal, this result further supports our observation that serum PTH levels do not correlate with lethality.
To investigate the possible role of thymic PTH in the lethality of Gcm2 mutants, we designed experiments to determine the ontogenesis of extra-parathyroid PTH-expressing cells. Since the parathyroids and thymus arise from the same embryonic structure, we tracked the process by which the parathyroid and thymus domains resolve into separate primordia in mice using in situ hybridization for Pth and Gcm2. At E12, Gcm2/Pth expression in the parathyroid/thymus common primordia specifically marked the anterior/dorsal Gcm2-positive parathyroid domain with a clear interface at the posterior/ventral Foxn1-positive thymus domain [3], [21](Figure 2A-I). At E13, the Gcm2/Pth-positive parathyroid domain had started to separate from the thymus domain, and some parathyroid cells were located outside the major parathyroid domain (Figure 2A-II). At E18.5, small clusters of parathyroid cells were located between the parathyroids and thymus or directly associated with the thymus, in some cases under the developing thymic capsule (Figure 2A-III and 2C). This phenotype was seen in all 11 E16.5-18.5 wild-type embryos on multiple genetic backgrounds (C57BL/6J, 129/C57BL/6J F1 hybrid, or 129S6; Figure 2A–2D), which indicates that this incomplete separation pattern is a common phenomenon in the mouse.
RT-PCR using cDNA made from total thymus and other organs from wild-type mice confirmed that co-expression of Gcm2 and Pth was detected only in the thymus (Figure 3A). Gcm2 and Pth expression could be detected as early as E13.5 in dissected whole thymus, when the thymus had just separated from the parathyroids, and at all later stages (Figure 3B).
If these misplaced Gcm2/Pth-positive cells are authentic parathyroid cells, Gcm2 should regulate their differentiation and survival [3]. As predicted, all misplaced parathyroid cells were ablated in Gcm2−/− mutants (Figure 2B, 2D and 2E). The thymic Pth expression level was also greatly reduced relative to wild-type, while the expression of the TEC marker Foxn1 was not affected (Figure 3C). These data suggest that misplaced parathyroid cells in the thymus are the primary source of thymic PTH in wild-type mice, and that these cells are absent in Gcm2−/− mutants.
To test whether a similar phenomenon occurs in human embryogenesis, we used whole-mount in situ hybridization for Gcm2 in early week 6 to mid week 8 human embryos or dissected parathyroids and thymic lobes. At early week 6, Gcm2 was expressed in the dorsal region of the 3rd and 4th pharyngeal pouches (Figure 4A, 4B; 2/2 embryos). By early week 7 the common parathyroid/thymic primordia (derived from the 3rd pharyngeal pouch) have detached from the pharynx. Throughout week 7, clusters of Gcm2 expressing cells were located in the anterior portion of the common primordium (4/4 embryos) and at the posterior tip (1/4 embryos) of the migrating elongated thymic structure (Figure 4C–4H). Similar to the phenomenon we found in mouse (Figure 2A-II), small clusters of ‘stray’ GCM2-positive cells were often present (Figure 4E, *). By late week 7, although Gcm2 positive cells were still attached to the common primordia, separate parathyroids were present, as well as small Gcm2 expressing clusters that may represent accessory parathyroids (Figure 4F–4H; 1/1 embryo).
Dissected parathyroids and thymic lobes from one side of early and mid week 8 embryos showed three major Gcm2 expressing parathyroids (black arrowheads) and a smaller Gcm2 expressing accessory parathyroid (red arrowhead) associated with a single thymic lobe (Figure 4I–4K; 3/3 embryos). Of the major parathyroid rudiments, one is clearly associated with the thymic primordium at late week 7 and therefore appears to derive from the 3rd pharyngeal pouch, while the other is clearly outside the common thymus-parathyroid primordium and thus most likely derives from the 4th pharyngeal pouch (note that GCM2 expression is clearly evident in the 4th pouch at week 6) (Figure 4I–4K). In addition, a smaller parathyroid rudiment was consistently observed associated with the posterior tip of the thymus domain of the common primordium at week 7 and week 8 (Figure 4D–4J), although at least in late week 7 this appeared to be present only in one of the two bilateral primordia (Figure 4H). Furthermore, as the parathyroid separates, some Gcm2 expressing cells are left attached to the upper cordlike thymic structure (Figure 4J, 4K, white arrows). These data demonstrate that similar to our observations in the mouse, ectopic parathyroids exist from week 7 in the human embryo, and that the presence of intrathymic parathyroids in adulthood may be in part due to incomplete separation from the thymus.
RT-PCR using total thymus cDNA from Gcm2−/− mice could still amplify Pth at high cycle numbers (Figure 3C), suggesting that the misplaced parathyroid cells were not the only source of thymic PTH. Quantitative RT-PCR using total thymus cDNA from wild-type and Gcm2−/− mice on a C57BL/6 genetic background showed that the second source of thymic PTH in the Gcm2−/− mice is about 1/350 of the level in the wild-type mice on the C57Bl/6 background (Figure 3D). We therefore investigated this Gcm2-independent source of thymic PTH expression.
The thymus is a complex immune organ composed of hematopoietic cell-derived thymocytes and multiple types of stromal cells [22]. TECs play a required role in the production of a self-restricted and self-tolerant T-cell repertoire through positive selection and negative selection [22]. Negative selection occurs in the medullary region, where medullary TECs (mTECs) promiscuously express many tissue-restricted self-antigens (TRAs) that are required for negative selection to establish central tolerance and prevent autoimmunity [23]. To test whether thymic PTH expression was due to TRA expression in mTECs, we performed RT-PCR on sorted TECs (Figure 5). TECs expressed both Foxn1 and Pth, and expression levels were similar in TECs sorted from wild-type controls and Gcm2−/− mutants (Figure 5B and 5C). We did not detect Gcm1 or Gcm2 expression in the purified TECs (Figure 5C), indicating that Pth expression in these cells is not controlled by Gcm2, and arguing against a previously proposed role for Gcm1 in regulating thymic PTH expression [4].
Pth expression was not found in other thymic cell types by RT-PCR or microarray analyses, including T cells, macrophages, and dendritic cells (Figure 5B) [24]. Microarray data from sorted mTECs or cTECs also showed that Pth transcripts were present only in mTECs [24]. We further confirmed the expression of Pth in mTECs using Rag2−/− mutant mice, which have a normal cortical structure but lack an organized medulla [25], [26]. Thymic Pth expression was greatly reduced in Gcm2−/−;Rag2−/− double mutants (Figure 5D), although not totally ablated, consistent with the incomplete block in mTEC differentiation in Rag2 mutants.
As a genetic test of the TEC origin of thymic Pth expression, we generated Gcm2−/−;Foxn1nu/nu double mutant mice that have no parathyroids and in which TEC differentiation is blocked [27]. We failed to detect any thymic Pth expression in the thymic epithelial rudiments of Gcm2;Foxn1 double mutants (Figure 5E). These results further supported the conclusion that thymic Pth expression has only two sources: misplaced authentic parathyroid cells that express Pth in a Gcm2-dependent manner; and differentiated mTECs that express Pth independent of Gcm2.
The initial report of the Gcm2 single mutant phenotype invoked the 100% neonatal lethality of Hoxa3 mutants, which are aparathyroid and athymic, in support of the proposal that thymus-derived PTH ameliorated the lethality phenotype of Gcm2 mutants [4]. As Hoxa3 mutants have a variety of other defects that could contribute to lethality [28], [29], we used the Gcm2−/−; Foxn1nu/nu double mutants as a more appropriate test of this possibility. These double mutants have a specific genetic deletion of both parathyroids and thymus, without any known potentially confounding phenotypes. In double heterozygote intercrosses, all genotypes were present in the expected Mendelian ratios at the newborn stage. Adult mice had reduced numbers of genotypes homozygous for the Gcm2 mutation (Table 2), consistent with the rate of lethality of Gcm2−/− mutants on this mixed genetic background. Surprisingly, compared with a survival rate of about 55% for Gcm2−/− mutants in these crosses, Gcm2−/−;Foxn1nu/nu double mutants had a lower survival rate of about 18% (Table 2). However, the ionized calcium and inorganic phosphorus concentrations in both newborn and adult mice were not significantly different between wild-type and Foxn1nu/nu mutant mice, or between Gcm2−/− mutants and Gcm2;Foxn1 double mutants (Figure 6C–6F). These results indicate that Foxn1-dependent Pth expression in mTECs does not contribute to serum calcium physiology. While the reason for the increased lethality of double mutants is as yet unclear, these data provide further evidence that the lethality phenotype of Gcm2−/− and Gcm2−/−Foxn1−/− mutants was not PTH-related.
The initial report of the Gcm2 null mutant phenotype showed that surgical removal of both the thymus and parathyroids from wild-type adults resulted in lethality [4]. As our data shows that parathyroid cells are normally associated with the thymus due to the incomplete organ separation during development, this result could have been due to the removal of thymus-associated parathyroids, rather than to the removal of thymus-produced PTH. Although there was no detectable serum PTH in most Gcm2 mutants (Figure 1C), we tested whether thymic PTH participates in endocrine function by determining whether the removal of the thymus from Gcm2−/− mutants would increase lethality on the 129/C57BL/6J hybrid background. First, we performed thymectomy surgery on newborn Gcm2−/− mutants on the 129S-C57Bl/6 genetic background. Unmanipulated and mock surgery groups from the same 129S6/C57BL/6J hybrid genetic background were used as controls. Thymectomized Gcm2−/− mutants did not show increased lethality (Table 1), and serum biochemistry did not show any difference in ionized calcium or inorganic phosphorus levels between surviving Gcm2−/− mutants with mock surgery and Gcm2−/− mutants with thymectomy (Figure 6A, 6B). These data, in combination with the analysis of Gcm2;Foxn1 double mutants, therefore demonstrate that the thymus does not provide any PTH-related endocrine function in mice.
Our data reveal two cellular sources of extra-parathyroid PTH. The first source is misplaced authentic parathyroid cells that arise during normal organogenesis, which express PTH in the same way as the parathyroid glands and are ablated in the Gcm2 null mutants. The second source is mTECs, which express PTH independently of Gcm2, but dependent on Foxn1-mediated TEC differentiation. We also define two different physiological functions for the PTH derived from these two different sources. We propose that parathyroid cells, including those in the main parathyroid glands and the misplaced parathyroid cells, are the only physiologically relevant postnatal source of serum PTH, and that the thymus has no contribution to serum PTH or calcium physiology. mTECs also express PTH, probably as a self-antigen, but this PTH does not contribute to serum PTH for endocrine function. This result is consistent with the lack of secretory machinery in these cells used in the parathyroid cells to secrete PTH into the circulation [30]–[32], and the likelihood that the PTH translated in the mTECs is degraded into short peptides to be used for negative selection.
Based on our observations in both mouse and human, the separation process of the parathyroids from the thymus results in multiple “micro-parathyroids” in addition to the main parathyroid glands. Parathyroid adenomas have been found in the human thymus, and have been shown to express Gcm2, indicating that intrathymic adenomas could be the result of uncontrolled growth of the misplaced parathyroid cells [5], [6], [33]. These misplaced parathyroid cells could receive signals from the inappropriate microenviroment, causing them to secrete high PTH or over-proliferate; alternatively, these small groups of parathyroid cells may respond inappropriately to homeostatic mechanisms.
Our analysis presents the first genetic marker study of human parathyroid development, and reveals new information about parathyroid development that differs from the original descriptions of human parathyroid organogenesis based on histological studies. Our results on the ontogeny of these extra-parathyroid PTH-expressing cells provides insight into understanding the etiology of some hyperparathyroid disorders caused by ectopic parathyroid glands and intrathymic parathyroid adenomas [5], [6]. It is widely accepted that in humans four parathyroids develop during embryogenesis, giving rise to the superior and inferior parathyroids, and that ectopic and supernumerary parathyroids, often associated with the thymus, can cause primary hyperparathyroidism due to hyperplasia, adenomas, and carcinomas [5], [34]. Our data indicate that more than four major parathyroid rudiments are present by week 7 in the human fetus and that accessory parathyroids are present in the majority of fetuses at week 7 to week 8, and are therefore more frequent than previously documented [12], [35]. Our ability to identify these additional parathyroid structures is due to the increased resolution of analysis provided by in situ hybridization. Morphological studies would not identify all of the smaller accessory parathyroids, and may have annotated some parathyroid primordia as other structures.
In our view, it is not yet possible to definitively determine the relationship of the four GCM2-positive structures at week 8 to those present at late week 7, since the thymus and parathyroid primordia are actively migrating at this stage in development and lineage tracing studies are not possible in human embryos. The two GCM2-positive structures associated with the anterior end of the thymus primordium at week 8 may correspond to the major parathyroid foci present at late week 7 (i.e. to the structures we assign as arising from the 3rd and 4th pharyngeal pouches), while those at the posterior tip of the thymus primordium correspond to the Gcm2-positive clusters at the posterior in week 7. However, it is also possible that these both anterior parathyroids arise from the 3rd pharyngeal pouch and that the 4th pouch-derived parathyroid is no longer associated with the thymus primordium and therefore is not dissected out along with the thymus. In our view, the question of which of the parathyroid rudiments present in human fetal development give rise to the inferior and superior parathyroids in the adult also remains open, and will require further detailed study for its resolution.
Our original interest in this project was piqued by the differences in lethality between the Gcm2 null mutants on different genetic backgrounds, from ∼30–60% on 129/C57BL/6J hybrid background (this report; [4]), to nearly 100% on the C57BL/6J background (this report). The original report of the Gcm2 null mutants proposed that that thymic PTH could in part rescue an aparathyroid lethality phenotype; the lethality of thyroid-parathyroid-thymectomy in adult wild-type mice, and of athymic and aparathyroid Hoxa3 null mutants were listed in support of this model [4]. It was later shown that Pth−/− mice can survive [7], [8], calling into question the assumption that aparathyroidism in mice would necessarily be lethal. Our data indicate that the lethality phenotype in Gcm2−/− mutants is not related to serum PTH levels; however, the question of why these mutants die is still not resolved. There are also unexplained differences in survival of different parathyroid-related mutants on the same background. A recent study from the Kovacs lab suggests that the measurable levels of PTH in some Gcm2−/− mutants at fetal stages may reflect Gcm2-independent PTH originating from the placenta. However, other than increased placental calcium transport, the Gcm2 and Pth null mutants have very similar phenotypes, both of which are milder than the Hoxa3 null when all are on the Black Swiss genetic background [19]. Hoxa3 null mutants have phenotypes similar to the PTH/PTHrP null, which also die at birth, suggesting that Hoxa3 may play some role in calcium physiology outside the parathyroid. In part, the answer may lie outside of calcium physiology. Hoxa3 mutants have other defects that could contribute to lethality [28], [29]. The increased postnatal lethality of the Gcm2;Foxn1 double mutants may also be due to as yet unidentified functions for both of these transcription factors in other tissues; expression of both Foxn1 and Gcm2 has been identified in the postnatal central nervous system (http://mouse.brain-map.org/brain/Foxn1.html; http://www.ncbi.nlm.nih.gov/projects/gensat/).
Two models have been proposed for the mechanism that regulates the promiscuous expression of TRAs in mTECs. The progressive restriction model proposes a mosaic of TRA expression in immature mTECs characteristic of multi-lineage differentiated cells of endoderm-derived organs, and expressed by the same tissue-specific regulators as in their ‘normal’ tissues [36]. In contrast, the terminal differentiation model proposes that some mTECs have an autonomous property to express TRAs by a different mechanism compared to their tissue-specific regulation, characterized by lower transcriptional levels and independence from tissue-specific transcriptional regulators [24]. This latter model is supported by single-cell PCR of individual mTECs, and by the analysis of casein beta gene expression in mTECs compared to mammary gland cells [37]. Given the pharyngeal endodermal origin of parathyroid cells, PTH should be a good candidate for the progressive restriction model, as discussed above. However, our data showed a much lower Pth expression level in mTECs and a Foxn1-dependent and Gcm2-independent pathway for PTH expression in mTECs, more consistent with the terminal differentiation model; Gcm2-dependent PTH in the thymus came exclusively from misplaced parathyroid cells. Microarray analysis indicates that PTH expression in mTECs is Aire-independent [38], consistent with immunolocalization studies [39]. It is still an open question whether the regulation mechanism for thymic PTH in mTECs is common to other Aire-independent TRAs.
All experiments using mice were carried out at UGA with the approval of the UGA Institutional Animal Care and Use Committee. First and second trimester human fetuses were obtained in collaboration with the Reproductive Biology Unit, Little France, Edinburgh. Ethical approval for use of human fetal tissue was granted by the Lothian University Hospitals NHS Trust and the Lothian Research Ethics Committee: Smith et. al. ‘Isolation and propagation of fetal stem cells’ LREC/2002/6/15. Consent was obtained from all donors, and the tissue was anonymized before being made available for research. Use and disposal of tissues are strictly regulated in accordance with conditions stipulated in the Ethics approval and in the University of Edinburgh Health and Safety regulations regarding use of human tissue. All experiments using human tissue were performed at the University of Edinburgh.
The generation and genotyping of Gcm2 null mutant has been described [4]. Gcm2 mutant mice on a 129/SvEv-C57BL/6J genetic background were backcrossed to C57BL/6J mice for more than 5 generations. These majority C57BL/6J Gcm2 mutant mice were then backcrossed to 129S6 mice (Taconic) to obtain 129S6/C57BL/6 F1 hybrids.
Foxn1-nude mice (Jackson Labs) and R26YFP reporter mice [40] were maintained on a C57BL/6J and 129SvJ hybrid background. C57BL/6J Rag2 null mutant mice were a generous gift from Dr. E. V. Rothenberg. The Foxn1Cre allele of Foxn1 was previously described [41].
Analysis of Gcm2;Foxn1 double mutants was done by mating Gcm2+/−; Foxn1+/nu males with Gcm2+/−; Foxn1+/nu females. A total of 182 one month old mice from Gcm2+/−; Foxn1+/nu mating were genotyped at weaning. Reduced survival of genotypes homozygous for the Gcm2 mutation was significant using the chi-square test. Since there were no survival defects detected in Gcm2+/− or Foxn1+/nu heterozygous mice, we combined heterozygous mice with wild-type mice as a control group.
In all crosses, for calculating the % survival, the survival of wild-type mice was set at 100%. For staging of embryos, noon on the day of the vaginal plug was designated as E0.5.
Embryos were staged according to the standard head/rump measurement and classified according to Carnegie stages. Embryos used in this study were from week 6 (Carnegie stage 16–17), week 7 (Carnegie stage 18–19) and early to mid-week 8 (Carnegie stage 20–21). Embryos were fixed in 4% PFA for 24 hours and stored at −20°C in 100% methanol until used for analysis.
Isolation of RNA and RT-PCR were performed as described [42]. Tissues were dissected from embryos, newborns, or adult mice and total RNA was isolated with Trizol. Genomic DNA was removed using DNase I. Reverse transcription was performed using SuperScript III Reverse Transcriptase (Invitrogen), then cDNA was subjected to PCR. The following primers were used: β-actin forward 5′-TGGAATCCTGTGGCATCCATGAAAC-3′, β-actin reverse 5′-TAAAACGCAGCTCAGTAACAGTCCG-3′, Pth forward 5′-CTGCAGTCCAGTTCATCAGC-3′, Pth reverse 5′-AAGCTTGAAAAGGTAGCAGCA-3′, Gcm2 forward 5′-CATCAATGACCCACAGATGC-3′, Gcm2 reverse 5′-GGCACTTCTTCTGCCTTCTG-3′, Foxn1 forward 5′-TGACGGAGCACTTCCCTTAC-3′, Foxn1 reverse 5′-GGGAAAGGTGTGGGTAGGTC-3′, Gcm1 forward 5′-TGAAAAACAAGCCCTTCAGC-3′ and Gcm1 reverse 5′-TCTGGCTTTGTCACAGATGG-3′. Both Gcm2 and Pth RT-PCR products were confirmed by sequencing.
Paraffin section in situ hybridization for Gcm2 and Pth was performed as described [3]. Staged embryos were fixed in 4% paraformaldehyde overnight and processed for paraffin embedding. 8-10 µm sections were hybridized with digoxigenin-labeled RNA probes at 0.5 µg/ml. Alkaline phosphatase-conjugated antidigoxigenin Fab fragments were used at 1∶5000. BM-purple (Roche) was used as a chromagen to localize hybridized probe. Nuclear fast red was used as a counterstain.
Whole-mount in situ hybridization on human fetal tissue was performed as described [21]. Gcm2 probes used were generated by PCR amplification from microdissected human fetal thymic/parathyroid tissue using the following primers: Gcm2F, 5′-GGGCCACCTCCTATGAAAAT-3′; Gcm2R, 5′-GCAGCCTCTAGGGATGTGAA-3′. NBT/BCIP (Roche) was used to localize the hybidized probe. Embryos were embedded in paraffin and sectioned after staining in whole mount.
Thymic stromal cell isolation was modified from a previously described method [43]. Thymi from Foxn1+/Cre;R26-YFP+/tg mice were dissected, minced into small pieces and agitated in RPMI1640 with 2% FBS to remove most thymocytes. The remaining tissue pieces were collected and resuspended in RPMI1640 and 0.2 mg/ml collagenase for 20 minutes at 37°C with gentle stirring. The tissue pieces were allowed to settle for 5 minutes, the supernatant was discarded, and the tissue was resuspended in dispase media (0.2 mg/ml of dispase, 0.2 mg/ml of collagenase and 25 ug/ml of DNaseI in RPMI 1640) for 20 minutes at 37°C with gentle stirring. The supernatant was discarded and the tissue chunks resuspended in fresh dispase media for 30–45 minutes at 37°C. The digested products were then passed through a 25 G needle, centrifuged at 800× g for 3minutes, resuspended in PBS containing 2% FBS and 5 mM EDTA, and then filtered through a 70 um cell strainer. The filtered cells were stained with anti-mouse CD45-PE (BD pharmingen) antibody before being subjected to sorting using a MoFlo cell sorter (Dako) to isolate PE-, YFP+ TECs. The yield of TECs was about 20,000 cells per adult thymus, with about 93% purity.
Total RNA from sorted TECs was extracted with the RNeasy Micro kit according to manufacturer's instructions (QIAGEN). Total RNA from whole thymi was isolated with Trizol (Invitrogen). First-strand cDNA was reverse transcribed using superscript III (Invitrogen). Quantitative PCR was performed on an ABI 7500 real time PCR system with Taqman universal PCR mix (Applied Biosystems). 18S rRNA VIC/TAMRA primer-probe (Applied Biosystems) was used as endogenous control. Pth FAM primer-probe (Assay ID: Mm00451600-g1) was purchased from Applied Biosystems. PCR was performed at 50°C, 2 min; 95°C, 10 min; 40 cycles of 95°C for 15 sec; 60°C for 1 min. The relative quantity of gene expression was determined using 7500 SDS software (Applied Biosystems).
Serum sample collection from E18.5 fetal, newborn or adult mice has been described [44], [45]. For fetal or newborn mice, the neck was incised to transect the carotid and jugular, and whole blood was collected into plain capillary tubes. For adult mice, blood samples were collected into capillary tubes from tail vein right after mice were sacrificed by cervical dislocation, or a cardiac puncture was used to obtain larger samples. Serum samples were prepared by centrifugation to remove blood cells, then stored at −20°C until assayed. The inorganic phosphorus and ionized calcium levels were measured using kits 117-30 (for phosphorus) and 140-20 (for calcium) from Diagnostic Chemicals Limited (Canada). Serum PTH was measured with a rodent PTH 1-34 Elisa kit with a detection limit of 1.6 pg/ml (Immutopics, San Clemente, CA). PTH values that were below the detection limit of 1.6 were reassigned a value equal to the detection limit.
Neonatal thymecotomy was performed as described [46]. Each newborn pup was chilled on ice for 1 minute, until unresponsive. A small incision was made in the center of the throat. The submandibular gland and muscle were moved aside with forceps, and the top portion of the sternum cut to expose the thymus. The thymus was removed with a kimwipe-covered toothpick, and the sternum and skin closed with surgical adhesive (3 M Vetbond, No. 1469SB, 3 M Animal Care Products, St. Paul, MN, USA). Pups were revived on a 37°C warming plate, then returned to their mother. Mock surgeries were performed without removing the thymus. All mice were allowed to grow until 1 month of age, then serum samples were prepared for serum biochemistry as described above.
|
10.1371/journal.ppat.1003776 | The Neonatal Fc Receptor (FcRn) Enhances Human Immunodeficiency Virus Type 1 (HIV-1) Transcytosis across Epithelial Cells | The mechanisms by which human immunodeficiency virus type 1 (HIV-1) crosses mucosal surfaces to establish infection are unknown. Acidic genital secretions of HIV-1-infected women contain HIV-1 likely coated by antibody. We found that the combination of acidic pH and Env-specific IgG, including that from cervicovaginal and seminal fluids of HIV-1-infected individuals, augmented transcytosis across epithelial cells as much as 20-fold compared with Env-specific IgG at neutral pH or non-specific IgG at either pH. Enhanced transcytosis was observed with clinical HIV-1 isolates, including transmitted/founder strains, and was eliminated in Fc neonatal receptor (FcRn)-knockdown epithelial cells. Non-neutralizing antibodies allowed similar or less transcytosis than neutralizing antibodies. However, the ratio of total:infectious virus was higher for neutralizing antibodies, indicating that they allowed transcytosis while blocking infectivity of transcytosed virus. Immunocytochemistry revealed abundant FcRn expression in columnar epithelia lining the human endocervix and penile urethra. Acidity and Env-specific IgG enhance transcytosis of virus across epithelial cells via FcRn and could facilitate translocation of virus to susceptible target cells following sexual exposure.
| HIV-1 causes a sexually transmitted disease. However, the mechanisms employed by the virus to cross genital tract tissue and establish infection are uncertain. Since cervicovaginal fluid is acidic and HIV-1 in cervicovaginal fluid is likely coated with antibodies, we explored the effect of low pH and HIV-1-specific antibodies on transcytosis, the movement of HIV-1 across tight-junctioned epithelial cells. We found that the combination of HIV-1-specific antibodies and low pH enhanced transcytosis as much as 20-fold. Virus that underwent transcytosis under these conditions was infectious, and infectivity was highly influenced by whether or not the antibody neutralized the virus. We observed enhanced transcytosis using antibody from cervicovaginal and seminal fluids and using transmitted/founder strains of HIV-1. We also found that the enhanced transcytosis was due to the Fc neonatal receptor (FcRn), which binds immune complexes at acidic pH and releases them at neutral pH. Finally, staining of human tissue revealed abundant FcRn expression on columnar epithelial cells of penile urethra and endocervix. Our findings reveal a novel mechanism wherein HIV-1 may facilitate its own transmission by usurping the antibody response directed against itself. These results have important implications for HIV vaccine development and for understanding the earliest events in HIV transmission.
| Sexual transmission of HIV-1 requires that virus establish infection across genital tract or intestinal tissue. Sexually transmitted infections, other causes of inflammation, and localized trauma may allow susceptible CD4+ target cells at skin or mucosal surfaces to become directly exposed to secretions from infected sexual partners [1], [2]. However, when skin and mucosa are intact, it remains unclear precisely how HIV-1 gains access to target cells. One possibility is that virus translocates between epithelial cells until susceptible cells are found either in or below the epithelium [3]. Alternatively, Langerhans cells may sample the surface, acquire virus, and move it to areas of abundant target cells [4], [5]. Finally, transcytosis of HIV-1 (i.e., movement through cells) has been studied as a potential mechanism to translocate virus from mucosal surfaces to deeper-lying CD4+ cells [6], [7], [8].
Transcytosis offers an explanation for movement of virus across epithelial cells forming tight junctions, which might normally exclude pathogens from moving beyond the surface. However, in vitro, only a very small amount of virus, usually less than 0.3% of a cell-free virus inoculum, finds its way through cells into the medium bathing basolateral surfaces ([9]). Interactions between HIV-1 Env and several host cell surface molecules, including glycolipids, heparan sulfate proteoglycans and gp340, have been proposed to play a role in transcytosis [10], [11], [12], [13], [14].
With the exception of the acute phase prior to development of anti-HIV-1 immune responses, semen, cervicovaginal, and rectal fluids from HIV-1-infected individuals contain antibodies against HIV-1 Env [15], [16], [17]. The concentration of Env-specific IgG present in such secretions varies considerably from person to person and is usually on the order of 100 to 1,000-fold less than concentrations found in plasma [18]. The presence of Env-specific IgG strongly suggests that some proportion of Env molecules on the surface of infectious virions in genital tract secretions is coated with IgG. Since HIV-1 is successfully transmitted sexually, the coating antibody is either of insufficient quantity or quality to neutralize virus infectivity upon contact with an uninfected partner.
Antibody in genital tract secretions of HIV-1-infected individuals could play a role in facilitating the transport of virus across mucosal epithelia. Such a role is made particularly plausible by the reported expression of the Fc neonatal receptor (FcRn) in human genital mucosal tissue [19]. FcRn is a heterodimeric receptor belonging to the MHC class I family of proteins [20], [21]. The expression of FcRn in endothelial cells is thought to be critical for IgG homeostasis in blood [22], and its expression in placental syncytiotrophoblasts is a key factor in transporting maternal IgG to the fetal circulation [23]. A characteristic of FcRn is its ability to bind the Fc region of IgG at acidic pH and to release it at neutral pH [24]. This pH-dependent binding allows the transport of intact IgG or of IgG immune complexes from luminal surfaces bathed in acidic fluids, for example, cervicovaginal secretions, to basolateral surfaces exposed to a neutral intracellular milieu [19].
Cervicovaginal secretions are maintained at acidic pH by acid-producing bacteria that make up part of the normal vaginal microbiota [25]. Although perturbations of normal microbiota, such as occur with bacterial vaginosis, raise the pH, the secretions generally remain in the acidic range [26], [27]. Semen rapidly neutralizes cervicovaginal secretions, but the extent of the pH change is variable. For example, a large amount of ejaculate may raise the pH to the neutral range, whereas a small amount may not [28], [29]. The pH of rectal secretions ranges from about 6.8 to 7.2 [30].
Given that HIV-1 in genital tract secretions may be complexed with IgG antibody, that female genital tract secretions are acidic, and that FcRn has been demonstrated in genital tract tissues, we evaluated the role of pH and antibody on transcytosis of HIV-1 through polarized epithelial cells.
To investigate the effect of low pH and antibody on HIV-1 transcytosis across epithelial cells forming tight junctions, we exposed the apical surface of HEC-1A cells to HIV-1 at pH 6.0 or 7.4 with or without HIV-1-specific IgG (HIVIG). Virus was quantified in the medium bathing the basolateral cell surface (“subnatant fluid”) by RT-PCR and, although detectable as early as six hours after exposure of virus to the apical cell surface, the quantity was greater at 12 hours (Figure S1) [8]. Thus, in subsequent experiments, transcytosis was measured at 12 hours. Using HIV-1US712, a clade B R5 clinical isolate, HIVIG enhanced transcytosis in a dose-dependent manner when virus and antibody were exposed to the apical surface at pH 6.0 (Figure 1A). There was no increase in transcytosis with HIVIG at pH 7.4 or with HIV-negative IgG (IVIG) at pH 6.0 or 7.4. We found similarly enhanced transcytosis using additional R5 as well as X4 and X4/R5 strains (Figure 1B). Importantly, transcytosis of four of five clade C transmitted/founder Env-pseudotyped viruses was enhanced in a pH and antibody-dependent manner (Figure 1C). Enhanced transcytosis with HIV-1-specific antibody at low pH also occurred with T84 colon carcinoma cells (Figure S2).
Since sexual transmission may occur with small amounts of virus, we investigated if pH- and antibody-dependent enhancement of transcytosis could occur at very low HIV-1 inocula. Using HIVIG or the anti-gp41 monoclonal antibody (mAb) 2F5, we found that transcytosis occurred with virus inocula as low as 2 pg of p24 (about 60,000 RNA copies) with HIVIG and 0.02 pg of p24 (about 500 RNA copies) with 2F5, amounts too small to be detectable in subnatant fluid in the absence of low pH and HIV-1-specific antibody (Table 1). These quantities of virus are within the range observed in seminal and cervicovaginal fluids of HIV-infected individuals [31], [32], [33].
The impact of both antibody and low pH suggested FcRn involvement [34], [35]. We knocked down FcRn in HEC-1A cells, verifying lower expression by flow cytometry and by Western blot (Figure S3A). The knock-down HEC-1A cells attained the same level of electrical resistance as did the wild-type cells (data not shown), indicating that FcRn knockdown did not affect the ability to form tight junctions. Unlike with wild-type HEC-1A cells, there was no enhanced transcytosis with FcRn-knockdown HEC-1A cells when either mAb 2F5 (Figure 2A) or polyclonal HIVIG (data not shown) were used. We also evaluated Fc mutants of the HIV-1 Env-specific mAb b12. A mutant designed to abrogate FcRn binding (I253A), markedly lowered transcytosis compared with wild-type b12 (Figure 2B) [36]. The second mutant (M428L), designed to bind with higher affinity to FcRn, increased transcytosis compared with wild-type b12 (Figure 2B) [37]. Binding to HIV-1JRFL gp120 (Figure S3B) and neutralization of HIV-1JRFL (Figure S3C) were nearly equivalent for the wild-type and Fc mutant versions of b12, indicating that the Fc mutations did not affect Fab-antigen binding. Blockade of FcRn with anti-FcRn antibody and inhibition of endosomal acidification by bafilomycin A1 also substantially reduced or eliminated enhanced transcytosis (Figure S3D–F), as did competition between the non-HIV-1-specific mAb Den3 and the anti-HIV-1 Env mAb VRC01 (Figure S3G). Consistent with other investigations of Fc-FcRn interactions, maximally enhanced transcytosis occurred at pH 5.5–6.0, with some enhanced transcytosis apparent at pH 4.5 and 6.5 (Figure S3H) [38], [39]. Since FcRn binds to IgG and not to IgA, we compared two different IgG1 mAbs, b12 and HGN194, with their IgA class-switched versions. Both IgG1 mAbs enhanced transcytosis of HIV-1JRFL pseudoviruses and SHIV1157ipEL-p at pH 6.0, whereas the IgA class-switched versions did not (Figure 2C and 2D). In fact, as reported, dimeric IgA1 HGN194 inhibited transcytosis [40]. Thus, enhanced transcytosis at low pH in the presence of specific antibody is mediated by IgG and is dependent on FcRn.
Using 50 µg/ml of VRC01 or Den3, transcytosis of IgG alone increased approximately 3 fold from about 0.4% at pH 7.4 to about 1.3% at pH 6.0 (Figure S4). However, the effect of FcRn-mediated transcytosis on IgG alone does not appear as strong as the effect on IgG immune complexes, where, for example, with complexes made with 50 µg/ml of VRC01 or 2F5, there was about an 8-fold increase in transcytosis under conditions allowing FcRn engagement. This difference may be due to the contribution of fluid phase uptake of IgG by the epithelial cells at both pH 6.0 and pH 7.4; IgG thus internalized can engage FcRn in acidic endosomes and be shuttled to the basolateral side of the cells [41]. The internalization of immune complexes, on the other hand, likely depends primarily on FcRn engagement at the surface of the cell at pH 6.0.
We evaluated the ability of HIVIG and a panel of mAbs with variable neutralizing activities to mediate pH-dependent transcytosis with fully infectious HIV-1JRFL; 50% inhibitory concentrations (IC50s) of the antibodies ranged from 0.06 to >50 µg/ml (Fig. 3E). Both poorly neutralizing antibodies (HIVIG and mAbs b6 and F240; Figure 3A) and neutralizing mAbs (4E10, 2F5, 2G12, VRC01, and b12; Figure 3B) enhanced transcytosis at pH 6.0. At 50 µg/ml, transcytosis correlated directly with mAb binding to HIV-1JRFL (Spearman rho = 0.75; p = 0.052) and inversely with the IC50 of the mAbs (Spearman rho = −0.71; p = 0.050) (Figure S5). At pH 6.0, all Env-specific mAbs and HIVIG mediated transcytosis of virus that infected TZM-bl cells (Figure 3C and D). However, there was a strong correlation between the amount of transcytosed infectious virus and the neutralizing activity (IC50) of the antibody that mediated the transcytosis (Spearman's rho = 0.86; p = 0.001). Virus whose transcytosis was mediated by poorly neutralizing antibodies HIVIG, F240 and b6, at least at concentrations of 100 and 50 µg/ml, was more infectious than virus which crossed the epithelial cells in a non-FcRn-dependent manner (i.e., in the presence of Den3 control mAb) (Fig. 3C). Conversely, transcytosis mediated by antibodies with the lowest IC50s, such as VRC01 and b12, resulted in less infectious virus than was observed with the Den3 control antibody (Figure 3D). Thus, strong binding activity results in more FcRn-dependent transcytosis, whereas strong neutralizing activity renders the transcytosed virus less infectious. This point is further illustrated by the ratio of percent-transcytosed:percent-infectious virus (Figure 3E). For example, for every infectious unit, about 30 times more virus is transcytosed with VRC01 than with HIVIG (Figure 3E). Note that independently of transcytosis, HIV-1JRLF infectivity on TZM-bl cells increased about 3.5-fold after incubation of virus for 12 hours at pH 6.0 compared with pH 7.4; however, IC50s were very similar (<15% difference) at the two pH values (pH comparisons done for 2F5 and VRC01 only; data not shown). Virus infectivity was essentially abrogated after a 12-hour incubation at pH 4.0 (data not shown).
We next determined whether IgG purified from cervicovaginal fluid and from seminal fluid could enhance transcytosis at low pH. Using IgG from cervicovaginal fluid of three HIV-infected women and from seminal fluid of three infected men, enhanced transcytosis occurred at IgG concentrations well within their expected range in genital tract secretions (Figure 4A and 4B) [18]. The ability of genital tract IgG to mediate transcytosis correlated strongly with infectious virus capture activity by the IgG (Spearman's rho = 0.94, p = 0.005; Figure 4C) and less so with binding to monomeric Env glycoprotein from the same virus strain (HIV-1US657) (rho = 0.65, p = 0.16; Figure S6). None of the genital tract IgGs were able to neutralize HIVUS657, the clinical R5 strain used in these experiments, at IgG concentrations as high as 50 µg/ml (not shown). Consistent with HIVIG and the non-neutralizing mAbs, higher concentrations of genital tract IgGs generally resulted in greater infectivity of the transcytosed virus (Figure 4D and 4E).
FcRn expression was previously reported in human uterine and vaginal epithelial cells [19]. Using immunohistochemistry to survey FcRn protein expression at various sites in the human genital tract, we detected abundant FcRn expression in columnar epithelial cells lining the human penile urethra (Figure 5A and 5D; Figure S7A and S7D) and endocervix (Figure 5B and 5E; Figure S7B and S7E). In contrast, little to no FcRn protein was observed in vaginal/ectocervical squamous epithelia, and expression occurred only in the basal epithelial layer (Figure 5C and 5F; Figure S7C). A similar staining pattern was observed in foreskin tissue (data not shown).
Female genital tract secretions are often acidic, and the secretions of HIV-infected individuals have antibody capable of coating virus contained in those secretions. These facts led us to explore the role of antibody and low pH on transcytosis of HIV-1 across epithelial cells. Our primary finding is that at acidic pH, IgG enhances transcytosis of HIV-1 clinical isolates, including transmitted/founder Env-pseudotyped strains. Moreover, antibody from both cervicovaginal and seminal fluid mediates enhanced transcytosis at low pH. The enhanced transcytosis is abrogated by blocking or knocking down FcRn, which is known to bind IgG and immune complexes at low pH and release them at neutral pH [24], [42]. We also establish that virus translocated across epithelial cells after incubation with antibody at low pH remains infectious. Although neutralizing antibodies generally promote more transcytosis, the transcytosed virus is relatively less infectious than virus whose transcytosis is mediated by non-neutralizing antibodies. Finally, we demonstrate abundant FcRn protein expression in columnar epithelial cells of the human endocervix and penile urethra, suggesting that these sites could play a major role in FcRn-mediated immune complex transcytosis.
Our results indicate that FcRn may be responsible for shuttling IgG-bound HIV-1 across epithelial cells in the genital tract. This is consistent with other studies that have highlighted a role for FcRn in immune complex shuttling across tissues [34], [35], [43]. Mice expressing human FcRn in intestinal epithelial cells were able to deliver IgG to the luminal intestinal surface, which could then bind to its cognate antigen and return the immune complex back to the lamina propria for presentation by dendritic cells to CD4+ T cells [34]. In addition, cytomegalovirus (CMV) applied to human placental explants from women with high anti-CMV neutralizing antibody activity was rapidly transcytosed across syncytiotrophoblasts and captured by villus macrophages [35]. Under these conditions, the virus did not replicate. However, in explants from CMV-seropositive women with low or undetectable neutralizing antibodies, virus replication readily occurred in cytotrophoblasts underlying an intact, uninfected syncytiotrophoblast layer. Thus, it appeared that neutralizing antibody inhibited infection after allowing virus to cross the syncytiotrophoblast layer. On the other hand, non-neutralizing antibody allowed or even promoted infection. Syncytiotrophoblasts express high levels of FcRn, and when FcRn on explants was blocked, IgG-virion complexes were not transported across the surface [35]. Just as we found with HIV-1, FcRn-mediated transcytosis of CMV occurred with both neutralizing and poorly neutralizing antibody, but transcytosed virus remained infectious only when complexed with poorly neutralizing antibody. Finally, immunohistochemical staining of placentas from in utero infections were consistent with this model of FcRn-mediated transcytosis [35].
To our knowledge, ours is the first study to investigate transcytosis using virus coated with HIV-specific antibody in an acidic environment that mimics that of the female genital tract. Our in vitro observations are applicable to male-to-female transmission via vaginal intercourse, where enhanced transcytosis could facilitate infection. In this regard, Li et al. reported FcRn expression and bidirectional IgG transport in a human vaginal tissue model [19]. Although we did not detect FcRn in the apical layers of vaginal epithelium, we did detect abundant FcRn expression in columnar endocervical epithelial cells. These cells may be exposed to acidic vaginal secretions where they occur at the cervical os. Furthermore, cervical ectopy, a common condition characterized by the extension of endocervical columnar epithelium into the ectocervix and upper vagina, has been implicated as a risk factor for HIV-1 infection [44], [45]. Prevalent in reproductive-age women, these cervical lesions are exposed to vaginal pH conditions and could provide portals for FcRn-mediated male-to-female HIV-1 transmission [46]. FcRn was also found, though not consistently, in basal epithelial cells of the vagina. These cells lie deep in the epithelium and are unlikely to come in contact with acidic secretions and HIV-1 immune complexes unless there were trauma or substantial thinning of the overlying squamous epithelium. It is important to note that seminal fluid can rapidly raise the pH of cervicovaginal secretions to levels which would not support immune complex-FcRn binding. However, the pH of cervicovaginal fluid following ejaculation is dependent on the quantity of the ejaculate and may stay within an acidic range [29]. Furthermore, HIV is present in preejaculate secretions and could be introduced into the female genital tract prior to ejaculation [47].
With respect to female-to-male transmission, the penis comes in contact with vaginal secretions that would remain at acidic pH at least until ejaculation, allowing time for exposure of penile tissues, including the foreskin and urethra, to IgG-coated virus at low pH [28], [29]. Our demonstration of abundant FcRn on human penile urethral epithelium supports a model where exposure to antibody-bound HIV-1 might lead to enhanced female-to-male transmission. It should be noted that the pH of vaginal secretions is typically about 4, which is below the pH required for Fc-FcRn binding [48]. However, there is substantial variability in normal vaginal pH [26], [48], and we did begin to observe enhanced transcytosis at pH 4.5 (Figure S3H). Furthermore, it is possible that there is some buffering effect of foreskin and urethral secretions. The foreskin, whose presence increases HIV infection rate, could trap secretions containing HIV-1 immune complexes and thereby allow greater urethral exposure to infected material within the pH range of Fc-FcRn binding [49]. Additionally, bacterial vaginosis, a condition associated with an increased risk of female-to-male (as well as male-female) HIV transmission, results in vaginal secretions ideal for Fc-FcRn binding [26], [27], [50]. Exposure of penile tissues to the pH range of Fc-FcRn binding may also occur after ejaculation, since complete neutralization of vaginal acidity may not occur immediately or at all [29]. It is also possible, though less likely, that FcRn mediates HIV transmission via the penis during insertive anal intercourse, where the penis may come into contact with slightly acidic rectal secretions [30].
The finding that IgG from cervicovaginal and seminal fluids obtained from HIV-infected individuals mediate enhanced transcytosis of infectious virus further suggests the biological relevance of our results. Cervicovaginal and seminal fluids are reported to contain an average of ∼3 µg/ml and up to ∼15 µg/ml of Env-specific IgG [18]. Four of the six samples we evaluated bound to infectious HIV-1 at 5 µg/ml. Moreover, all of our samples mediated transcytosis at ≤12.5 µg/ml of total IgG, well below total IgG concentrations found in genital secretions of HIV-infected men and women [18]. Even during acute HIV infection, when the risk of transmission to an uninfected partner is highest, 23 of 23 subjects (100%) were reported to have anti-gp41 IgG antibodies and 40% had anti-gp120 IgG antibodies in cervicovaginal and seminal fluids [17]. Anti-gp41 IgG levels were on average 11-fold higher than gp41-specific IgA levels; anti-gp41 IgM was found less frequently and in lower quantity. Thus, HIV-1 immune complexes are likely to occur in mucosal secretions, are likely to contain predominantly IgG, and under acidic conditions, would be subject to FcRn-mediated transcytosis in an exposed host.
The relevance of our findings is also supported by our demonstration that transcytosis of transmitted/founder strains of HIV-1 Env pseudotyped virus is enhanced by antibody. We are currently evaluating whether transmitted/founder strains, in comparison with chronic strains, are preferentially transcytosed, which would be consistent with a report showing a higher sensitivity of clade B transmitted/founder strains to anti-Env antibody binding [51].
Our findings represent a new model of antibody-dependent enhancement (ADE) of HIV-1 infection. Previous studies have demonstrated ADE in vitro due to FcγR- or complement-mediated mechanisms or to modulation of the interaction of gp120 with CCR5 [52], [53], [54]. Here we demonstrate that enhancement in vitro occurs at the level of transcytosis across epithelial cells and involves FcRn. In vivo, Ig isotype, as well as neutralizing activity, are likely to play a determining role in whether an antibody might protect from or enhance infection. As demonstrated recently, intrarectally applied dIgA1 HGN194 mAb, but less so the IgG1 version, prevented SHIV infection following intrarectal challenge [40]. In vitro, the dIgA1 inhibited transcytosis, whereas we now show that the IgG1 version enhances transcytosis at pH 6.0. Another study showed that, compared to irrelevant- and no-antibody controls, there was an increase in the number of transmitted/founder SHIV variants when vaginal challenge followed systemic or local infusion of a non-neutralizing IgG1 mAb [55]. Clearly, other studies have found that IgG with neutralizing activity can prevent lentivirus infection after vaginal challenge [56], [57]. Thus, whereas a strong vaccine-induced neutralizing IgG response may protect, non-neutralizing IgG or waning titers of neutralizing IgG present in an acidic lumen might enhance transcytosis across mucosal barriers while allowing infection of susceptible target cells. However, whether an antibody protects, enhances or has no effect is likely to depend on the potency and breadth of antiviral activity, the viral strain, the inflammatory state of the exposed individual, and genetic factors—such as FcγR polymorphisms—that might influence antibody function [58]. Finally, if FcRn-mediated transcytosis applies in vivo, our results would strengthen the argument for a mucosal IgA response to vaccination—though not at the exclusion of a strong IgG neutralizing or other anti-viral response—since IgA can inhibit transcytosis, would not engage FcRn, and mediates only uni-directional translocation of immune complexes from the subepithelial space into external secretions [40], [59].
Some studies have reported that anti-HIV-1 Env IgG antibodies can inhibit transcytosis [9], [60], [61]. One of these studies found that polyclonal anti-HIV Env IgG inhibited transcytosis of cell-free virus on HEC-1 cells, whereas none of 13 mAbs did; in fact, some of the mAbs might have increased transcytosis, although by no more than about 50% [61]. To our knowledge, none of these studies was carried out under the acidic conditions that characterize female genital tract secretions.
Our results suggest that FcRn might facilitate infection in hosts without pre-existing antibody or with a non-neutralizing IgG response to prior infection (which would result in secondary infection) or to vaccination. However, FcRn could also play a beneficial role in preventing infection after exposure. FcRn mediates the bidirectional transcytosis of IgG, and in immunized individuals, could provide a conduit for antibodies to neutralize virus as shown for herpes simplex virus type 2 [19]. In addition, IgG immune complexes can prime CD4+ and CD8+ T cells in an FcRn-dependent manner, and FcRn targeting may be a useful mucosal immunization strategy [62], [63], [64].
In summary, we have demonstrated that FcRn mediates enhanced transcytosis of HIV-1 in the presence of low pH and HIV-1-specific antibody. We have also shown that FcRn is present on epithelial cells in areas of the genital tract that are potentially exposed to HIV-1 during sexual intercourse. Our findings point toward a novel mechanism by which the sexual transmission of HIV-1 may be facilitated.
This research was approved by the Institutional Review Boards at the University of California, Irvine, Boston University, and the University of Alabama, Birmingham. Subjects from whom specimens were collected for study purposes provided written informed consent.
Human Endometrial Carcinoma (HEC-1A) cells (ATCC) were propagated in Modified McCoy's 5a Medium, and Human Colon Carcinoma (T84) cells (ATCC) in Dulbecco's modified Eagle's medium; media were supplemented with 2.5 mM L-glutamine (Gibco, Invitrogen Technologies), 1% Penstrep (Cellgro Mediatech Inc.) and 10% FBS (Atlas Biologicals) and maintained at 37°C with 5% CO2. TZM-bl cells (NIH AIDS Reagent Program) for infectivity assays were propagated in RPMI 1640 supplemented with L-glutamine, Penstrep and 10% FBS as above. Five primary clinical HIV-1 strains, HIV-1US657, HIV-1US712, HIV-1JRFL, HIV-1HT593, and HIV-1HT599 were obtained from the NIH AIDS Reagent Program. SHIV1157ipEL-p, provided by Ruth Ruprecht, was grown in rhesus peripheral blood mononuclear cells [65].
HIVIG (IgG derived from pooled plasma of HIV-infected individuals) and IgG1 monoclonal antibodies (mAbs) 2F5, 4E10, 2G12, F240, b6, and VRC01 were obtained from the NIH AIDS Reagents Program. IVIG (Gamunex, Taleris Biotherapeutics) was commercially acquired. mAb b12 and control mAb Den3 were provided by Dennis Burton and Brian Moldt, and control mAb Fm-6 was a gift of Wayne Marasco (Dana-Farber Cancer Institute); b12 and the control mAbs are IgG1. Generation and purification of dimeric and monomeric IgA2 versions of b12 (dIgA2 b12 and mIgA2 b12) are described elsewhere [66]. Briefly, the IgG constant region in pDR.12 (IgG b12) was replaced with the constant region of IgA2. IgA2 b12 was expressed in CHO-K1 cells with human J chain and purified by Protein L affinity matrix (Pierce). mIgA b12 and dIgA b12 were isolated by size exclusion chromatography. IgG1 HGN194 (a human mAb against HIV-1 Env V3), dIgA1 HGN194, and dIgA2HGN194 were provided by Davide Corti and Antonio Lanzavecchia [67]. HGN194 variants were constructed as follows: human J chain precursor (accession number NP_653247), IgA1 (allele IGHA1*01, accession number J00220) and IgA2 (allele IGHA2*01, accession number J00221) constant region nucleotide sequences were codon optimized and synthesized by Genscript. Constant regions were cloned into a mammalian expression vector used for subcloning of the HGN194 VH region. The HGN194 VH and VL chain were codon optimized and synthesized by Genscript and cloned into an IgG1 and Ig-lambda expression vector. MAbs HGN194 dIgA1, dIgA2, and IgG1 were produced by transient transfection of 293 freestyle cells with polyethylenimine and expression plasmids encoding corresponding heavy and light chains (in the case of dIgA1 and dIgA2, the J chain expression plasmid was included). Supernatant fluid from transfected cells was collected after 7–10 days of culture. HGN194 dIgA1, dIgA2, and IgG1 were affinity purified by Peptide M (dIgA1 and dIgA2) or Protein A (IgG1) chromatography. Purified Abs were quantified by ELISA using dIgA1 and dIgA2 or IgG1-specific Abs (Southern Biotech). Purity and polymeric state of dIgA1 and dIgA2 were confirmed by native-PAGE analysis and gel filtration chromatography. The presence of dIgA1 and dIgA2 associated J-chain was confirmed by Western blot from native and SDS-PAGE gels. Sera from 20 Zambian clade C-infected subjects (obtained from Zdenek Hel, University of Alabama, Birmingham) were pooled for IgG isolation using the Pierce Melon Gel IgG Spin Purification Kit (Thermo Scientific) according to the manufacturer's instructions. Env-specific IgG, determined as for CVL and seminal fluid (see below), was 0.98% of total IgG. Sera from five uninfected individuals were pooled and processed for IgG isolation in the same manner.
Fc mutants designed to enhance (M428L) or reduce (I253A) mAb b12 binding to FcRn were constructed as follows: briefly, the b12 variable regions were PCR-amplified from pDR12 and cloned into the pγ1HC and pκLC vectors [68], [69]. Amino acid substitutions were introduced by QuikChange site-directed mutagenesis (Stratagene, La Jolla, CA). Constructs were verified by sequence analysis before transiently expressed in FreeStyle 293 cells (Invitrogen, Carlsbad, CA) and purified by protein A affinity chromatography (GE Healthcare, United Kingdom).
Antibodies were tested for neutralizing activity against indicated HIV-1 strains using TZM-bl cells.
Half-area 96-well plates (Corning) were coated with 5 µg/ml (250 ng/well) of goat anti-human Fc antibody and incubated over night at 4°C. Plates were then washed with PBS and blocked with 4% non-fat dry milk for 1 hour at room temperature (RT). After washing, capture antibodies were added at 5 µg/ml (250 ng/well), and plates were incubated an additional hour at RT. Next, virus was added to washed plates (20 ng p24/well) and incubated for 3 hours at 37°C. Unbound virus was removed by washing with PBS. Subsequently, 1×104 TZM-bl cells/well were added in the presence of 10 µg/ml DEAE dextran and incubated for 48 hours at 37°C. Cells were then washed, lysed, and developed with luciferase assay reagent according to the manufacturer's instructions (Promega). Luminescence (relative light units) was measured using a Synergy 2 microplate luminometer (BioTek).
We measured binding of antibodies either to virus directly coated on ELISA-plate wells or to solubilized Env. For the direct virus binding assay, plates were coated with HIV-1JR-FL (20 ng p24/well) for 2 hours at 37°C, washed with PBS and blocked with 4% non-fat dry milk in PBS. After 1 hour at 37°C, plates were washed, antibodies were added in serial dilutions and incubated for 1 hour at 37°C. Detector antibody (horse radish peroxidase-labeled goat anti-human Fc) was added to the washed plate and incubated for 45 min at 37°C. Finally, plates were washed, developed (TMB solution, Life Technologies), and read at 450 nm using a plate reader (BioTek).
The soluble Env binding assay was performed as previously described with some modifications [70]. Briefly, wells were coated with 250 ng of a gp120 Env specific anti-C5-antibody (D7324 [Aalto Bioscience]), washed and blocked with 4% non-fat dry milk. Serial dilutions of detergent-solubilized HIV-1JR-FL (starting at 150 ng p24) was added and incubated for 2 hours at 37°C. Plates were then incubated with a constant concentration of antibodies (1 µg/mL) for 1 hour at 37°C followed by detection and development steps as described above.
Five R5 clade C transmitted/founder Env pseudotyped strains were constructed as described [71], [72]. Briefly, rev-vpu-env cassettes from the transmitted founder strains were cloned into pcDNA 3.1D/V5-HIS TOPO® expression vector. The pseudotyped viruses were then produced by co-transfecting 293T cells with pcDNA 3.1(rev-vpu-env), pNL4-3.lucR-E-, and fugene 6 (Roche).
Cervicovaginal lavage (CVL) and seminal fluid were collected from HIV-1-infected patients and healthy volunteers at the University of Alabama, Birmingham. All subjects gave written consent in accordance with an IRB-approved protocol. CVL was collected from one 34 year-old uninfected women and from three infected women (age 29 to 46 years) with CD4+ lymphocyte counts of 458/mm3, 181/mm3 and 498/mm3 and plasma viral loads of 14100 copies/ml, 824 copies/ml and 88 copies/ml, respectively. Viral loads were not measured in the CVL fluid specimens. Two of the women (with the lower plasma viral loads) were receiving anti-retroviral therapy. Briefly, CVL fluid was obtained by flushing the cervix and vagina with 5 ml sterile saline, and the wash was collected into tubes with protease inhibitors ([73]). Seminal fluid was obtained from two uninfected men (ages 25 and 40 years), and from three infected men (age 43 to 53 years) by masturbation (58). CD4 counts in the infected men were 404/mm3, 336/mm3 and 407/mm3 and plasma viral loads were <100 copies/ml, 8092 copies/ml and 6750 copies/ml, respectively; only one of these subjects (with viral load of 11 copies/ml) was receiving anti-retroviral therapy. Seminal fluid was assayed for HIV-1 RNA by PCR, but none was detected. The cervicovaginal and seminal fluids were centrifuged and supernatant fluids aliquoted and frozen at −80°C until assayed. Total IgG was determined by ELISA [74]. IgG isolation was accomplished by incubating samples with Protein G-Sepharose (GE Healthcare Bio-Sciences Corp.) followed by elution of bound IgG according to manufacturer's instructions. The IgG preparations were concentrated and dialyzed against DPBS using Amicon Centrifugal Filter Units (Millipore Corp.). The IgG preparations from the two uninfected men were pooled to obtain sufficient quantity for experiments; all other IgG preparations were tested individually.
Env-specific IgG binding levels in seminal and CVL fluids were quantified by ELISA. Detergent-solubilized Env from HIV-1US657 was captured by a polyclonal sheep anti-gp120 antibody (D7324, Aalto Bio Reagents Ltd). Sample IgG and an anti-gp120 mAb standard (b6) were serially diluted, added to wells, washed, and detected by anti-human IgG (gamma)-HRP (Sigma-Aldrich, A6029) antibody. Plates were subsequently developed, stopped, and read at OD450 nm. The concentrations of Env-specific IgG in the seminal and CVL IgG samples were calculated using the mAb b6 standard and are reported as a percent of total IgG in each sample.
Anti-Env IgG ranged from 0.9 to 2.6% of total IgG in the seminal fluid specimens and from 0.1 to 0.6% of total IgG in the CVL fluids specimens (Figure S6).
Transcytosis assays were conducted using reproductive tract-derived (human endometrial carcinoma [HEC-1A]) or intestinal tract-derived (human colonic carcinoma [T84]) cells. HEC1-1A or T84 cell monolayers were created on 0.4 µm polyethylene terephthalate membrane hanging transwell inserts (Millipore). Cell viability was >95% at the time of plating. Electrical resistance across the membrane, which ranged from 400–450 mOhms/cm2 at the start of the transcytosis assay, confirmed monolayer integrity. Resistance was re-measured after the transcytosis assay in more than 50% of wells and ranged from 450–480 mOhms/cm2. HIV-1 alone or with antibody was added to monolayers in media buffered to pH 6.0 or 7.4. After 12 hours, fluid in the lower chamber (“subnatant fluid”), maintained at pH 7.4, was collected and used to measure viral RNA copy number and infectivity. In the absence of cell monolayers, about 69% of the virus inoculum was present in the lower chamber of the wells after 12 hours.
Viral RNA was extracted from cell-free subnatant fluid using PureLink Viral RNA Mini Kits (Invitrogen) or NucleoSpin RNA Virus extraction kits (Macherey Nagel Inc.), according to the manufacturers' instructions. Quantitative one-step real-time RT-PCR of extracted HIV-1 viral RNA was done using Quantitect SYBR Green RT-PCR kits (Qiagen GmbH) and that of SHIV1157ipEL-p with Rotor Gene Probe RT-PCR kits according to the manufacturers' instructions. HIV-1gag primers: SK462 d(AGTTGGAGGA-CATCAAGCAGCCATGCAAAT) and SK431 d(TGCTATGTCAGTTCCCCTTGGTTCTCT) (AnaSpec Inc.). SIV-1 gag primers: d(GGG AGA TGG GCG TGA GAA A) and d(CGT TGG GTC GTA GCC TAA TTT T). SIV-1 gag probe: d(TCA TCT GCT TTC TTC CCT GAC AAG ACG GA) (Integrated DNA Technologies, Inc.).
150 µl of subnatant fluid was used to infect 1×104 TZM-bl cells. TZM-bl cells were lysed 2 days post-infection with 1X Cell Culture Lysis Reagent (Promega Corp.), and luciferase activity was determined by chemiluminescence using Luciferase Substrate (Promega Corp.).
HEC-1A cells were transduced with FcRn shRNA Lentiviral Particles (Santa Cruz Biotechnology Inc.) following manufacturer's protocol. Cells were selected in medium containing 5 µg/ml Puromycin dihydrochloride (Sigma-Aldrich Inc.), and FcRn expression was verified by flow cytometry using rabbit polyclonal anti-FcRn antibody (Santa Cruz Biotechnology Inc.), normal rabbit IgG (negative control) and FITC-goat anti-rabbit IgG F(ab′)2 secondary antibody (Jackson ImmunoResearch Laboratories Inc.) (Figure S3A). Cytofix/Cytoperm Plus Kits (BD Biosciences) were used to fix, permeabilize and stain cells. Knockdown of FcRn was also confirmed by western blot using rabbit anti-FcRn antibody (Novus Biologicals) (Figure S3A). Wild-type and knockdown cells had similar viability. Neither wild-type nor knockdown HEC-1A cells stained for FcγRIIa or FcγRIIIa (not shown).
HEC-1A cells were incubated with 50 µg/ml rabbit polyclonal anti-FcRn IgG (H-274; Santa Cruz Biotechnology Inc.) or normal rabbit polyclonal IgG for 1 hour at pH 7.4 before exposing the apical surface to HIV-1US712 and HIVIG, b12, IVIG or Synagis. Similarly, HEC-1A cells were incubated with 0.1 µM bafilomycin A1 (Santa Cruz Biotechnology Inc.) for 1 hour prior to HIV-1 and antibody exposure. Transcytosis was then carried out as above.
Cervical tissue, which included portions of endocervix and upper vagina (ectocervix), was obtained from 10 women aged 31–50 undergoing hysterectomy for nonmalignant conditions. Vaginal tissue was also obtained from women undergoing vaginal repair (n = 6, aged 44–78 years). Penile tissue, including urethra (n = 16) and foreskin (n = 2), was harvested at autopsy from 16 men aged 34–73 with no history of hormonal or immunosuppressive medications. Tissues were processed within 60 minutes of surgical removal. Samples were either embedded in Tissue-Tek Optimal Cutting Temperature Compound (Sakura Finetek U.S.A., Inc.) and rapidly frozen and stored at −70°C (frozen sections) or were fixed in formaldehyde and processed for paraffin embedding.
The alkaline phosphatase immunohistology technique was described previously ([75]). Two anti-FcRn antibodies were used: 1) Anti-FcRn antibody purified from rabbit serum raised against α2 (88–177aa) and α3 (1782-247aa) domains of human FcRn (provided by Neil Simister, Brandeis University) for use on frozen sections (Figure 5), and 2) rabbit anti-FcRn antibody obtained from Novus Biologicals for use on paraffin sections following citrate buffer (pH 6.0) antigen retrieval (Figure S7). Sections were blocked with serum-free protein solution, and optimally diluted primary FcRn antibodies or rabbit IgG (negative control) were added and incubated for 1–2 hours at RT. Binding of antibodies to FcRn in tissues was visualized using an alkaline phosphatase detection system that stains positive cells bright red. Sections were counterstained with hematoxylin and cover-slipped using aqueous mounting medium.
Differences in amounts of transcytosed or infectious virus between conditions were compared using Kruskal-Wallis or repeated-measures ANCOVA. For repeated-measures ANCOVA, the percentage of transcytosed or infectious virus was logit-transformed and normality evaluated using the Shapiro–Wilk test. Correlations between continuous variables were evaluated by Spearman's rho. Two-tailed p-values are reported.
|
10.1371/journal.pcbi.1005310 | Customized Regulation of Diverse Stress Response Genes by the Multiple Antibiotic Resistance Activator MarA | Stress response networks frequently have a single upstream regulator that controls many downstream genes. However, the downstream targets are often diverse, therefore it remains unclear how their expression is specialized when under the command of a common regulator. To address this, we focused on a stress response network where the multiple antibiotic resistance activator MarA from Escherichia coli regulates diverse targets ranging from small RNAs to efflux pumps. Using single-cell experiments and computational modeling, we showed that each downstream gene studied has distinct activation, noise, and information transmission properties. Critically, our results demonstrate that understanding biological context is essential; we found examples where strong activation only occurs outside physiologically relevant ranges of MarA and others where noise is high at wild type MarA levels and decreases as MarA reaches its physiological limit. These results demonstrate how a single regulatory protein can maintain specificity while orchestrating the response of many downstream genes.
| Bacteria can sense and respond to stress in their environment. This process is often coordinated by a master regulator that turns on or off many downstream genes, allowing the cell to survive the stress. However, individual genes encode products that are diverse and optimal expression for each gene may differ. Here, we focus on how expression of diverse downstream genes is optimized by targets of the multiple antibiotic resistance activator MarA. Using single-cell experiments and computational modeling we show that downstream genes process MarA signals differently, with unique activation, noise, and information transmission properties. We find that each downstream gene’s response depends critically on the level of the input MarA. Furthermore, by swapping parts of the regulatory elements of genes we were able to create novel responses. This suggests that these properties can be readily tuned by evolution. Our findings show how a network with diverse downstream genes can be used to process the same command to achieve many distinct outputs, which work together to coordinate the response to stress.
| Genetic networks often feature master regulators that control suites of downstream genes. This type of architecture is particularly common in stress response; examples include Msn2 and Crz1 in Saccharomyces cerevisiae, σB in Bacillus subtillis, and the multiple antibiotic resistance activator MarA in Escherichia coli [1–4]. Each of these regulators controls tens to hundreds of diverse downstream targets with widely varying functional roles. For example, MarA regulates small RNAs, metabolic enzymes, efflux pumps, and regulatory proteins [1,5]. This diversity of gene products raises two questions: First, how is one signal from an upstream regulator decoded differently by multiple downstream targets? Second, what are the potential benefits and tradeoffs that define how these genes respond? To address this, we examined three key properties: activation, transmitted noise, and transmitted information. Although the properties are linked, they can be adjusted in a variety of ways to tailor the response of individual genes.
Activation is set by molecular details of the transcription factor and the DNA to which it binds. The resulting transfer function is often described by the sigmoidal Hill function [6]. Here we focus on MarA, which regulates over 60 downstream targets by binding to a well-characterized degenerate binding sequence in their promoters known as the ‘marbox’ [5,7]. Previous studies have mapped the activation properties of genes controlled by MarA, finding that the amount of MarA needed to turn on expression varies greatly among its targets, with a 19-fold difference in the dissociation constant (Kd) between genes [1]. The majority of downstream genes are only weakly activated unless MarA is overexpressed far beyond physiologically relevant levels [1]. This is due to the high dissociation constants of the downstream promoters [8], and provokes the question of why these genes are regulated by MarA at all.
Noise provides insight into why this may be the case. Recent studies at the single-cell level suggest that although most cells only express low levels of MarA, levels are high in a small subset of the population due to cell-to-cell differences in gene expression [9,10]. These single-cell differences allow a subpopulation of cells to survive antibiotic exposure, and survivors can recolonize after the stress has passed [9]. Therefore, noise in MarA may turn on expression of costly downstream genes only in a small subset of the population to hedge against future uncertainty. This role for noise in stress response is observed frequently, such as in sporulation and competence in B. subtilis, which allow subpopulations of cells to survive periods of extreme stress [11–13]. When a noisy input controls downstream targets that have high dissociation constants, low-level fluctuations can be filtered, while larger signals are transmitted [14,15].
The relationship between the level of the input and the ability to activate downstream genes defines whether information is transmitted between input and output. This can be quantified using the metric channel capacity, which is the theoretical limit of how well information can be passed through a network (similar to the diameter of a pipe, or channel) [16,17]. Channel capacity depends both on the bounds of possible input values for the natural system and on the shape of the activation curve [17–19]. For instance, if the realistic range of MarA values are constrained such that a downstream gene is only weakly activated, then the channel capacity is low. As the range of MarA values widens, the channel capacity may either increase, if there is a corresponding increase in downstream gene expression, or remain low, if expression does not change, as in the case where the response is saturated. Therefore the ability to transmit information depends on both the input distribution and where these values fall on the activation curve.
Expression of individual downstream genes can be tailored to balance activation, noise, and information transmission in any number of ways. For instance, certain genetic regulatory elements transmit information near the channel capacity [20,21], while others lose considerable information due to noise or active filtering [22,23]. Therefore, gene regulation may be tailored to balance these properties.
Using MarA and its downstream genes as a case study in multi-gene regulation, we investigated how expression is tailored at the single-cell level. We identified two qualitative classes of promoters: amplifying and filtering. The amplifying promoters generate diversity and have a high channel capacity at low levels of MarA. In contrast, the filtering promoters have low noise and only transmit information when MarA is high. These qualitative differences in promoter classes correlate with the functional differences in the gene products they control. We created chimeric promoters by swapping the MarA binding sequences between downstream promoters from these groups and easily altered their quantitative characteristics. Therefore, this binding sequence can serve as an evolutionary target for tuning output response.
We first asked how individual genes respond to MarA at the single-cell level. To do this, we transformed E. coli MG1655 ΔmarRAB with a plasmid bearing an IPTG-inducible version of marA transcriptionally-fused to red fluorescent protein (rfp). We cotransformed cells containing the inducible marA-rfp plasmid with a second plasmid containing green fluorescent protein (gfp) linked to a promoter for a MarA-controlled downstream gene. We then simultaneously measured RFP and GFP levels in individual cells using microscopy. By adding IPTG, we increased MarA (measured by RFP), which activated expression of the downstream promoter (measured by GFP). IPTG induction allowed us to capture a broad range of MarA levels (S1 Fig). These data provide single-cell resolution measurements of downstream gene expression as a function of MarA.
Initially, we quantified expression from the promoter PmicF in response to MarA. micF encodes a small RNA that represses the outer membrane porin OmpF, decreasing vulnerability to a number of stressors including antibiotics and osmotic shock [24]. Previous population-level analysis of PmicF has indicated that the promoter has a relatively low Kd, suggesting that it should turn on with low levels of MarA [1]. Consistent with this, PmicF expression increased in response to MarA induction and eventually saturated (Fig 1A). We eliminated the possibility that this result was due to artificial crosstalk between the RFP and GFP channels by constructing a control strain where the PmicF reporter was cotransformed with a plasmid with inducible rfp and no marA and observed no spurious crosstalk effects (S2 Fig).
Consistent with previous work on noise propagation [25], we noted that noise in GFP expression changed along the PmicF activation curve. Diversity in single-cell expression is highest at low levels of MarA. To quantify this effect, we calculated the normalized coefficient of variation as a measure of transmitted noise (Fig 1B). Analytically, this value is proportional to the local slope of the activation curve [19]. In other words, transmitted noise is highest where the Hill function is the steepest.
To investigate where on this curve physiological levels of MarA fall, we conducted experiments using the PmicF reporter in two genetic backgrounds. First, we measured the lower bound of the PmicF response using wild type E. coli MG1655. These represent the natural levels of PmicF under unstressed conditions with basal MarA expression. Second, we measured the upper bound of the PmicF response by using a strain we denote MarA+, where the chromosomal copies of both MarR binding sites are inactivated. The marRAB operon is induced when repression by MarR is inhibited, therefore by inactivating these repressor binding sites this strain expresses the maximum physiologically realistic level of MarA. Measurements from these two strains have distinctly different levels of PmicF reporter expression (Fig 1C).
To confirm that these bounds on the physiological levels of MarA were appropriate we also subjected wild type cells to several chemical stresses. We first used salicylate, the canonical inducer of the marRAB operon, which causes a conformational change in MarR that prevents it from repressing marRAB expression [26]. Using 1 and 3 mM salicylate, we found expression to fall within the upper and lower bounds established by the wild type and MarA+ strains. Additionally, the quinolone ciprofloxacin can indirectly inhibit MarR by increasing intracellular copper levels [26]. Exposure to sublethal levels of ciprofloxacin (2 and 4 μg/L) also resulted in intermediate levels of PmicF expression. Together, these results outline the biologically relevant range of PmicF expression.
Although PmicF expression changes dramatically as MarA sweeps across physiologically relevant values, bulk measurements of other MarA-activated genes have suggested that many MarA-regulated genes are not strongly activated [1]. Thus, we next tested expression of PinaA, which has a Kd ten times higher than PmicF [1]. While the exact role of inaA is unknown, it is a pH-inducible gene involved in stress response [27]. In sharp contrast to PmicF, PinaA shows a gradual response to MarA, which never saturates over the tested range (Fig 1D). We quantified transmitted noise and our results show good agreement with the theoretical prediction based on the slope of the fitted Hill function (Fig 1E). In contrast to PmicF, PinaA noise levels remain low across all MarA values. Paralleling the experiments with PmicF, we used the wild type and MarA+ genetic backgrounds to establish the physiologically relevant upper and lower bounds for PinaA expression (Fig 1F). The two distributions are noticeably closer together than in the case of PmicF, which is expected given the gradual slope of the Hill function.
Our initial results with PmicF and PinaA reveal that even with a common upstream regulator, there can be categorical differences in downstream gene expression. These differences include population-level response characteristics such as the shape of the activation curve and also single-cell level effects such as variability in gene expression. These effects are related since transmitted noise depends on the slope of the activation curve.
To explore how different characteristics of downstream gene activation influence transmitted noise, we used a computational model to simulate systems with different values for the dissociation constant (Kd) and Hill coefficient (n). We simulated a noisy activator regulating five downstream genes with different Kd values (Fig 2A). In addition, we included a control not regulated by the activator. We then calculated the transmitted noise for each, and the mean values of our stochastic simulations show excellent agreement with our analytical solutions (Fig 2B).
Given the sigmoidal shape of the Hill function, Kd alone can determine whether a promoter filters or amplifies a given input signal. Moreover, Kd is related to the strength of binding between the transcription factor and its associated binding site. This parameter is easily altered by mutations in the binding site and is therefore a potential evolutionary tuning knob. Indeed, the marbox has substantial variation among the myriad of genes regulated by MarA [8].
In contrast to Kd, altering the Hill coefficient n primarily affects the magnitude and shape of the noise response (Fig 2C and 2D). Genes with high n values have high transmitted noise over narrow input ranges, while lower n values correspond to lower, broader responses. We note that other parameters, such as activation and degradation rates, can also influence transmitted noise (Supplementary Information and S3 Fig). In general terms, Kd controls the activator levels where the transmitted noise is highest, while n primarily affects the magnitude of the transmitted noise.
We next asked whether the activation and transmitted noise profiles of a diverse set of MarA-regulated genes varied as a function of n and Kd as in our computational simulation. We expanded our single-cell studies to include a total of six MarA-regulated promoters: PmicF and PinaA discussed previously, and the promoters for efflux pump genes PacrAB and PtolC, superoxide dismutase PsodA, and PmarRAB. We selected these genes based on their diverse responses to MarA at the population level [1]. For each, we used IPTG-inducible MarA and measured activation and transmitted noise in the downstream promoters (Fig 2E and 2F). Of the six genes we measured, we observed a range of expression profiles that fall broadly into two groups. First, the ‘amplifying’ group, which includes PmicF and PmarRAB, saturates over the examined range of MarA inputs. These promoters have lower Kd values and larger n values. For these genes, low-level fluctuations in MarA will become large fluctuations in the downstream gene. In contrast, the ‘filtering’ group includes PacrAB, PinaA, PsodA, and PtolC. These genes do not saturate over the MarA range we tested and have lower n values. In this group, low-level fluctuations in MarA are filtered in the downstream gene, attenuating both the signal and the noise. These two classes of genes have categorically different transmitted noise profiles (Fig 2F). The amplifying genes have high transmitted noise peaks at low levels of MarA and drop sharply as MarA increases, while the filtering genes have low, broad transmitted noise curves.
While the previous results illustrate the differences in activation and transmitted noise, these findings need to be placed in the context of the biologically relevant levels of MarA. To investigate this we quantified the channel capacity of each promoter. Channel capacity is defined as the maximum mutual information—the potential of the input to inform the output [28]. The channel capacity reflects the ideal distribution of inputs through a target channel for maximizing mutual information [17]. To keep our calculations biologically grounded, we constrained the possible inputs by using estimated endogenous MarA levels. This is critical to our analysis because the physiologically relevant ranges of MarA are limited, and in some cases only span a narrow section of the downstream gene’s activation curve (Fig 1D and 1F). In order to quantify channel capacity for a given promoter, we divided our analysis into two cases that correspond to wild type and MarA+ levels. By mapping the distribution of GFP values that corresponds to wild type and MarA+ through our inducible system, we were able to estimate the MarA distribution that produced each downstream response (S4 Fig). Together, these two distributions represent physiologically relevant estimates of unstressed and stressed MarA levels (Fig 3A and 3B).
To provide intuition into the channel capacity of a promoter, we considered how MarA levels that correspond to wild type and MarA+ cells affect genes with amplifying versus filtering properties. At low levels of MarA, amplifying genes like PmicF transmit information well, proportionally mapping input to output. In contrast, the filtering genes like PinaA map the same input to a narrow band of output (Fig 3C). The difference in width between the input and output distributions in the filtering gene corresponds to information loss. In contrast, under MarA+ conditions the channel capacity of the filtering gene PinaA increases, while the amplifying gene PmicF has a lower channel capacity since the promoter saturates and high MarA values all map to the same output (Fig 3D).
We asked how the channel capacity varied for the six MarA-regulated genes under wild type and MarA+ conditions. We calculated the channel capacity as a function of Kd and n for the two genetic backgrounds (Fig 3E and 3F). Using parameters derived from experimental data, we plotted the location of each of the downstream genes on the channel capacity heat map. We found that for wild type MarA levels, downstream genes within the filtering class display lower channel capacity than those in the amplifying class (Fig 3E). As the input increases to MarA+ levels, we observed a shift and the filtering genes increased channel capacity, while amplifying genes decreased (Fig 3F).
Our calculations for channel capacity quantify what the maximum mutual information is for a bounded range of MarA inputs. Calculating the actual mutual information requires precise knowledge of the MarA input distributions. To estimate this, we calculated mutual information between the MarA distributions from the wild type and MarA+ strains and the downstream gene expression distributions produced by these inputs (S5 Fig). Despite the low channel capacity under wild type conditions for many downstream genes, the MarA input distribution is optimized to transmit information at near channel capacity for the filtering genes. However, these data also suggest that the wild type MarA distribution may not be taking full advantage of the amplifying promoters PmicF and PmarRAB, though we note that the results are very sensitive to the input distributions (S6 Fig), which are produced here as estimates.
The amplifying or filtering characteristics of a promoter are determined by how MarA binds. Therefore, it is likely that altering the MarA binding sequence would have a profound effect on the activation profile of the promoter it regulates, and in turn, its transmitted noise and channel capacity. Moreover, although we observed two general classes of promoters, it may be possible to generate promoters with intermediate properties.
To investigate this, we created chimeric promoters by swapping the marbox sequences between PmarRAB and PacrAB. We constructed two chimeric promoters, which we denote Pam and Pma. In the Pam chimera we started with PacrAB and replaced its marbox with that from PmarRAB; in the Pma chimera PmarRAB has the PacrAB marbox. We quantified the activation and transmitted noise of these chimeric promoters as before (Fig 4A and 4B). We found that both chimeric promoters have Kd and n parameter values that fall between the two natural promoters, and the corresponding channel capacity is also intermediate as a result (Fig 4C and 4D). This shows that activation, and the transmitted noise and information properties that depend on it, are readily tunable through marbox mutations. This sequence could serve as an ideal target for evolutionary adaptation.
Genetic networks where one master regulator controls multiple downstream genes can efficiently respond to stress by customizing the response of individual genes based on their diverse functions. Using the multiple antibiotic resistance activator MarA as a case study, we show that its downstream targets are individually tailored in the way they respond to MarA and how they transmit noise and information.
Understanding the physiological context of MarA proved to be critical; for instance, we found downstream genes that amplified signals under only wild type levels of MarA (PmicF and PmarRAB) and also examples that only show a response under high, non-physiological MarA conditions (PacrAB, PinaA, PsodA, and PtolC). These results argue that studies of stress response genes should be coupled with a concrete understanding of the appropriate cellular context. In our experiments with MarA we determined that there were two qualitative classes of downstream genes, which serve to increase variability or transmit critical signals. Ultimately, a cell’s ability to survive stress depends upon expression of multiple downstream genes in a coordinate fashion. The flexibility of MarA responses could balance multiple demands on a particular gene’s expression including cost, desired expression level, and noise. Further, the dissociation constant Kd and Hill coefficient n, are critical in setting a gene’s response. Our work with the chimeric promoters illustrates that these parameters are readily changed by altering the sequence of the marbox. We note that there are two MarA homologs, SoxS and Rob, that may play additional regulatory roles, further underscoring the need for context-dependent measurements [29].
While our experiments focused on activation, the underlying analysis can be extended to other classes of regulation. For instance, transmitted noise is proportional to the local slope of repressors just as it is in activators. As such, this study serves as a framework for contrasting how one gene controls many in the context of noise and information, unconstrained by the method of regulation. In the future it will also be interesting to examine the role of feedback, as previous research has shown that positive feedback has the potential to increase signal transmission without transmitting the associated noise [30].
Throughout our experiments, we show that each of the downstream genes behave differently in wild type and MarA+ strains. These two conditions correspond to approximations for stressed and unstressed states. Both PmarRAB and PmicF control regulatory molecules and therefore these amplifying genes may serve to increase diversity in unstressed conditions. As cells shift to stressed conditions, expression of these genes saturates. This could signify the transition from a bet-hedging state, where diversity is favored, to a state where all cells consistently express the target gene products. In contrast, the filtering genes only engage under high levels of MarA. Efflux pumps and other gene products controlled by these promoters are often costly to the cell and should only be expressed when needed [38].
We have demonstrated the flexibility of multi-gene regulation in stress response networks through a collection of single-cell experiments, computational simulations, and analytical analysis. Our results show how multiple downstream genes can display customized expression given the same input. Straightforward changes in promoter sequence can be used to change activation, noise, and information transmission properties, allowing for a diverse set of possible outcomes that can be tailored to optimize expression of specific gene products. Our findings on the plasticity and specificity of the MarA network provide insight into the role that master regulators can play in diverse stress response networks.
All plasmids were derived from the BioBrick library described in [31]. In order to construct the downstream gene reporter plasmids (denoted PacrAB, PinaA, PmarRAB, PmicF, PsodA, and PtolC) we placed the promoter from each upstream of super-folder green fluorescent protein (gfp) (AddGene #63176). These transcriptional reporters were constructed on a plasmid containing the kanamycin resistance marker and low-copy SC101 origin of replication (from [31]).
For the activator/reporter experiments, we cotransformed the GFP reporter plasmid with a second plasmid containing the ampicillin resistance marker and medium-copy p15A origin of replication (pBbA5k from [31]). This plasmid places either marA-rfp or rfp under the control of the IPTG-inducible lacUV5 promoter. marA-rfp is a transcriptional fusion of marA and rfp.
We used three strains for experiments: wild type E. coli MG1655, and genetic variants ΔmarRAB and MarA+. The ΔmarRAB strain is described in [9]. In MarA+, we used transversion mutations that annihilate the MarR binding sites in the chromosomal copy of the marRAB promoter, preventing repression of the operon.
We constructed the chimeric reporter plasmids using PacrAB and PmarRAB with marbox sequences from [7].
Further details on plasmids and strains is provided in Supplementary Information.
Cultures were inoculated from single colonies and grown overnight at 37°C with 200 rpm shaking in LB medium with 30 μg/ml kanamycin (reporter-only) or 30 μg/ml kanamycin and 100 μg/ml carbenicillin (activator/reporter). Overnight cultures were diluted 1:100 in selective LB. For the activator/reporter experiments, we added 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, or 500 μM IPTG and grew cultures for four hours (Supplementary Information and S1 Fig). For the reporter-only experiments, we grew cultures for two hours before adding either salicylate (1 or 3 mM) or ciprofloxacin (2 or 4 μg/L), then grew them for an additional two hours. Wild type and MarA+ reporter-only strains were grown for four hours without the addition of inducers.
For microscopy images, we placed cells on 1.5% MGC low melting temperature agarose pads [32]. We used a Nikon Instruments Ti-E microscope to image the cells at 100× magnification. Three images were taken of each pad to ensure that at least 100 cells were imaged under each growth condition. Custom MATLAB scripts were used to extract fluorescence data from individual cells. For the activator/reporter experiments, fluorescence values for cells from all IPTG levels were combined and then binned according to their RFP levels (S1 Fig).
To quantify activation of each downstream gene, we used Hill functions:
B=αβ((AKd)n1+(AKd)n)+cβ
(1)
where A is MarA and B is the downstream gene product. α is the promoter strength, Kd is the dissociation constant, n is the Hill coefficient, c is the basal expression level, and β is the degradation and dilution rate. The functions for calculating the normalized level of downstream gene product (used in Fig 2) subtract the basal expression and normalize (Supplementary Information).
We calculated transmitted noise using two independent methods in this study. First, transmitted noise is calculated from experimental data as the ratio of the noise in the downstream gene output (B) over the noise in the MarA input (A) [30]. We note that this quantity is sometimes referred to as ‘noise amplification’.
Noise in an individual gene is calculated as the coefficient of variation, which is the standard deviation (σ) divided by the mean (μ) for a given gene product level [30,33]. However, we needed to account for noise sources not coming directly from fluctuations in MarA, such as those from intrinsic and extrinsic noise [34]. The term S is equal to the noise of the downstream gene without regulation by MarA, and is necessary to fit the analytical solution to the experimentally measured transmitted noise. S does not vary as a function of MarA and is the sum of intrinsic and extrinsic noise sources that are independent of upstream gene regulation. Statistics were determined by bootstrap resampling of one third of the population 100 times.
Analytically, transmitted noise is equal to the local slope of the activation curve, normalized by the values of the function about that point [19]. Mathematically, this is equal to the slope of the logarithmic transform of the Hill function [30,35].
We used this function to calculate the analytic solutions in all noise plots.
Because the transmitted noise is equal to the local slope of the activation curve, Hill functions and transmitted noise curves share the same parameters [19]. We simultaneously fit both curves to experimental data using a differential evolution algorithm with a custom fitness function [36]. For fitness function and exact values of fitted parameters, see Supplementary Information.
We modeled the input A and the downstream products B by:
A˙=αA−βA+IA
(4)
B˙=α(AKd)n1+(AKd)n−βB
(5)
where α is the promoter strength, Kd is the dissociation constant, and n is the Hill coefficient. In addition, αA is the production rate of A and β describes protein degradation and dilution for both A and B. We simulated intrinsic noise of the input protein IA using an Ornstein-Uhlenbeck process [37]. The intrinsic noise of the input has a standard deviation of α. The correlation time of this noise calculated as Tint/ln(2), with Tint of 5 minutes [32]. As with the experimental data, the level of input protein (αA) was varied through a range of possible inputs (30 log spaced values). The parameters used and details of the stochastic simulation are given in Supplementary Information.
The channel capacity (I*) is dependent on the relationship between input and output and is calculated using the functions from [17,19,20,28]:
I*(A;B)=log2(Z)+X
(6)
Z=∫AminAmax[(dB/dA)2B+A0*A*(dB/dA)2]12dA
(7)
X is a constant that is independent of the parameters of the downstream promoters. It is introduced by the small noise approximation implicit in this calculation of channel capacity. Amin and Amax describe the minimum and maximum input values, which we determine from experimental data by mapping the output from wild type and MarA+ to the data from the activator/report experiments (S4 Fig) using the 5th to 95th percentiles from these distributions. A0 is a scaling term for the concentration of the activator to match experimental results. For further details see Supplementary Information.
|
10.1371/journal.ppat.1002104 | Hemoglobin Promotes Staphylococcus aureus Nasal Colonization | Staphylococcus aureus nasal colonization is an important risk factor for community and nosocomial infection. Despite the importance of S. aureus to human health, molecular mechanisms and host factors influencing nasal colonization are not well understood. To identify host factors contributing to nasal colonization, we collected human nasal secretions and analyzed their ability to promote S. aureus surface colonization. Some individuals produced secretions possessing the ability to significantly promote S. aureus surface colonization. Nasal secretions pretreated with protease no longer promoted S. aureus surface colonization, suggesting the involvement of protein factors. The major protein components of secretions were identified and subsequent analysis revealed that hemoglobin possessed the ability to promote S. aureus surface colonization. Immunoprecipitation of hemoglobin from nasal secretions resulted in reduced S. aureus surface colonization. Furthermore, exogenously added hemoglobin significantly decreased the inoculum necessary for nasal colonization in a rodent model. Finally, we found that hemoglobin prevented expression of the agr quorum sensing system and that aberrant constitutive expression of the agr effector molecule, RNAIII, resulted in reduced nasal colonization of S. aureus. Collectively our results suggest that the presence of hemoglobin in nasal secretions contributes to S. aureus nasal colonization.
| Staphylococcus aureus is a medically important human pathogen that is found in the nasal passages of approximately 1/3 of the population. The nose serves as a reservoir for spread of this pathogen and predisposes the host to potential infection. Factors contributing to S. aureus nasal colonization are only beginning to be elucidated. The collection and analysis of human nasal secretions provided evidence that the presence of hemoglobin in nasal secretions can promote S. aureus nasal colonization. Hemoglobin reduced expression of the S. aureus agr quorum sensing regulatory system known to be involved in surface colonization, and it was found that induction of the agr system reduced nasal colonization. These findings suggest that individuals experiencing frequent nosebleeds would be prone to S. aureus colonization and epidemiological data supports these findings. By understanding host factors and bacterial molecular mechanisms involved in nasal colonization we may one day be able to design novel decolonization strategies.
| Staphylococcus aureus is a human commensal and the causative agent of many serious acute and chronic infections [1]. The primary reservoir for S. aureus is the nasal cavity [2], [3]. Asymptomatic colonization occurs in approximately 20% of the normal population, another 60% are transiently colonized and the remaining 20% appear to be rarely or never colonized [4], [5]. Why some individuals are prone to colonization while others resist colonization is not understood. S. aureus nasal colonization is a known risk factor for several infections including bacteremia [6], [7], postoperative infections [8], and diabetic foot ulcer infections [9]. Treatment with the topical antibiotic mupirocin has proven to be effective at reducing nasal colonization and the risk of postoperative infection [3], [6]. However the appearance of mupirocin resistance threatens this nasal eradication strategy [10]. Therefore an improved understanding of nasal carriage is needed to foster development of new strategies to reduce colonization and subsequent infection.
S. aureus nasal colonization likely involves both host and bacterial determinants. Studies analyzing patterns of nasal carriage suggest that host factors may influence S. aureus nasal colonization [3]. Different carriage rates have been observed among different ethnic groups and families [11], [12]. Furthermore, human nasal secretions show variability in supporting growth or having antimicrobial activity against S. aureus [13]. Nasal host receptors may also vary among individuals as S. aureus adherence to desquamated epithelial cells from carriers is significantly greater than for non-carriers [14]. Bacterial products influencing colonization have also been identified. These include sortase A, teichoic acid, clumping factor B, capsule, iron-regulated surface determinant A, alkyl hydroperoxide reductase, catalase, and the autolysin SceD [15]–[21]. In addition, recent evidence has shown that polymicrobial interactions likely play a role in S. aureus nasal colonization [22], [23]. Taken together these studies suggest that nasal carriage is a multifactorial process that is influenced by host determinants, bacterial products, and polymicrobial interactions.
The goal of this work was to determine if components of human nasal secretions are capable of promoting S. aureus colonization. Human nasal secretions were collected and examined for their ability to promote S. aureus surface colonization. Analysis of nasal secretions that promoted S. aureus surface colonization revealed that hemoglobin was both necessary and sufficient for this activity. Hemoglobin promoted surface colonization in a multitude of S. aureus strains. Furthermore, the addition of hemoglobin to a cotton rat model reduced the inoculum size necessary to establish S. aureus nasal colonization. Finally, hemoglobin was found to inhibit induction of the agr quorum sensing system and a construct expressing the agr effector molecule RNAIII displayed decreased nasal colonization. Our findings suggest that nasal secretions containing hemoglobin have the ability to modulate S. aureus gene expression and increase nasal colonization.
Animal work in this study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Committee on Use and Care of Animals (UCUCA) of the University of Michigan (Permit Number:10394). All efforts were made to minimize pain and discomfort during the procedure. Work involving collection of nasal secretions from human subjects was approved by the University of Michigan Institutional Review Board, approval number IRB00001996. Written informed consent was provided by study participants.
The bacterial strains used in this study are described in Table 1. Strains of Escherichia coli were grown in Luria-Bertani broth or Luria agar plates, and growth medium was supplemented with ampicillin (100 µg/ml) or chloramphenicol (10 µg/ml) as needed for maintenance of plasmids. Except where noted, S. aureus strains were grown in tryptic soy broth (TSB) or tryptic soy agar (TSA). For selection of chromosomal markers or maintenance of plasmids, S. aureus antibiotic concentrations were (in µg/ml): erythromycin (Erm) 10; chloramphenicol (Cam) 10. All reagents were purchased from Fisher Scientific (Pittsburg, PA) or Sigma (St. Louis, MO) unless otherwise indicated. The human proteins used were purchased from Sigma at the highest available purity: hemoglobin (H7379), IgK (K3502), carbonic anhydrase (C4396), orosomucoid (G9885), IgG (I4506), albumin (A4327), hemopexin (H9291), transferrin (90190), lactoferrin (L0520), plasminogen (P7999), myoglobin (M0630), fibronectin (F2006), and collagen (C7521). Apohemoglobin was prepared using the cold acid acetone precipitation method [24].
Restriction and modification enzymes were purchased from New England Biolabs (Beverly, MA). All DNA manipulations were performed in E. coli strain DH5α. Oligonucleotides were synthesized at Integrated DNA Technologies (Coralville, IA). Plasmids were transformed into S. aureus RN4220 by electroporation and then purified and moved to other indicated S. aureus strains by electroporation. Plasmid pALC2073-RNAIII was made by PCR amplification of RNAIII using primers GTTGTTGAATTCTTCATTACAAAAAAGGCCGCGAGCTTGGGA and GTTGTTGGTACCAGATCACAGAGATGTGATGGAAAATAGTTG. The PCR product was digested with KpnI and EcoRI and ligated into the pALC2084 vector that had been digested with the same restriction enzymes.
The cotton rat nasal colonization model described by Kokai-Kun was utilized in this study [25]. Briefly, S. aureus was grown overnight in TSB, harvested by centrifugation, washed and resuspended in phosphate buffered saline (PBS) or PBS supplemented with protein (hemoglobin, myoglobin, or apohemoglobin (5 mg/ml)) or/and anhydrotetracycline (200 ng/ml) as indicated. Cotton rats were anesthetized, and a 10-µl aliquot containing 1×108 or 1×105 colony forming units (CFUs) was intranasally instilled drop-wise equally between the two nostrils. After 5 days the animals were sacrificed and the noses were surgically removed. The noses were placed in 1 ml of PBS containing 0.5% Tween-20, homogenized, and dilution plated onto TSA supplemented with 7.5% NaCl and spectinomycin (200 µg/ml) to determine CFUs. All animal experiments were conducted in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the University of Michigan University Committee on Use and Care of Animals (UCUCA) approval number 10394.
Nasal secretions were collected from a convenience sample of 20 adult patients in a rhiniology clinic within the University of Michigan Department of Otolaryngology and 10 healthy volunteers not visiting a clinic (University of Michigan Institutional Review Board approval number IRB00001996). Written informed consent was provided by study participants. Clinic patients were seen for evaluation of a variety of concerns, including nosebleeds, nasal blockage, rhinitis, and sinusitis. Nasal examination was performed as part of the routine clinical evaluation. For experimental purposes, the anterior nare was swabbed and the swab was streaked onto Mannitol Salt Agar to determine colonization by S. aureus. Secretions were collected by thoroughly swabbing the anterior and posterior nasal passageways with a sterile cotton swab (Remel BactiSwab). The tip of the swab was cut from the shaft, placed in a ridged eppendorf tube and centrifuged to obtain secretions. Volumes of secretions obtained varied from ∼50 µl to ∼200 µl. Secretions were stored at −20°C. Collected secretions were vortexed and sonicated to break up clumps then passed through a 0.22 µm syringe filter. To determine major protein composition of the secretions, 15 µl of samples were separated by 10% sodium dodecyl sulfate (SDS)-PAGE and stained with Sypro Ruby (Biorad). Visible bands were identified by in gel trypsin digestion and subsequent LC-MS/MS analysis (MS Bioworks, Ann Arbor, MI). Statistical analysis was utilized to determine the major protein component of each excised band and the major protein constituents were used for further experiments. The value for the abundance measurement is the Normalized Spectral Abundance Factor (NSAF) [26]. Western blots to detect the presence of hemoglobin in nasal secretions were done using the primary antibody anti-Human Hemoglobin whole antiserum produced in rabbit (Sigma H4890) at a dilution of 1∶100 in 5% milk. The secondary antibody was anti-rabbit IgG produced in goat, conjugated to peroxidase (Sigma A0545), used at a concentration of 1∶10,000 diluted in 5% milk.
Microtiter plate biofilms and flow cell biofilms were grown as previously described [27]. The growth medium for microtiter biofilms was 66% TSB supplemented with 0.2% glucose or with 10% (volume) nasal secretion or indicated concentration of protein. Microtiter plates used for this assay were cell culture treated 96 well (Costar 3596) or 384 well (Nunc 164688) plates. Flow cell biofilms were grown in 2% TSB with 0.2% glucose or 5 µg/ml hemoglobin. Confocal scanning laser microscopy and image analysis was performed as described previously [27]. Adherence assays to collagen and fibronectin were performed as previously described [28], [29]. Briefly, 96-well plates (Nunc 265301) were coated with the following proteins in PBS: Type IV collagen from human placenta (20 µg/ml; Sigma), fibronectin from human plasma (1 µg/ml; Sigma), or Bovine Serum Albumin (10 mg/ml; Fisher). 100 µl of protein was added to each well and incubated overnight at 4° with constant agitation. The wells were then washed three times with 1% BSA in PBS for 20 minutes at 37° with agitation. S. aureus was grown overnight in Brain Heart Infusion with 50 µg/ml of Human Hemoglobin. Following a one hour incubation at 37°C with agitation, wells were washed three times with PBS to remove non-adherent bacteria, then fixed with 2.5% glutaraldehyde in PBS for two hours at 4°C. Bacteria were stained with .1% crystal violet for 30 minutes at room temperature then washed three times with water. The stain was extracted by incubation with .2% Triton X-100 for 30 minutes at room temperature. The absorbance at 570 nm was read on a microtiter plate reader (Tecan Infinite M200).
To monitor expression from the RNAIII promoter (P3), an overnight culture of the appropriate reporter strain was inoculated into TSB with Cam and grown to an optical density at 600 nm of 0.05. In triplicate, 475 µl of reporter culture was aliquoted into test tubes and indicated amounts of hemoglobin (or other protein) were added. The tubes were shaken at 250 rpm at 37°C for 12 hours. Both cell density (optical density at 595 nm) and green fluorescent protein (GFP) fluorescence (excitation at 485 nm, emission at 535 nm) were measured in a Tecan Infinite M200 (Research Triangle Park, NC) microtiter plate reader by removing 100 µl from each tube and assaying in a microtiter plate (Corning 3606).
S. aureus type 1 autoinducing peptide (AIP) was prepared as previously described using strain AH426 [27]. Briefly, an overnight preculture of expression strain AH426 was prepared and inoculated into 100 ml of Luria-Bertani broth with Amp. The culture was grown at 37°C with shaking until an optical density at 600 nm of 0.5 was reached, and IPTG (isopropyl-ß-d-thiogalactopyranoside) was added to a 0.5 mM final concentration. The culture was grown with shaking at 30°C for 3 h, and the cell pellets were stored at −70°C. Cell pellets were resuspended in 20 ml chitin binding buffer consisting of 100 mM phosphate buffer, pH 7.0, with 500 mM NaCl, 1 mM EDTA, 150 µl protease inhibitor cocktail (Sigma; catalog number P8465), and 0.5 mM phenylmethylsulfonyl fluoride. The cell suspension was lysed through two passes in a French press, and insoluble material was removed by centrifugation at 19,000 rpm for 30 min at 4°C in a Beckman JA-20 rotor. The supernatant was removed, 4 ml equilibrated 50% chitin beads (New England Biolabs) was added, and the resin suspension was mixed gently at room temperature for 30 min. The chitin resin was removed by centrifugation at 500×g for 5 min. The supernatant was removed, and the resin was washed three times for 5 min with 25 volumes of chitin binding buffer. The resin suspension was poured into a 10-ml column and allowed to settle by gravity (2-ml final resin volume), and the resin was equilibrated with three column volumes of elution buffer [100 mM phosphate, pH 7, 50 mM NaCl, 1 mM EDTA, 1 mM tris(2-carboxyethyl)phosphine (TCEP)]. Gravity flow from the column was stopped, and the resin was left sealed at room temperature for 15 h. Following incubation, fractions were eluted and assayed for activity or saved at −20°C. To determine AIP concentration the following was done: A Sep-Pak Plus cartridge (Waters, Milford, MA) was conditioned according to the manufacturer's instructions. To remove TCEP, an AIP sample from an intein purification was bound to the cartridge, washed with 20 ml of water with 0.1% trifluoroacetic acid (TFA), and eluted with 2 ml of 60% acetonitrile with 0.1% TFA. The concentration of the AIP was determined using assays with 5,5′-dithio-bis-(2-nitrobenzoic acid) (DTNB), also called Ellman's reagent (Pierce, Rockford, IL). The thiolactone ring was opened with 1 M (final concentration) NaOH and neutralized with HCl, and DTNB assays were performed before and after base treatment. For the assays, a 1-ml reaction mixture was prepared with 100 mM Tris-HCl, pH 8, and 0.1 mM DTNB (prepared fresh) and different amounts of untreated and base-treated AIP were added. The reaction mixtures were incubated for 10 min at room temperature, and the absorbance was measure at 412 nm. The concentrations were determined with an extinction coefficient of 13,600 M−1 cm−1, and the prebase reading was subtracted to get the final AIP concentration.
AIP activation assays were performed by incubating 100 nM of AIP with a human hemoglobin solution (5 µg/ml in PBS) for 2 hours at 30°C. Hemoglobin was removed from indicated samples and nasal secretions by adding 2 µl anti-hemoglobin antibody produced in rabbit (Sigma) with a 1 hour incubation, followed by addition of 5 µg of goat anti-rabbit IgG Magnetic Beads (New England Biolabs) with a 1 hour incubation at room temperature with periodic rocking. Tubes were then placed in a magnetic tube separation rack for 30 minutes, after which time the supernatant was collected. The AIP activation assay was performed by growing overnight cultures of S. aureus reporter strain (AH462) in TSB and subculturing 1∶50 into TSB plus 0.2% glucose supplemented with AIP; AIP+hemoglobin; or AIP pretreated with hemoglobin that was removed by antibody pulldown as described above. Cultures were grown in test tubes at 37°C with shaking at 250 rpm. Cell density and fluorescence were monitored at 8 hours after inoculation.
Statistics were performed using a 1-way analysis of variance (ANOVA). Results are expressed as mean ± standard error of the mean, unless otherwise indicated.
To identify host factors that contribute to S. aureus nasal carriage, we collected nasal secretions from patients at the University of Michigan Otolaryngology Clinic and analyzed them for the ability to promote S. aureus surface colonization. Tryptic soy broth was supplemented with filtered nasal secretions (10% of total volume) from 9 subjects and analyzed for the ability to promote S. aureus surface colonization in triplicate wells of a 384 well plate. 384 well plates were utilized for this assay to reduce the amount of volume necessary, as a limited volume of secretion (∼50–200 µL) was collected from each subject. Secretions 1, 2, and 5 significantly increased S. aureus surface colonization (Figure 1). Protease treatment of secretions 1, 2, and 5 with pepsin-coated beads resulted in the loss of ability to promote surface colonization, suggesting a protein component was responsible for induction of surface colonization. Nasal swabs taken at the time of secretion collection revealed that individuals 1, 2, 3, 5, and 7 were colonized with S. aureus (Figure 2A).
The nasal secretions were next examined by sodium dodecyl sulfate polyacrylamide gel electrophoresis (SDS-PAGE) and all visible protein bands from secretion 1 were identified by mass spectrometry (Figure 2A). Secretion 1 contained plasminogen, lactoferrin, transferrin, hemopexin, albumin, IgG, orosomucoid, carbonic anhydrase, IgK, and hemoglobin (alpha and beta subunits) (Figure 2B). Next, these proteins were analyzed individually to determine if any were capable of promoting S. aureus surface colonization (Figure 2C). We found that purified hemoglobin promoted S. aureus surface colonization. None of the other proteins tested promoted S. aureus surface colonization.
Hemoglobin concentrations of 10 µg/ml and higher were found to significantly increase S. aureus surface colonization in microtiter plate assays and flow cell assays (Figure 3A and 3B). Other heme containing proteins, such as myoglobin, or heme itself did not possess the ability to promote S. aureus surface colonization (Figure 3A). In addition, apohemoglobin promoted surface colonization in the same concentration range as hemoglobin, suggesting iron is not the cause of this phenomenon.
Adherence assays were also performed to determine if growth in the presence of hemoglobin could influence initial attachment to the human extracellular matrix proteins collagen and fibronectin (Figure 3C). S. aureus grown in the presence of hemoglobin attached to both collagen and fibronectin at significantly higher levels than S. aureus grown in the absence of hemoglobin. These results are consistent with previous reports suggesting that growth on blood agar plates is ideal for expression of surface binding proteins by S. aureus [29], [30].
Because nasal secretions are complex mixtures of protein, sugars, and salts, hemoglobin was specifically depleted by immunoprecipitation to test the hypothesis that hemoglobin was a necessary factor in inducing bacterial surface colonization. Additional nasal secretions were collected and one that contained hemoglobin and promoted S. aureus surface colonization was processed as described above to remove clumps and debris. Half of this secretion was left untreated while the other half was incubated with anti-hemoglobin antibody that was then immunoprecipitated with a secondary antibody conjugated to magnetic beads. Figure 4A shows SDS-PAGE analysis of the untreated (lane 1) and hemoglobin immunoprecipitated nasal secretion (lane 2). Immunoprecipitation of hemoglobin significantly reduced the ability of this nasal secretion to promote S. aureus surface colonization (Figure 4B). Of note, one protein band that was not hemoglobin nonspecifically immunoprecipitated in this experiment so we cannot rule out the possibility that this unidentified protein could have an impact on the colonization assay.
We next collected nasal secretions and performed nasal swabs on 11 additional individuals visiting the clinic and 10 healthy individuals to determine if a correlation between the presence of hemoglobin in nasal secretions and S. aureus colonization persist in an additional sampling (Figure 2D and 2E). Six of the secretions from clinic patients (Figure 2D lanes 1,3,4,6,9,11) had detectable hemoglobin by western blotting and five of these secretions came from individuals determined to be nasally colonized with S. aureus. Analysis of healthy volunteers revealed detectable hemoglobin in nasal secretions from three individuals and all three of these volunteers were nasally colonized with S. aureus (Figure 2E lanes 3,5,6).
To determine if the presence of hemoglobin could influence S. aureus nasal colonization, we utilized the cotton rat model [25]. Nasal instillation with 10 µl of a S. aureus suspension at a density of 1×108 colony forming units resulted in reproducible colonization after 5 days (Figure 5). A reduced inoculum, 10 µl of a S. aureus suspension at a density of 1×105 colony forming units, resulted in no isolation of nasal S. aureus after 5 days. However, if the reduced inoculum (1×105) was suspended in a solution of hemoglobin at 5 mg/ml, nasal colonization was reproducibly observed 5 days after instillation. This high concentration of hemoglobin was utilized in an attempt to keep hemoglobin present in the nasal passageway for as long as possible. Myoglobin supplementation of the reduced inoculum (1×105) did not significantly increase S. aureus nasal colonization. Supplementation with apohemoglobin resulted in robust nasal colonization similar to the hemoglobin condition. These results suggest that the presence of hemoglobin in nasal secretions, but not all heme containing proteins, can increase the likelihood of S. aureus nasal colonization given a small inoculum.
Biofilm formation and lack of biofilm dispersal often correlate with reduced expression of the agr quorum sensing system [27], [31], [32], [33]. Induction of the agr system results in the increased production of several secreted virulence factors including proteases, hemolysins, and toxins [34], [35]. Because a recent report by Schlievert et. al. [36] demonstrated that hemoglobin found in menses inhibits production of secreted exotoxins, we hypothesized that hemoglobin would inhibit agr expression. To determine if hemoglobin was affecting agr expression we followed expression of the quorum sensing responsive promoter fusion, P3-GFP, in the presence of increasing concentrations of hemoglobin (Figure 6A). Hemoglobin significantly inhibited expression from the P3 promoter measured after 12 hours of growth at concentrations from 10–100 µg/ml. Control experiments with myoglobin did not result in reduced expression from the P3 promoter but supplementation with apohemoglobin inhibited P3 expression, suggesting the activity is specific to the hemoglobin peptide rather than any heme containing protein.
We reasoned that hemoglobin could antagonize activation of the agr P3 promoter either by sequestering autoinducing peptide (AIP) or directly binding to the cell surface and interfering with the agr activation process. To delineate between these two possibilities we incubated purified AIP with hemoglobin, removed the hemoglobin by immunoprecipitation, and assayed the ability of the remaining solution to activate expression from the P3 promoter (Figure 6B). A mixture of hemoglobin and AIP at a concentration of 50 nM was unable to activate expression of the P3 promoter compared to AIP alone. However, when AIP concentration levels were increased this inhibition was overcome. If hemoglobin was depleted by immunoprecipitation from a mixture containing 50 nM AIP the remaining solution possessed P3 activation activity, suggesting that hemoglobin does not sequester AIP. In addition, S. aureus biofilms grown in the presence of hemoglobin dispersed upon addition of AIP (Figure 6D).
There are four types of agr quorum sensing systems among S. aureus strains. Each agr system (agr-I through agr-IV) recognizes a unique autoinducing peptide structure (AIP-1 through AIP-4). These agr types can be divided into three cross-inhibitory groups: agr-I/IV, agr-II, and agr-III. Hemoglobin was found to inhibit induction of the P3 promoter in each agr class and in multiple strains (Figure 6C). We also examined mutants in receptors known to bind hemoglobin (isdB and isdH) [37] and found no difference in the ability of hemoglobin to inhibit agr quorum sensing this these genetic backgrounds. These results suggest the ability of hemoglobin to inhibit quorum sensing is spread across all agr types and does not occur by binding of the AIP peptide or via binding to known hemoglobin receptors (IsdB or IsdH).
Recent work by Burian et. al. has shown that expression of the agr quorum sensing system is minimal during nasal colonization [38], [39]. Based on this finding and the knowledge that hemoglobin can promote S. aureus nasal colonization and inhibit the expression of the agr quorum sensing system, we hypothesized that expression of RNAIII, the agr quorum sensing system effector molecule, would reduce nasal colonization. To test this hypothesis we utilized a plasmid containing promoterless RNAIII cloned behind a tetracycline inducible promoter, pALC2073-RNAIII. A similar construct has previously been shown to reduce expression of surface associated adhesins and reduce biofilm formation in the presence of tetracycline [40]. In the rat nasal colonization model, presence of pALC2073-RNAIII resulted in reduced nasal colonization, even in the presence of hemoglobin (Figure 7). These results suggest that repression of the agr quorum sensing system is necessary for efficient S. aureus nasal colonization and factors capable of inhibiting agr expression may promote nasal colonization.
Nasal carriage of S. aureus is a major risk factor for developing a range of infections in both clinical and community settings [6]–[9]. Nasal colonization is multifactorial, likely involving both host and bacterial determinants [11]–[21], [41], [42]. Current strategies for eradication of S. aureus nasal carriage include the use of topical mupirocin; however mupirocin resistance is appearing suggesting that susceptibility may not be long lasting [43], [44]. Therefore, further advances in the control of S. aureus colonization are needed and will depend on an in-depth understanding of both host and bacterial determinants of carriage. In this work, we describe a host factor, hemoglobin, and a bacterial molecular mechanism, expression of the agr quorum sensing system, that influence S. aureus nasal colonization.
The collection and analysis of human nasal secretions from nine volunteers revealed that secretions from three subjects had the ability to significantly promote S. aureus surface colonization (Figure 1). Protease digestion of these secretions eliminated their ability to promote surface colonization, suggesting the involvement of proteinaceous factors. Major protein components of one of the secretions were identified and individual analysis of each protein revealed that hemoglobin possessed the ability to promote S. aureus surface colonization (Figures 2 and 3). Depletion of hemoglobin from a nasal secretion resulted in reduced S. aureus surface colonization (Figure 4). Furthermore, in a rat model the presence of hemoglobin reduced the size of the inoculum necessary to achieve nasal colonization (Figure 5). Finally, we found that hemoglobin prevented expression of the agr quorum sensing system (Figure 6) and the aberrant constitutive expression of RNAIII resulted in reduced nasal colonization (Figure 7). Collectively these results suggest that hemoglobin is a host factor whose presence in nasal secretions contributes to S. aureus nasal colonization.
Several S. aureus factors have been described as determinants for nasal colonization, including: wall teichoic acid, the iron-regulated surface determinant A, catalase, alkyl hydroperoxide reductase, clumping factor B, sortase A, and the autolysin SceD [15]–[21]. Here we show that the aberrant constitutive expression of RNAIII, the agr quorum sensing effector molecule, results in reduced nasal colonization. This finding correlates well with recent data that agr is not expressed during nasal colonization in the cotton rat model or in humans [38], [39]. At this time it is unclear which agr regulated factors influence nasal colonization. Recent work has demonstrated that the Staphylococcus epidermidis secreted protease Esp is capable of preventing and eradicating S. aureus nasal colonization [22]. Considering that the expression of several secreted proteases are up-regulated by the agr regulatory system and the expression of agr-regulated proteases results in biofilm dispersal [27], [35], we speculate that S. aureus protease expression may result in reduced nasal colonization. Further work is needed to elucidate the role of agr-regulated factors in nasal colonization.
The mechanism by which hemoglobin influences S. aureus agr expression and nasal colonization is not clear. Hemoglobin is an iron (heme) containing protein, but other iron or heme containing proteins did not produce the same effect and apohemoglobin elicited the same responses as hemoglobin. Our data suggest that hemoglobin is not able to bind and sequester autoinducing peptides, so hemoglobin likely acts by binding to the cell surface. Mutations of known hemoglobin binding proteins, IsdB and IsdH, exhibited no difference in their response to hemoglobin, suggesting they are not involved in this phenomenon. Agr inhibition also occurs in multiple strains and agr types. Hemoglobin has regions of high positive charge that could promote interaction with negatively charged phospholipids and interfere with activation of membrane histidine kinases such as AgrC. Indeed, work by Scheilvert et al. demonstrated that exotoxin regulation via the SrrA-SrrB two component system was affected by the presence of hemoglobin [36].
Our findings are consistent with a recent study that found individuals experiencing epistaxis (nose bleeds) are more likely to be nasally colonized with S. aureus than control individuals [45]. Some nasal secretions used in our study were collected from patients experiencing epistaxis, nasal blockage, rhinitis, or sinusitis, which presumably increased the likelihood of hemoglobin being present in their secretions. However microepistaxis is thought to be common and can result from digital trauma (i.e., nose picking), nasal or sinus infections, dry ambient air, and from the use of topical nasal medications such as antihistamines and corticosteroids [46], [47]. Sampling of nasal secretions from 10 healthy individuals revealed the presence of hemoglobin in secretions from 3 individuals, all of whom were nasally colonized by S. aureus. Therefore it seems possible that hemoglobin is commonly present in nasal secretions and we propose this could contribute to S. aureus nasal colonization. A large sampling of healthy individuals is needed to determine if there is a significant correlation between S. aureus nasal colonization and the presence of hemoglobin in nasal secretions.
The presence of hemoglobin may also play an important role in S. aureus colonization of other body sites. Recent work has demonstrated that staphylococcal exotoxins were not produced when the organism was cultured in human menses and hemoglobin was identified as the inhibitory factor [36]. It seems plausible that hemoglobin present in vaginal secretions and menses could have profound effects on S. aureus vaginal colonization. Exposure to hemoglobin could also influence S. aureus bloodstream infections by limiting exotoxin production and promoting biofilm infections such as infective endocarditis.
Why would S. aureus have evolved to down regulate the agr quorum sensing system and colonize surfaces in the presence of hemoglobin? Since S. aureus preferentially uses hemoglobin as an iron source [48], it is likely advantageous to form a surface associated community where the scarce resource, iron, is available. Another non-exclusive possibility is that the production of some agr regulated secreted proteases could cleave hemoglobin, releasing antimicrobial peptides. Others have shown hemoglobin peptide fragments found in menses and ticks possess antimicrobial activity against S. aureus [49], [50]. Therefore by not producing agr-regulated proteases when hemoglobin is present, S. aureus may avoid the generation of these antimicrobials. In any case, S. aureus has evolved to respond to one of the most abundant proteins found in humans and this interaction could have profound effects on S. aureus colonization and pathogenesis.
|
10.1371/journal.ppat.1005434 | Establishment of a Wolbachia Superinfection in Aedes aegypti Mosquitoes as a Potential Approach for Future Resistance Management | Wolbachia pipientis is an endosymbiotic bacterium estimated to chronically infect between 40–75% of all arthropod species. Aedes aegypti, the principle mosquito vector of dengue virus (DENV), is not a natural host of Wolbachia. The transinfection of Wolbachia strains such as wAlbB, wMel and wMelPop-CLA into Ae. aegypti has been shown to significantly reduce the vector competence of this mosquito for a range of human pathogens in the laboratory. This has led to wMel-transinfected Ae. aegypti currently being released in five countries to evaluate its effectiveness to control dengue disease in human populations. Here we describe the generation of a superinfected Ae. aegypti mosquito line simultaneously infected with two avirulent Wolbachia strains, wMel and wAlbB. The line carries a high overall Wolbachia density and tissue localisation of the individual strains is very similar to each respective single infected parental line. The superinfected line induces unidirectional cytoplasmic incompatibility (CI) when crossed to each single infected parental line, suggesting that the superinfection would have the capacity to replace either of the single constituent infections already present in a mosquito population. No significant differences in fitness parameters were observed between the superinfected line and the parental lines under the experimental conditions tested. Finally, the superinfected line blocks DENV replication more efficiently than the single wMel strain when challenged with blood meals from viremic dengue patients. These results suggest that the deployment of superinfections could be used to replace single infections and may represent an effective strategy to help manage potential resistance by DENV to field deployments of single infected strains.
| Dengue fever is a viral disease transmitted by Aedes aegypti mosquitoes and more than 30% of the world’s population is at risk. The control of dengue virus (DENV) transmission has been problematic as no vaccines or drugs are effective against the four serotypes. Vector control of mosquitoes during epidemics is considered the only option to prevent transmission. Recently, a novel biocontrol method using the endosymbiotic bacterium Wolbachia has been developed in which DENV replication is significantly inhibited in Wolbachia-infected Ae. aegypti. This bacterium also induces a reproductive phenotype called cytoplasmic incompatibility that allows rapid invasion of uninfected mosquito populations. Like any control method, evolutionary responses are expected of the system that might limit its future effectiveness. Here we report the generation and characterization of a superinfected Ae. aegypti line containing two Wolbachia strains (wMel and wAlbB). We show that stable Wolbachia superinfections are more effective at blocking dengue than single infections. Superinfections also demonstrate a cytoplasmic incompatibility phenotype that should enable them to replace single infections in the field. This represents a potential mechanism for resistance management in regions where single infections have already been deployed.
| The endosymbiotic bacterium Wolbachia pipientis was first discovered in 1924 by Marshall Hertig and Burt Wolbach in ovaries of the mosquito Culex pipiens [1]. Wolbachia is a Gram-negative, obligate endosymbiont that is maternally transmitted [2]. It is estimated that around 40–75% of all arthropod species are infected with Wolbachia [3, 4] and the phenomenal success of this bacterium has been attributed to its ability to manipulate the reproductive biology of its host to provide it with a vertical transmission advantage in host populations [5]. These manipulations include feminization, parthenogenesis, cytoplasmic incompatibility (CI) and male-killing [6, 7]. Of these reproductive phenotypes, CI is probably the best studied and describes the phenomenon of early embryonic death resulting from crosses between an infected male and uninfected female or in crosses involving two different Wolbachia strains [7, 8].
More recently, Wolbachia has been shown to limit pathogen replication, in particular the enveloped, positive single-stranded RNA viruses such as dengue (DENV), yellow fever (YFV) and chikungunya (CHIKV) [9–12]. Wolbachia also inhibits additional human pathogens transmitted by mosquitoes including filarial nematodes [13] and malaria parasites [14–16]. The mechanism of pathogen inhibition by Wolbachia is still being investigated, but blocking has been linked to priming of the host innate immune system and competition for limited resources between pathogens and Wolbachia [17, 18].
The ability of Wolbachia to limit pathogen replication has led to the field deployment of Ae. aegypti transinfected with two Drosophila Wolbachia strains, wMel and wMelPop-CLA [19, 20]. wMelPop-CLA is a pathogenic strain that grows to high densities in insect hosts and infected adult insects have significantly reduced lifespan [21]. In contrast, the closely related wMel strain is avirulent and grows to a lower density in most insect tissues. Correspondingly, total DENV inhibition in whole adult wMel-infected mosquitoes is lower than in wMelPop-CLA infected mosquitoes [12]. However, key to the success of such an approach is the use of Wolbachia strains that can successfully invade wild mosquito populations through the action of CI. The wMelPop-CLA Wolbachia strain imposes significant fitness costs to Aedes mosquitoes including reducing fecundity and egg longevity [9, 12, 22, 23]. Although the wMelPop-CLA strain has a stronger inhibitory effect on total DENV replication in whole mosquito bodies, the significant fitness costs were predicted to prevent invasion of wild mosquito populations [24]. Semi-field cage experiments revealed that the wMel strain would likely invade wild mosquito populations at a faster rate than the virulent wMelPop-CLA strain [12]. Based on these findings, the wMel strain was released into two suburbs of Cairns, Australia in 2011 and reached fixation in mosquito populations within a few months [19].
The avirulent Wolbachia strain wAlbB, transinfected from closely related Aedes albopictus mosquitoes, also inhibits DENV replication in Ae. aegypti with smaller fitness costs than wMelPop-CLA [25]. If avirulent Wolbachia strains such as wMel or wAlbB induce the most favourable phenotypic effects for establishment in wild mosquito populations, the potential long-term development of resistance to the inhibitory effects on DENV must be considered. A strategy to overcome the potential development of DENV resistance to either the wMel or wAlbB strains in wild mosquito populations is to release a superinfected line that would ‘sweep over’ the existing single infection. In this study, we describe the generation of an Ae. aegypti mosquito line co-infected with Wolbachia strains wMel and wAlbB. The CI attributes of this superinfected line, named wMelwAlbB, indicate the superinfection should replace either single infection in a population and as such provide a potential mechanism to address resistance if it were to develop. In addition, the superinfected strain shows fitness costs compatible with a successful field deployment and inhibition of DENV that is predicted to have a large impact on dengue transmission in human populations.
Total Wolbachia density in the superinfected Ae. aegypti line was determined using qPCR and primers specific for the gene encoding the Wolbachia surface protein (wsp) in conjunction with the Ae. aegypti rps17 gene to ‘normalise’ for differences in mosquito size. After infection densities had stabilized by generation 18 (G18), the total Wolbachia density in the wMelwAlbB line was higher than in either parental line and comparable to the virulent wMelPop-CLA strain (Fig 1A).
The tissue localization within adult female mosquitoes of both the wMel and wAlbB Wolbachia strains in the superinfected line was determined by fluorescence in situ hybridisation (FISH) in formaldehyde-fixed, paraffin-embedded tissue sections using specific probes against wMel (labelled in red) and wAlbB (labelled in green) (Fig 1B). The Wolbachia tissue tropism in the superinfected line was compared with the wMel and wAlbB strains in the parental, single infected lines. We confirmed the specificity and lack of cross-reactivity of the wMel and wAlbB FISH probes by using both probes against each of the parental lines. No Wolbachia signal was detected in wAlbB mosquitoes when using the wMel probe, and vice versa.
Our FISH studies demonstrated the coexistence of both strains in various tissues within the adult female mosquito body. As expected for maternally transmitted symbionts, both wMel and wAlbB strains were particularly abundant in the ovaries (Fig 1B). In addition, both strains were also found to co-localise in somatic tissues such as fat body, nervous tissue (e.g. thoracic ganglia), Malpighian tubules and salivary glands (S1 Fig). The density of wAlbB in all these tissues was similar in the wMelwAlbB line as in the single wAlbB-infected line. However, wMel was more abundant in the Malpighian tubules, fat body and muscle from the super infected line than in the parental wMel line. The density of wMel in salivary glands appeared to be similar in the super infected Ae. aegypti line as in the single wMel line.
Interestingly, the wMel and wAlbB Wolbachia strains showed quite distinct localisation patterns in ovaries of superinfected wMelwAlbB line females. Whereas wMel was found evenly distributed throughout the whole egg chamber (nurse cells and oocyte), wAlbB was concentrated in the posterior end of the egg chamber that contains the oocyte (Fig 1B). This is similar to the tropism observed in each the parental lines (Fig 1B). These differences in tropism could represent different patterns of binding of these two strains to the host microtubules and dynein as well as kinesin-1 that appear to drive the movement of Wolbachia into the oocyte during oogenesis [26, 27].
Maternal transmission was determined from crosses between wMelwAlbB infected females and uninfected wild type males. We observed 100% maternal transmission for wAlbB across all generations and 97%, 98% and 100% transmission for wMel across generations G12, G14 and G17 respectively (Table 1).
Cytoplasmic incompatibility (CI) was determined by setting up a series of reciprocal crosses between wild type, wMel, wAlbB and wMelwAlbB infected mosquitoes. Viable offspring from each of the crosses was used to determine the level of CI induced by the wMelwAlbB line. Egg hatch rate percentages from different crosses are summarised in Table 2. Crosses between wMelwAlbB infected females and wild type males as well as males infected with wMel, wAlbB and wMelwAlbB resulted in viable offspring while the reciprocal crosses resulted in no viable offspring.
To determine the mosquito fitness costs of Wolbachia superinfection, the longevity (Fig 2) and fecundity and egg survival (Fig 3) of the superinfected line were compared to both uninfected mosquitoes as well as each parental infected line.
To test the extent to which DENV replication is relatively inhibited in the wMelwAlbB line, we first challenged wild type, wMel, wAlbB, wMelPop-CLA and wMelwAlbB infected mosquitoes with DENV-2 using intrathoracic injections. A DENV-2 strain ET300 was injected at a titre of 104 genome copies/mL and mosquitoes were incubated for 7 days. Positive strand DENV-2 RNA genome copies were detected and quantified in whole mosquito bodies using qRT-PCR. Consistent with previous findings, we saw a significant ~ 1 log reduction of DENV-2 genome copies in wMel and wAlbB, whilst in wMelPop-CLA mosquitoes, DENV-2 genome copies were dramatically reduced by ~ 4 logs (Fig 4A). No significant differences in DENV-2 copies between the wMelwAlbB superinfected line and each of the parental lines were observed (Fig 4A). However, DENV-2 infection rates (calculated as the percentage of DENV-2 infected mosquitoes of the total injected) in wMelwAlbB were consistently lower (69%) than both wMel (89%) (Fisher’s exact test, p = 0.034) and wAlbB (100%) (Fisher’s exact test, p>0.0001).
We next challenged wild type, wMel, wAlbB and wMelwAlbB mosquitoes with DENV-2 (ET300) by oral feeding. Defribrinated sheep blood was inoculated with 107 DENV genome copies per ml and 5–6 day old females from each line were allowed to feed for 2 hours using artificial feeders. Fully fed females were selected and incubated for 14 days. Positive strand DENV-2 RNA genome copies were detected and quantified in whole mosquito bodies using qPCR. We found a significant ~1.5 log reduction in DENV-2 genome copies in wMel, wAlbB as well as wMelwAlbB mosquitoes compared to wild type. No significant difference in DENV-2 genome copies between the three Wolbachia-infected lines were found (Fig 4B). We did observe non-significant, lower DENV infection rates in the wMelwAlbB infected line (15%) as compared to the wMel (30%) (Fisher’s exact test, p = 0.41) and wAlbB (35%) (Fisher’s exact test, p = 0.24) infected lines (Fig 4B).
We then assessed the susceptibility of wild type, wMel and wMelwAlbB mosquitoes to DENV infection after feeding on human viremic blood from 43 dengue patients admitted to the Hospital for Tropical Diseases in Ho Chi Minh City, Vietnam. Two feeds were excluded from analysis; a flow chart describing the number of blood fed mosquitoes, their survival and the final cohorts for analyses are described in S2 Fig. The characteristics of the 41 blood donor patients are shown in S1 Table. DENV-1 and DENV-4 were the predominant infecting serotypes in the patient donors (88% of infectious feeds).
The wMel and superinfected wMelwAlbB lines had lower frequencies of DENV infection than wild-type mosquitoes in abdomens and saliva (Fig 5 and Table 3). Across all time points, a total of 42.65% of wild-type mosquitoes had infectious saliva versus 6.57% for wMel and 2.89% for wMelwAlbB (adjusted odds ratio (OR) 0.065; 95% CI = 0.038–0.112; p <0.001 for wMel, and OR 0.025; 95% CI = 0.014–0.043; p < 0.001 for wMelwAlbB versus wild-type) (Table 3). wMelwAlbB further reduced the risk of females having infectious saliva compared to wMel-infected females (OR = 0.377; 95% CI = 0.196–0.725; p = 0.003).
In addition, Wolbachia-infected mosquito strains also had significantly lower concentrations of DENV RNA in their abdomen and salivary gland tissues compared to wild-type mosquitoes (Fig 6A and 6B and S2 Table). wMelwAlbB blocked DENV infection in the salivary glands more efficiently than wMel (Fig 6B). Collectively, these data generated using clinically-relevant virus challenge methods, suggest that the wMelwAlbB strain delivers an incrementally improved DENV blocking phenotype compared to wMel.
Wolbachia has been shown to inhibit pathogen replication in both natural and transinfected insects [9–12, 18, 20]. Combined with Wolbachia’s remarkable evolutionary adaptations to ensure rapid spread and transmission, [5, 28] this bacterium holds promise as an effective biocontrol agent against mosquito-borne diseases such as dengue [20]. Trials with the wMel strain of Wolbachia have shown its establishment and spread in both semi-field [12] and wild populations of Aedes aegypti mosquitoes [19]. However, not all Wolbachia strains are suitable for use in biocontrol strategies. The virulent wMelPop-CLA strain, for example, results in greater overall inhibition of DENV replication in adult female mosquitoes than the avirulent Wolbachia strains, but imparts significantly higher fitness costs [11, 12, 29]. Preliminary trials in Australia and Vietnam in which the wMelPop-CLA strain was released into wild mosquito populations indicate that these fitness costs prevented successful establishment [30].
Modelling projections suggest the establishment of Wolbachia strains in dengue endemic settings will result in a substantial reduction in disease burden [31]. The persistence of an inhibitory effect on DENV replication within wild Wolbachia-infected mosquitoes will be key to the success of any release program. Laboratory vector competence experiments with field (F1) wMel-infected Ae. aegypti mosquitoes, collected one year following field release, indicated very low levels of DENV replication and dissemination [32], demonstrating the persistence of the virus inhibition phenotype. The potential evolution of DENV resistance to Wolbachia’s inhibitory effects must be considered if this biocontrol strategy can be sustainable on a long-term basis. However, the ability to predict the likelihood of resistance development in virus populations will require a greater understanding of the mechanisms of Wolbachia-mediated viral inhibition. Host immune stimulation has been shown to result in antiviral effects in Ae. aegypti [10, 25, 33] but this is not universal for all Wolbachia-mediated antiviral inhibition [34–36]. The density and tissue tropism of Wolbachia strains in insect hosts appears to be the most important factors [12, 37, 38] and competition for shared host resources such as cholesterol has been shown to influence the strength of Wolbachia-induced antiviral effects [17]. High density Wolbachia strains in Drosophila flies provide strong inhibitory effects on insect viruses despite a long-term evolutionary association [11, 39]. Thus, the non-specific nature of the anti-viral environment in Wolbachia-infected Ae. aegypti tissues, coupled with the dominant evolutionary process of purifying selection in DENV populations[40], such that minor variant viruses that arise within individual hosts are lost because they are not infectious to both humans and mosquitoes, creates significant barriers to the emergence of DENV strains that are resistant to Wolbachia. Nonetheless, the association between density and viral inhibition in these natural Wolbachia-host endosymbiotic relationships suggest resistance is less likely to develop for Wolbachia strains that grow to high densities in transinfected insect hosts. Therefore, a superinfection that results in a cumulative higher density Wolbachia infection would be predicted to reduce the potential for DENV resistance development in Ae. aegypti.
In the event DENV does evolve resistance to either the wMel or wAlbB strains in wild mosquito populations, one potential option would be to release a superinfected line that would ‘sweep over’ the existing single infection. For this resistance management strategy to be effective, favourable CI spread dynamics would be needed for a superinfected line to replace existing single Wolbachia infections in wild mosquito populations. The crossing patterns induced by wMelwAlbB (Table 2) indicate that either the wMel or wAlbB strain could be replaced by a superinfection in wild mosquito populations.
The density of Wolbachia strains in transinfected Ae. aegypti mosquitoes is also correlated with mosquito fitness costs [12]. The additive density of Wolbachia strains in the superinfected line, measured at G18 when the line was stable, was comparable to the virulent wMelPop-CLA strain (Fig 1B). Despite the superinfected line resulting in a cumulative high density Wolbachia infection, the effects on the majority of mosquito fitness parameters were very similar to that observed for the single infected wMel line. Under laboratory conditions superinfected males and females had a marginally shorter adult lifespan than uninfected wild type mosquitoes (~10% reduction). The observed effects on adult mosquito longevity of the superinfected line are significantly less than those for the virulent wMelPop-CLA strain, which reduces the lifespan of adult Ae. aegypti mosquitoes by approximately ~50% [9].
In our study, no differences in the number of eggs laid by females (fecundity) from the superinfected line compared to wMel, wAlbB or wild type mosquitoes were observed. Under semi-field conditions, the virulent wMelPop-CLA strain reduced fecundity of Ae. aegypti females by ~60% [12], which may have contributed to the inability of this strain to invade wild mosquito populations [41]. Minimal fecundity costs should increase the potential of the superinfected line to ‘sweep over’ existing single infections in wild mosquito populations. In contrast, survival of eggs from superinfected females during periods of embryonic quiescence was significantly lower than either parental line or wild type mosquitoes. Following two months of storage, ~50% of superinfected eggs were still viable. Although the hatch rates for the superinfected line were lower than that observed for the wAlbB- infected line, the hatch rates were very similar to that of the wMel infected line. Furthermore, hatch rates are still within the average 2-month survival rates (40–60%) for Ae. aegypti eggs during dry seasons [42, 43]. Further experiments under semi-field conditions will be needed to fully determine if the effect on embryonic quiescence is likely to impact the ability of the superinfected line to invade uninfected wild mosquito populations. The results of field releases to date (using wMel) suggest this is unlikely to be a major obstacle to establishing superinfections in the field. The wMel strain successfully invaded wild mosquito populations [19] and the infection remains stable in these release areas [44] despite the observed reduction on embryo hatch rates under laboratory conditions.
The release of a superinfected line for virus resistance management would require the co-infection to provide strong inhibitory effects on DENV replication. Vector competence experiments carried out under laboratory conditions indicated all Wolbachia lines significantly reduced DENV replication as previously reported [12, 25], however the superinfected line provided the greatest resistance. After oral feeding on fresh human viremic blood, the most relevant model to assess mosquito susceptibility to DENV, very few superinfected mosquitoes had infectious virus in their saliva and viral RNA concentrations were substantially reduced in mosquito tissues. These data give reassurance that any population replacement strategy with the superinfected line would be expected to deliver stronger inhibition of DENV transmission than is conferred by wMel.
In summary, the generation and characterisation of a superinfected line with the desired phenotypic effects to replace single wild infections provides a potential mechanism to overcome the emergence of DENV resistance. Both Wolbachia strains are stably maintained in the line with minimal mosquito fitness effects. Importantly, DENV replication is inhibited to a greater extent in the superinfected line compared to both parental lines. The observed CI phenotype induced by the superinfected line is of particular significance as it would enable the line to be released “on top of” existing wMel or wAlbB field releases in dengue endemic areas.
Wolbachia-uninfected Ae. aegypti eggs were collected from Cairns (Queensland, Australia) in 2013 (JCU wild type). The Wolbachia-infected wMel and wAlbB mosquito lines have been described previously [12, 25]. All Ae. aegypti mosquitoes were reared and maintained as described in [9] with the following modification. For hatching, eggs were placed in hatching water (distilled H2O, boiled and supplemented with 50 mg/L fish food [Tetramin]) and allowed to hatch for 24 h. Larvae were subsequently reared at a set density of ~150 in 3 L of distilled water as described in [9]. To prevent genetic drift between wild type and the Wolbachia infected mosquito lines used for analyses, females from each generation of the infected lines were backcrossed with a small proportion (10%) of uninfected field collected male mosquitoes.
Embryonic microinjection, isofemale line establishment and selection for stably-infected lines were done as previously described [9]. In short, the wMel strain was purified from wMel-infected mosquitoes and microinjected into the posterior-pole of wAlbB-infected preblastoderm embryos using methodology previously described [12]. Surviving G0 adult females (~600) from microinjection were mated to wild type males and blood fed for oviposition of the G1 generation. G0 females that laid fertile egg batches were screened using quantitative PCR as described by [17] and primers specific for wMel (forward primer: 5’-CAAATTGCTCTTGTCCTGTGG-3’, reverse primer: 5’-GGGTGTTAAGCAGAGTTACGG-3’) and wAlbB (forward primer: 5’-CCTTACCTCCTGCACAACAA-3’, reverse primer: 5’-GGATTGTCCAGTGGCCTTA-3’). For each sample, quantitative PCR amplification of DNA was performed in duplicate with a LightCycler 480 II Instrument (Roche) using LightCycler 480 SYBR Green I Master (Roche) according to the manufacturer’s protocol. From the ~600 females screened, 21 wMel positives were identified and pooled into two lines. The female progeny from both lines of superinfected females were mated to uninfected field-males for 5 generations (G0-G4) before the lines were considered stably infected with both strains of Wolbachia. One line was selected for further characterisation.
Wolbachia density and distribution in the superinfected line was compared to each of the parental strains using qPCR and fluorescence in situ hybridisation (FISH). Quantitative PCR to determine the total relative Wolbachia densities of infected lines was performed as described by [17] using primers specific to the gene coding for the Wolbachia surface protein (wsp) (forward primer 5’ GCATTTGGTTAYAAAATGGACGA-3’, reverse primer 5’- GGAGTGATAGGCATATCTTCAAT-3’), as well as the Ae. aegypti rps17 gene (forward primer 5’-TCCGTGGTATCTCCATCAAGCT-3’, reverse primer: 5’-CACTTCCGGCACGTAGTTGTC-3’).
Wolbachia was localized in sections of paraffin-embedded 5–7 day old female mosquitoes by FISH, as described in [10], except that only one probe against 16S rRNA was used against each strain and their concentration was increased by 10-fold to improve the signal. wMel was detected using the probe MelPopW6: 5’-GCTTAGCCTCGCGACTTTGCAG-3’, labelled with Alexa 594 dye (red), whereas wAlbB was localized using AlbBW5: 5’-CTTAGGCTTGCGCACCTTGCAA-3’, labelled with Alexa 488 dye (green). 16S rRNA is highly conserved between wMel and wAlbB, therefore the probe was designed against a part of the gene that includes several SNPs. We confirmed the specificity and lack of cross-reactivity of each probe by testing them against the single infected lines (wMel and wAlbB). Both probes were added simultaneously to the wMel, wAlbB and wMelwAlbB mosquito sections in order to obtain the images. DAPI was also used to stain total DNA.
The propagation and maintenance of dengue virus serotype 2 (DENV-2) ET300 was carried out as previously described [18]. For adult microinjections, 40 Ae. aegypti female mosquitoes were anesthetized by briefly exposing them to -20°C. The mosquitoes were subsequently injected intrathoracically with 50 nL of virus solution (104 genomic copies/ml in RPMI [Sigma-Aldrich] media) using a pulled glass capillary and a handheld microinjector (Nanoject II, Drummond Sci.). Injected mosquitoes were incubated for 7 days (40 mosquitoes per cup) at 26°C with 65% relative humidity and a 12h light/dark cycle. For feeding experiments with DENV-2 (ET300) infected blood, 80 Ae. aegypti female mosquitoes were placed in 500 mL plastic containers, starved for 25 hours and allowed to feed on a 50:50 mixture of defibrinated sheep blood and tissue culture supernatant containing 107 genome copies/mL of DENV-2. Feeding was done through a piece of desalted porcine intestine stretched over a water-jacketed membrane feeding apparatus preheated to 37°C for approximately three hours. Fully engorged mosquitoes were placed in 500 mL containers and incubated for 14 days at 26°C with 65% relative humidity and a 12h light/dark cycle.
To quantify DENV-2 genomic copies, total RNA was isolated from DENV-2 injected mosquitoes using the Nucleospin 96 RNA kit (Macherey-Nagel). DENV-2 qPCR analysis was done using cDNA prepared from individual mosquitoes according to [10]. Statistical significance for differences in DENV titres between treatments was determined using a one-way ANOVA with Tukey’s multiple comparison tests (Graph Pad Prism 6c).
Cohorts of 3–5 day old mosquitoes were allowed to feed on fresh, viremic blood from 43 NS1 rapid test-positive patients admitted to the Hospital for Tropical Diseases, in Ho Chi Minh City, Vietnam. Mosquitoes were fed via membrane feeders for a maximum of 1 hour. Fully engorged mosquitoes were placed in 250 mL containers and incubated at 28°C/80% humidity with a 12h light/dark cycle. Mosquitoes were harvested from each blood fed cohort 10, 14 and 18 days later. Detection of infectious virus in the saliva of each mosquito was as described previously [45]. Statistical analyses were performed with the statistical software R, version 3.1.3 (R Foundation for Statistical Computing, Vienna, Austria). Marginal regression models for binary (infected/uninfected mosquitoes) and continuous (tissue viral load) outcomes were fitted using generalized estimating equations with working exchangeable correlation structure to account for potential within-patient correlation.
Blood feeding by volunteers (Monash University human ethics permit no CF11/0766-2011000387) for this study was approved by the Monash University Human Research Ethics Committee (MUHREC). All adult volunteers provided informed written consent; no child participants were involved in the study.
The protocol for feeding mosquitoes with viremic human blood was reviewed and approved by the Ethics Committee of Hospital for Tropical Diseases (HTD), Ho Chi Minh City, Vietnam (approval number CS/ND/12/16), and the Oxford University Tropical Research Ethics Committee (OxTREC) (approval number OxTREC 30–12). All enrolled subjects provided informed written consent.
|
10.1371/journal.pgen.1005508 | A NIMA-Related Kinase Suppresses the Flagellar Instability Associated with the Loss of Multiple Axonemal Structures | CCDC39 and CCDC40 were first identified as causative mutations in primary ciliary dyskinesia patients; cilia from patients show disorganized microtubules, and they are missing both N-DRC and inner dynein arms proteins. In Chlamydomonas, we used immunoblots and microtubule sliding assays to show that mutants in CCDC40 (PF7) and CCDC39 (PF8) fail to assemble N-DRC, several inner dynein arms, tektin, and CCDC39. Enrichment screens for suppression of pf7; pf8 cells led to the isolation of five independent extragenic suppressors defined by four different mutations in a NIMA-related kinase, CNK11. These alleles partially rescue the flagellar length defect, but not the motility defect. The suppressor does not restore the missing N-DRC and inner dynein arm proteins. In addition, the cnk11 mutations partially suppress the short flagella phenotype of N-DRC and axonemal dynein mutants, but do not suppress the motility defects. The tpg1 mutation in TTLL9, a tubulin polyglutamylase, partially suppresses the length phenotype in the same axonemal dynein mutants. In contrast to cnk11, tpg1 does not suppress the short flagella phenotype of pf7. The polyglutamylated tubulin in the proximal region that remains in the tpg1 mutant is reduced further in the pf7; tpg1 double mutant by immunofluorescence. CCDC40, which is needed for docking multiple other axonemal complexes, is needed for tubulin polyglutamylation in the proximal end of the flagella. The CCDC39 and CCDC40 proteins are likely to be involved in recruiting another tubulin glutamylase(s) to the flagella. Another difference between cnk11-1 and tpg1 mutants is that cnk11-1 cells show a faster turnover rate of tubulin at the flagellar tip than in wild-type flagella and tpg1 flagella show a slower rate. The double mutant shows a turnover rate similar to tpg1, which suggests the faster turnover rate in cnk11-1 flagella requires polyglutamylation. Thus, we hypothesize that many short flagella mutants in Chlamydomonas have increased instability of axonemal microtubules. Both CNK11 and tubulin polyglutamylation play roles in regulating the stability of axonemal microtubules.
| Cilia are specialized projections found on the surface of eukaryotic cells. They play crucial sensory functions, as well as motile functions needed for clearing airways or propelling cells. Ciliary motility is perturbed in the inherited disease, Primary Ciliary Dyskinesia (PCD). Two coiled coil domain-containing (CCDC39 and CCDC40) proteins are needed for the assembly of multiple key structures/complexes that are required for generating ciliary motility. Using the unicellular green alga, Chlamydomonas, we have identified a kinase (CNK11) that when mutated is able to partially rescue the short flagella phenotype of the ccdc39 and ccdc40 mutants as well as mutants lacking axonemal dyneins or the N-DRC complex. In addition, CCDC40 is required for tubulin polyglutamylation at the proximal end of flagella. We suggest that substructures like dynein arms and the N-DRC, which are needed for motility, play a second role in stabilizing the axonemal microtubules and are needed for proper length control. The polyglutamylase, TTLL9, and the kinase, CNK11, play roles in stabilizing the axonemal microtubules based on their ability to partially rescue the short flagella phenotypes of multiple mutants.
| Defects in ciliary assembly and function cause a wide range of human diseases and syndromes called ciliopathies. Primary ciliary dyskinesia (PCD) is diagnosed by defects in ciliary motility, and is associated with a genetically heterogeneous group of recessive disorders [1]. Mutations causing PCD have been identified in genes encoding axonemal dynein subunits [2, 3], dynein assembly factors [4–6], and dynein docking/adaptor factors [7, 8]. The nexin-dynein regulatory complex (N-DRC) is an axonemal structure critical for the regulation of dynein motors and for connecting doublet microtubules to each other. Loss-of-function mutations in DRC1 (CCDC164/PF3) and DRC3 (CCDC65) cause severe defects in assembly of the N-DRC structure and result in defective ciliary movement in humans and Chlamydomonas reinhardtii [6, 9, 10]. PF2, which encodes DRC4, was used to identify 11 proteins in the N-DRC [10]. Mutations in CCDC39 and CCDC40 cause altered ciliary beating with the disorganization of the axoneme that includes the displacement of the peripheral outer doublets, as well as central pair microtubules, radial spokes and inner dynein arm defects [11–15]. Loss-of-function mutations in CCDC39 and CCDC40 in Chlamydomonas lead to short flagella, irregularly spaced radial spokes, absence or reduction of N-DRC components and inner dynein arm proteins [16, 17]. CCDC39 and CCDC40 mutations in children lead to earlier and more severe lung disease than in PCD patients with outer dynein arm mutations [18].
In Chlamydomonas, there are many mutations that can lead to short flagella. Partial reduction in IFT proteins (IFT144 (FLA15) and IFT139 (FLA17)) or motors such as the kinesin-2 motor FLA10 or cytoplasmic dynein result in short flagella [19–21]. Changes in the cytoplasmic pool of tubulins and flagellar precursor proteins also affect flagellar length [22, 23]. In addition, the simultaneous loss of multiple substructures, such as the dynein arms, radial spokes, and the N-DRC, result in short flagella [24–26]. LeDizet and Piperno isolated a suppressor (ssh1) that increased the flagellar length in double mutant strains that lacked outer and inner dynein arms without restoring the missing structures [26]. A recent study identified a deletion in the TPG2 gene as the causative mutation in the ssh1 strain [27]. TPG2 encodes FAP234, a flagellar protein that forms a complex with a tubulin polyglutamylase TTLL9/TPG1 [28, 29]. Tubulin polyglutamylation adds multiple glutamates to both α- and β-tubulin subunits along microtubules in cilia/flagella, basal bodies, and neuron axons [30–32]. Several tubulin tyrosine ligase-like (TTLL) proteins carry out the polyglutamylation process. Tubulin polyglutamylation can affect microtubule assembly, stability, and motility [32]. In Chlamydomonas, tpg1 affects polyglutamylation of α-tubulin specifically and shows a flagellar motility defect [29]. Both tpg1 and tpg2 mutations suppress the short flagella phenotype found in mutants that lack multiple axonemal dynein species [27].
NIMA-related protein kinases have been found in eukaryotes and their functions are related to regulation of cell cycle, cilia length, and microtubule stability [33–38]. Currently, there are 11 NIMA-related protein kinases identified in Chlamydomonas [33] and only two of them have been functionally characterized [35, 36]. A null mutant of the NIMA-like protein kinase CNK2 in Chlamydomonas has slightly longer flagella and defective flagellar disassembly. The cnk2-1 mutant has decreased tubulin turnover at the flagellar tip, which suggests that a reduced rate of flagellar disassembly is compensated by reduced rate of assembly [36]. The CNK2 protein, together with a MAP kinase (LF4), respond to flagellar length signals and block assembly and promote disassembly, respectively [36]. Thus, they provide input to the balance between assembly and disassembly of axonemal microtubules and flagellar length.
In this study, we identify a novel NIMA-related protein kinase CNK11 that rescues the short flagella phenotype found in several N-DRC mutants, as well as mutants lacking dynein arms. In addition, we discovered that the polyglutamylation defect caused by tpg1 could not rescue the CCDC40 mutant. Instead, the CCDC40 mutation in the tpg1 background has narrower distribution of polyglutamylated tubulin at the proximal end of flagella. The microtubule stabilizing drug paclitaxel is able to rescue the short flagella phenotype in CCDC39/CCDC40 mutants but this rescue fails in the presence of cnk11 or tpg1.
The pf7 and pf8 mutants were first isolated as mutants with no flagella or short flagella with a motility defect [17, 39] (Fig 1), and mapped to chromosome 17 [40, 41]. Using whole genome sequencing in combination with our SNP and short insertion/deletion library, we identified the causative mutations in both pf7 and pf8 mutant strains [42] (Table 1). A nonsense mutation in Cre17.g698365 (CCDC40) is responsible for the pf7 mutant phenotype; a nonsense mutation in Cre17.g701250 (CCDC39) leads to the pf8 mutation (Table 2). We performed BAC rescue to confirm they are the causative mutations (Fig 1A). Forty-one independent transformants that contain BAC DNA 17F4, which carries the CCDC40 gene, showed restoration of both flagellar length and motility in pf7; 20 independent transformants that contain BAC DNA 31N18, which carries the CCDC39 gene, restored flagellar length and motility in pf8. For each rescue, we analyzed 16 independent transformants and the transformed BAC DNA cosegregates with rescue in all transformants. Independently, Oda et al. showed that pf7 and pf8 encode CCDC40 and CCDC39 [16].
The fla12 mutant was isolated as a temperature-sensitive flagellar assembly mutant [43] that was previously mapped to chromosome 17 [40]. The fla12 cells shorten their flagella gradually and become immotile after the temperature is raised from 21°C to 32°C (Fig 1B). We used the same whole genome sequencing approach to identify a L845P change in CCDC39 in fla12 (Tables 1 and 2). The transgene that rescued the pf8 mutant was introduced into fla12 by meiotic crosses. In 12 independent progeny, the transgene restores normal flagellar length and motility in all strains at 32°C.
Chlamydomonas offers the ability to use suppressor analysis to find genes that restore function to motility mutants [44, 45]. After UV mutagenesis of the pf7 mutant, we screened for swimming cells and recovered 31 independent strains. PCR/enzyme digestion and Sanger sequencing revealed that all 31 strains are intragenic revertants (Table 3). Using the same strategy, we isolated 34 revertants of pf8 and 4 revertants of fla12 (Table 3), all are intragenic events.
Subsequently, we performed two independent UV mutagenesis screens on pf7; pf8 double mutants to isolate extragenic suppressors. In contrast to nitrogen-starved, autolysin-treated cells that assemble ~ 2 μm flagella (Fig 1A), the pf7; pf8 mutant cells in nitrogen-containing medium are mostly aflagellate. The first UV mutagenesis screen led to isolation of three independent strains (pf7; pf8; cnk11-1, pf7; pf8; cnk11-2, and pf7; pf8; cnk11-3) and the second UV mutagenesis screen identified three additional strains (pf7; pf8; cnk11-4, pf7; pf8; cnk11-5, and pf7; pf8; sup2D). All six strains show partial suppression of the aflagellate phenotype of pf7; pf8, but do not suppress the motility defect (Fig 1A). None of them is linked to either pf7 or pf8. They each contain one suppressor mutation based on crosses to the pf7; pf8 parent; the aflagellate phenotype segregates 2:2. The suppressor mutations in five of the strains (cnk11-1 to cnk11-5) are tightly linked to one another (S1 Table). Whole genome sequencing (Table 1) revealed that the five strains each carry a mutation in Cre07.g339100. The causative mutation in the sixth suppressor, sup2D, is currently under analysis.
In the five strains carrying mutations in Cre07.g339100, two nonsense mutations (cnk11-1 and cnk11-2), a frame shift (cnk11-3 and cnk11-5), and a missense mutation (cnk11-4, Table 2) were identified (Fig 2). Using dCAPs markers designed to each mutation (S2 Table), we observed linkage between suppression and the mutant allele in each strain. Cre07.g339100 encodes a 2903 aa (amino acid) protein with a NIMA-like protein kinase (NEK) domain (aa 582–921). This protein is different from the 11 NEKs (CNK1—CNK10, and FA2) that have been previously annotated in Chlamydomonas [33]. Thus, we name it CNK11 (Chlamydomonas NIMA-like protein kinase 11). Using the conserved protein kinase domain, we constructed a phylogenetic tree with using the kinase domains found in 77 NEKs from Arabidopsis, Aspergillus, C. elegans, Chlamydomonas, Dictyostelium, Drosophila, human, mouse, Trypanosoma, rice, Xenopus and zebrafish (S1 Fig and S3 Table). The tree reveals that CNK11 is phylogenetically different from any of the known NEK classes.
In a previous report [16], Oda et al. showed that the pf7 and pf8 single mutants assemble reduced amounts of two N-DRC proteins; DRC2 and DRC4. In our analysis of isolated axonemes, we found that the single mutants assemble reduced amount of DRC1, DRC4, DRC5, DRC7, and DRC11 and completely lack DRC2, DRC3, and CCDC39 (Fig 3A). In addition, the amount of each N-DRC proteins in a pf7; pf8 double mutant and a pf2; pf7; pf8 triple mutant is comparable to that found in the single mutants, with the exception that no DRC4 protein is found in the triple mutant. This is expected given that the pf2 mutant lacks DRC4 (Fig 3A) [10]. The similarity of the single and double mutants is also expected given the co-assembly of CCDC39 and CCDC40 [16]. Transformation of wild-type PF7 or PF8 into the corresponding mutant restores the N-DRC proteins (Fig 3A).
In axonemes from the pf7; pf8; cnk11-1 mutant, the N-DRC proteins are not restored (Fig 3A). This suggests that cnk11 mutants do not suppress the flagellar length defect of pf7; pf8 via assembly of N-DRC proteins (Fig 3A). In addition to the N-DRC proteins, we asked if the pf7 and pf8 mutants affect other axonemal proteins (Fig 3A). Tektin, which is a microtubule binding protein, is diminished in the ida6 (DRC1) and pf3 (DRC2) mutants, which lack the inner dynein arms species e [46]. Tektin is missing in the single, double, and triple mutants (Fig 3A). Because loss of tektin is associated with the loss of dynein e, we suggest that these mutants are likely to lack dynein e as well. The proximally localized minor dynein heavy chain, DHC11 [47], is missing in the single, double, and triple mutants. DIC3/ IC140, which is the intermediate chain for the I1/f inner dynein arm [48, 49], is reduced in pf8, the double and triple mutants, but not in the pf7 mutant. This is one of the few difference found between pf7 and pf8, and was independently validated [16]. DLE2/centrin, which is part of the b, e, and g inner dynein arm complex [50, 51], is slightly reduced in the single mutants. There is no reduction of RSP16, which is one of the radial spoke proteins [52]; or DIC2/IC69, an intermediate chain in the outer dynein arm [53, 54]. DII1/p28 is only slightly reduced based on 2 independent preparations. Similar results were observed by Oda and colleagues [16]. The ribbon proteins, Rib72 and Rib43a, were first identified by their insolubility in various extraction protocols [55, 56]. There is no loss of these two proteins in the pf7 or pf8 mutant compared to wild-type or other N-DRC mutants. LF5, which is a CDKL5 homolog involved in length control that localizes to the proximal 1 μm of the flagella [57], is increased in the single, double, and triple mutants. Since we load equal amounts of protein in each sample, the increase is likely to be due to increased representation of proteins at the proximal end where LF5 localizes. The pf7 and pf8 rescued strains resemble wild-type axonemes and restore all proteins to wild-type levels. The triple mutant pf7; pf8; cnk11-1 shows similar losses to the pf7; pf8 preparations. It indicates that the cnk11 mutant suppresses the pf7; pf8 length phenotype by means other than restoration of axonemal proteins.
In N-DRC mutants such as pf2 and pf3, the presence of 0.1 mM ATP leads to splaying of individual outer doublet microtubules in the medial and distal regions of the isolated full-length axonemes [10]. The proximal end of these axonemes remained intact. In contrast, wild-type axonemes remain intact throughout the whole length. Bower and colleagues concluded that the N-DRC provides some but not all of the resistance to microtubule sliding between doublets. This helps to maintain optimal alignment of doublets for productive flagellar motility [10]. Given pf7 and pf8 axonemes either lack or have reduced amounts of most N-DRC proteins tested, we examined isolated axonemes exposed to 0.1 mM ATP (Fig 4). To verify the proximal end showed splaying, we used antibodies to LF5 that localizes to the proximal end [57]. We observed little or no splaying of the doublet microtubules in wild-type axonemes. In the mutants, we observed splaying in the medial and distal regions of the single, double, and triple mutants that was similar to the splaying observed in N-DRC mutants [10]. The doublets in the proximal 1 μm end remain intact, similar to the observation found in N-DRC mutants. Thus, we conclude that CCDC39 and CCDC40 are not required for holding the microtubules together in the proximal region.
Given the temperature-sensitive fla12 mutant carries a missense mutation in CCDC39, we asked whether the temperature shift affects axonemal proteins and flagellar length in this mutant. Four hours after temperature shift from 21°C to 32°C, fla12 cells contain less CCDC39, DRC1, DRC3, DRC4, DRC7, and tektin while maintaining normal levels of DIC2 and DIC3 (Fig 3B). This suggests that the missense mutation affects the thermal stability of CCDC39, which in turn leads to reduction of N-DRC and axonemal proteins as observed in the null CCDC39 mutant (Fig 3A).
At 21°C, the fla12 mutant assembles slightly shorter flagella (~6.4 μm) when compared to wild-type (~8.5 μm). Eight hours after the temperature shift from 21°C to 32°C, the average flagellar length of fla12 is ~1.6 μm. The flagellar length of the fla12; cnk11-3 double mutant (~6.5 μm) is similar to the single fla12 mutant at 21°C. However after temperature shift, the flagellar length is ~3.9 μm, which is significantly longer than fla12 cells at the same time point (Fig 1B). This indicates that similar to the partial suppression of the short flagella phenotype of pf7; pf8 double mutant, cnk11 can partially suppress the temperature-sensitive short flagella phenotype of fla12.
Intraflagellar transport (IFT) was monitored previously in numerous flagellar assembly mutants [58]. Analysis of the fla12; pf15 double mutant suggested that the velocity of anterograde and retrograde IFT was increased over control velocities, and the frequency of IFT particles was also higher. We reanalyzed IFT by TIRF (Total Internal Reflection Fluorescence) microscopy using GFP-tagged IFT20 [59] in the fla12 mutant. Data obtained from 4 FLA12 and 5 fla12 cells indicate that anterograde (Fig 5C) and retrograde (Fig 5D) IFT velocities in fla12; IFT20-GFP cells are identical to those in FLA12; IFT20-GFP cells. Thus, we conclude that there is no IFT velocity defect in the fla12 mutant. There are at least two possible explanations for the disagreement between these two studies. In the fla12; pf15 study, the pf15 mutation disrupts the p80 subunit of katanin [60]. It is possible that there is some synthetic interaction between katanin and CCDC39 that affects IFT velocity and number. Alternatively, the multiple backcrosses of fla12 before the TIRF study could have removed another mutation that affected IFT.
To ask about the specificity of the cnk11 suppressor, we introduced the cnk11-1 mutation into the pf2 (DRC4) [10] and pf3 (DRC1) [61] mutants through meiotic crosses. Mutants in DRC4 are missing N-DRC proteins as well as dyneins a and c [10]. The pf2 mutant has an average flagellar length of ~5.2 μm (Fig 6A). In comparison, the pf2; cnk11-1 double mutant has an average flagellar length of ~8.6 μm (Fig 6A), which is comparable to the average flagellar length found in wild-type CC-125 cells (~8.9 μm; Fig 1A) and significantly longer than pf2 flagella. The pf3 mutant obtained from the Chlamydomonas Resource Center (CC-1026) has an average flagellar length of 7.4 μm (Fig 6A), slightly shorter than in wild-type cells. Mutants in DRC1 are missing N-DRC proteins, tektin, and RSP13 [10]. PCR on progeny from a meiotic cross between CC-1026 and cnk11-1 revealed that over 8 kb of genomic DNA on chromosome 7 is deleted. The deleted region includes most of the CNK11 gene and at least half of the adjacent gene Cre07.g339104 (Fig 2). Therefore, the strain CC-1026 should be annotated as pf3; cnk11-6. A meiotic cross between pf3; cnk11-6 to wild-type CC-124 cells allowed the isolation of a pf3; CNK11 strain. The average flagellar length of pf3; CNK11 cells is ~4.8 μm (Fig 6A), which is comparable to the length observed in pf2 cells. In addition, the pf3; CNK11 cells have flagellar lengths that are more variable (ranging from <1 μm to >8 μm) than the pf2 (mostly 3~7 μm), pf7, and pf8 (both mostly 1~3 μm) cells (S2 Fig). In conclusion, the cnk11 mutant rescues the short flagella phenotype of CCDC39 and CCDC40 mutants as well as two N-DRC mutants.
The pf22 and pf23 mutants were first isolated as paralyzed flagella mutants and both have short flagella [62]. The PF22 gene encodes a conserved cytoplasmic protein (DNAAF3) that is essential for the assembly of both outer and several inner dynein arms [4]. The pf23 mutant lacks inner dyneins a, c, d, and f [25, 63]. The outer dynein arm mutant, oda2, and inner dynein arm mutant, ida3, both display slow motility with normal flagellar lengths. The oda2; ida3 double mutant is paralyzed with very short flagella (Fig 6B) as has been observed for many oda; ida double mutants [26]. The N-DRC is not affected in any of these mutants. To ask whether the cnk11 suppressor can restore normal flagellar length in these mutants, we introduced cnk11 mutations into pf22, pf23, and oda2; ida3 mutants. The average flagellar length of pf22 is ~4.0 μm and ~6.4 μm for pf22; cnk11-5 (Fig 6A). The average flagellar length of pf23 is ~3.5 μm and ~7.2 μm for pf23; cnk11-5 (Fig 6A). The oda2; ida3 cells have an average flagellar length of ~0.5 μm; and, the oda2; ida3; cnk11-1 triple mutant has an average flagellar length of ~5.9 μm (Fig 6B). Thus, cnk11 mutations rescue the short flagella mutant phenotype of dynein arm deficient mutants, which lack multiple axonemal dynein species and presumably have unstable axonemal microtubules. Similar to the effect of cnk11 on pf7 and pf8, the cnk11 mutations do not rescue the motility defects found in the dynein arm deficient mutants. In addition, the cnk11 mutations do not rescue the temperature-sensitivity flagellar assembly of the kinesin-2 motor mutant, fla10, or the IFT mutants, fla15 and fla17, after 8 hrs at 32°C.
In human cell lines, knockdown of NEK4, a NIMA-like kinase, confers paclitaxel resistance and show defects in repolymerizing microtubules after nocadozole treatment [37]. In Arabidopsis thaliana, nek4, nek5, and nek6 all show hypersensitivity to paclitaxel [38]. Various NEK proteins play a role in microtubule stability. We tested the cnk11-1 allele on paclitaxel media with concentrations from 5 to 60 μM. The mutant strain behaved identically to the wild-type controls and we did not observe resistance or hypersensitivity as judged by cell division and cell size [64].
We then asked whether the addition of 10 μM paclitaxel for 30 minutes, a dosage that does not causes arrest of cell division in wild-type cells [64], would have any effect on flagellar length. In wild-type and cnk11-1 cells, which have normal flagellar lengths, there is no change (Fig 6B). We examined pf2, and the temperature-sensitive kinesin mutant fla10-1 which has about half-length flagella when grown at 28°C [19]. The addition of paclitaxel has no effect on either mutant (Fig 6B). In the short flagella mutants pf7, pf7; pf8, and oda2; ida3, paclitaxel conferred increased flagellar length. In contrast, paclitaxel does not lead to further elongation of flagella of these mutants when the cnk11 mutation is present (Fig 6B). This suggests that CNK11 and paclitaxel could act via the same mechanism to stabilize axonemal microtubules in these short flagella mutants.
Kubo et al. showed that both tpg1 and tpg2 can rescue the short flagella phenotype found in pf23 and pf28; pf30 [27]. PF28 is an allele of ODA2 (the gamma dynein heavy chain) and PF30 is an allele of IDA1 (1-alpha dynein heavy chain, I1/f complex). This result is similar to the effect of cnk11 on pf22, pf23, and oda2; ida3 (Fig 6). Therefore, we asked whether the tpg1 mutation can rescue the short flagella phenotype found in the pf7; pf8 double mutant. The TPG1 gene maps to chromosome 17 at 0.51 Mb, between the PF7 (chromosome 17 at 0.33 Mb) and PF8 (chromosome 17 at 0.74 Mb) genes. The short distance among these three genes makes it extremely hard to generate the pf7; pf8; tpg1 triple mutant by meiotic recombination since it would require two crossovers in an interval of only 4 map units. Given the pf7 mutant behaves similarly to the pf7; pf8 double mutant, we analyzed pf7 and pf7; tpg1 instead of pf7; pf8 and pf7; pf8; tpg1.
To our surprise, the tpg1 mutation does not rescue the short flagella phenotype found in pf7. Instead, the pf7; tpg1 mutant has a more severe flagella phenotype than the pf7 mutant. In nitrogen-free medium, ~85% of pf7 cells have flagella. In contrast, only ~48% of pf7; tpg1 cells have flagella. Measurement of flagellated cells in both strains showed no significant difference in the flagellar lengths between pf7 and pf7; tpg1 (Fig 6B).
We asked whether polyglutamylation of tubulin is altered in the pf7; tpg1 mutant by both immunoblots and immunofluorescence. In wild-type cells, tubulin in axonemal microtubules is polyglutamylated. A polyclonal antibody (Poly E) that specifically recognizes tubulin with three or more glutamates reveals much stronger signal intensity in α-tubulin than in β-tubulin in wild-type axonemes [29]. The signal intensity of polyglutamylated α-tubulin relative to polyglutamylated β-tubulin is significantly reduced in pf7 and pf8 (Fig 7A). Similar to the findings by Kubo et al. [29], we noticed significant reduction of α-tubulin polyglutamylation but not β-tubulin polyglutamylation in tpg1. A significant reduction of α-tubulin polyglutamylation was observed in both pf23; tpg1 and cnk11-1; tpg1 mutants, but not in the pf23 or cnk11-1 mutants. However in the pf7; tpg1 mutant, polyglutamylated α-tubulin remains (Fig 7A). The change of relative signal intensities between polyglutamylated α- and β-tubulins found in pf7, pf8, and pf7; tpg1 is not due to their short flagellar lengths, since the short flagellar length mutant oda2; ida3 has stronger signal intensity in α-tubulin than in β-tubulin, as found in wild-type axonemes (Fig 7A). Immunoblots with an anti-TPG2/FAP234 antibody [28] show that no TPG2 protein is detected in the axonemes of any strain carrying the tpg1 mutant (Fig 7A). It suggests that the presence of small amount of polyglutamylated α-tubulin in pf7; tpg1 is not due to the recruitment or recovery of the TPG1-TPG2 complex in the axoneme. The abundance of TPG2/FAP234 is not significantly affected by flagellar length or the pf7 and pf8 mutations.
By immunofluorescence, the polyglutamylated tubulin detected by the polyE antibody shows signal along the entire length of the axoneme in wild-type and cnk11-1 cells (Fig 7B). As observed previously, the signal in tpg1 is concentrated at the proximal end [29], and we observe that the polyglutamylated tubulin signal is only ~1.5 μm in length (Fig 7C). The cnk11-1; tpg1 and pf23; tpg1 double mutants have a similar stretch of polyglutamylated tubulin signal regardless of their flagellar lengths (Fig 7B). The pf7; tpg1 cells are strikingly different, the polyglutamylated tubulin signal is reduced to ~0.5 μm, which is significantly shorter than in the single or other double mutants (Fig 7B & 7C). This result is different from what we observed in the immunoblots, in which the polyE signal of α-tubulin is more abundant in pf7; tpg1 than in tpg1. It is likely that the difference is due to using isolated flagella that include both the microtubule axoneme and the flagellar membrane/matrix for the immunoblot and using axonemes that have the membrane/matrix fraction removed by detergent for immunofluorescence. Polyglutamylation of α-tubulin but not β-tubulin is associated with soluble tubulin heterodimers [65]. Thus the difference in polyE abundance between the two techniques is likely due to the removal of the soluble polyglutamylated α-tubulin in the immunofluorescence experiments. Combining the immunoblot and immunofluorescence results suggests that PF7/CCDC40 is needed for polyglutamylation at the proximal end of the microtubule axonemes.
Next we asked whether paclitaxel has any effect on tpg1. As might be expected for flagella with normal length, neither tpg1 nor cnk11-1; tpg1 is affected by treatment with paclitaxel for 30 minutes (Fig 6B). However, no increase in flagellar length is observed after paclitaxel treatment of the pf7; tpg1 mutant. We suggest that paclitaxel does not increase flagellar length in strains with the cnk11 or tpg1 mutations.
Given that the NIMA-related kinase cnk2-1 mutant affects the disassembly rate of flagella, we asked whether cnk11-1 affects the rates of assembly and/or disassembly. We first compared the rates of flagellar assembly after flagella amputation by pH shock in wild-type (CC-125) and cnk11-1 cells (Fig 8A). Within 30 minutes following flagellar amputation, the assembly rate of cnk11-1 cells was ~0.23 μm/min, which is not significantly faster than the rate of wild-type cells (~0.20 μm/min). This is very similar to rates observed in the cnk2-1 cells by Hilton et al. [36]. However, the assembly rate in cnk11-1 cells reduced significantly within the next 90 minutes, and resulted in slightly shorter flagella than in wild-type cells (Fig 8A). We conclude that the overall assembly rate during flagellar regeneration is not affected in cnk11-1 cells.
Another way to measure the dynamic of flagellar assembly is to test the incorporation rate of new tubulin subunits at the flagellar tip. When a pair of Chlamydomonas cells mate, they form a quadriflagellate cell (QFC), which has two pairs of flagella. Tubulin subunits are added at the tip of the flagella, using subunits from the cytoplasm [19]. The two pairs of flagella can be distinguished by using one parent that carries an epitope-tagged HA-tubulin gene (Fig 8B insert, green), while the other parent lacks this gene. Both pairs of flagella are visualized with an antibody to acetylated α-tubulin (Fig 8B insert, magenta). The flagella from the parent with the tagged α-tubulin are stained with an antibody to the HA tag. Newly incorporated tubulin on the unlabeled flagella is visualized with the antibody to the HA tag. We mated two wild-type strains and tracked the incorporation of HA-tubulin subunits at 30, 60, 90 minutes after mating (Fig 8B, magenta). The length of incorporated HA-tubulin at the tip of flagella gradually increased along time and reached ~0.48 μm at 90 minutes. We mated two parents with the cnk11-1 mutation and observed more incorporation of HA-tubulin subunits at 60 and 90 minutes (Fig 8B, blue). The length of incorporated HA-tubulin at the tip was ~0.80 μm at 90 minutes, which suggests a rate that is nearly twice as fast as in wild-type QFCs. Since the length of the flagella did not increase, we suggest that the cnk11-1 mutation increases tubulin turnover at the flagellar tip.
Kubo et al. showed that the tpg2 mutant has slow tubulin turnover at the flagellar tip [27]. We performed the same assay on the tpg1 mutant. The incorporation length of HA-tubulin in tpg1 cells was ~0.16 μm at 90 minutes, significantly lower than that in wild-type or cnk11-1 cells (Fig 8B, purple). The incorporation length of HA-tubulin in cnk11-1; tpg1 cells was ~0.17 μm at 60 minutes but dropped to ~0.05 μm at 90 minutes (Fig 8B, black). Therefore, the faster turnover rate of HA-tubulin observed in cnk11-1 flagella is suppressed by the tpg1 mutation.
The addition of 1-isobutyl-3-methylxanthine (IBMX) causes gradual disassembly of flagella in wild-type cells. To ask whether flagellar disassembly is affected by cnk11, we compared the flagellar shortening rates in wild-type (CC-125) and cnk11-1 cells (Fig 8C). The disassembly rates of CC-125 and cnk11-1 cells within the first 30 minutes were both ~0.10 μm/min, similar to the rate Hilton et al. reported [36]. The disassembly rate of tpg1 (~0.11 μm/min) was similar to wild-type and cnk11-1. It was slightly reduced in cnk11-1; tpg1 (~0.09 μm/min, Fig 8C). Thus, neither cnk11-1 nor tpg1 mutation affects the flagellar disassembly rate.
Regulation of ciliary and flagellar length is extremely important to the proper function of these organelles in different organisms. In humans, ciliary length defects are observed in multiple ciliopathies that include Bardet-Biedl syndrome, nephronophthisis, Joubert syndrome, polycystic kidney disease, and Meckel-Gruber syndrome [66]. In Chlamydomonas, flagellar length defects affect both motility and mating [67]. Flagellar length can be regulated by multiple factors, including the rates of flagellar assembly and disassembly [66, 68], availability of IFT proteins, motors, and structure proteins [20], as well as factors that affect the stability of axonemal microtubules [69].
In mutant screens performed by McVittie and others, three pf7 and five pf8 alleles were identified [17, 70]. The mutants show abnormalities in the organization of the axoneme and radial spokes [17]. Oda and colleagues localized CCDC39 and CCDC40 using tagged genes together with cryo-EM tomography to show that these proteins serve as docking sites along the doublet microtubules for axonemal structures, which include the radial spokes, the N-DRC and all of the inner dynein arms [16]. Our immunoblots with DLE2/centrin suggest that not all of the inner arms are missing in pf7 and pf8 since centrin associates with three inner dynein heavy chains (b, e, and g) and is only slightly reduced. Although the pf7 and pf8 mutants have paralyzed flagella, their dyneins are functional based on our sliding/splaying assays. Both single mutants and the double mutant show splaying of the microtubules (Fig 4) that is similar to the splaying observed in the N-DRC mutants [10]. Thus, the paralysis is likely to be due to the microtubule and radial spoke disorganization that regulate the coordinated behavior of the dynein arms, as hypothesized by both McVittie and Oda et al. [16, 17]. The splaying experiments also suggest that the link in the proximal 2 μm does not rely on CCDC39/40, DRC1, DRC4, or the inner dynein arm I1/f (Fig 4 and [10]). Bui and colleagues [71] identified rod-like circumferential interdoublet linkers in the proximal axoneme that are clearly structural different from the N-DRC structure. We assume that these structures are retained in the pf7 and pf8 mutants, but they have not been examined.
In all patients diagnosed with PCD with CCDC39 or CCDC40 mutations, the changes result in premature truncation of the protein, which suggests that null alleles are associated with the phenotype [11]. Unexpectedly, the long-term prognosis of children with CCDC39 or CCDC40 mutations is worse than for other PCD patients, and similar to patients with cystic fibrosis [18]. These alleles would be similar to the mutations in pf7 and pf8 that have premature termination alleles. We also identified a conditional allele, fla12, in the PF8/CCDC39 locus (Table 1). The leucine to proline change occurs in an unstructured region of the C-terminus of the protein and the leucine is not conserved in other organisms. At the permissive temperature (21°C) the flagella are slightly shorter. This missense mutation leads to reduced CCDC39 and other DRC proteins at the restrictive temperature (Fig 3B). After 8 hours at the restrictive temperature, the phenotype of fla12 cells resembles the phenotype of pf8 cells. The flagella are immotile and short. It is possible that missense alleles in CCDC39 in humans may have a less severe phenotype that only slightly alters the motility and would not have been grouped together with the more severe null alleles associated with PCD [11, 12, 14]. Our fla12 revertants (Table 3) indicate that the leucine can be replaced by a variety of amino acids. This suggests that the change to a proline undermines the protein function and leads to the short and paralyzed flagella at 32°C. Given that the speeds of anterograde and retrograde IFT in fla12 shows no difference compared to those found in wild-type cells, it suggests that IFT is unlikely to play a role in flagellar length control in this CCDC39 mutant.
Post-translational modification of tubulin, which includes polyglutamylation and polyglycination, affects axonemal microtubule stability. Suryavanshi et al. and Kubo et al. showed in Tetrahymena and Chlamydomonas that loss of polyglutamylation on the B-tubule is likely to affect the activity of inner arm dyneins [72, 73]. A decrease in tubulin polyglutamylation in mouse airway cilia changes the curvature of the cilia as well as the asymmetry of the beating [74]. Overexpression of the polyglutamylation enzyme (TTLL6) in Tetrahymena destabilizes axonemal microtubules [75]. Knockdown of the glycination enzymes (TTLL3 and TTLL8) causes instability and results in short or absent mouse ependymal cilia. Polyglutamylation changes the binding affinities of a number of microtubule associated proteins and motors [76], and promotes microtubule severing [77]. Thus, the presence of polyglutamylation may affect microtubule stability in a variety of ways. Polyglutamylation like acetylation of tubulin is associated with long-lived microtubules [76, 78].
Because the loss of CCDC39 and CCDC40 affects the level of polyglutamylation, we examined the tpg1 mutation in TTLL9. Loss of α-tubulin polyglutamylation in tpg1 causes a motility defect due to the loss of tektin but no change in flagellar length ([29] and Fig 6B). Thus, reduction in tubulin polyglutamylation in the pf7 and pf8 mutants cannot be solely responsible for the short flagella in the CCDC39/40 mutants. The tpg1 mutation in combination with either pf7 or inner dynein arm deficient mutant pf23 has very different consequences. The tpg1 mutant partially rescues the flagellar length defect in pf23 (Fig 6A) but leads to more aflagellate cells with pf7 and no change in the length of the remaining flagella. By immunoblots, the level of polyglutamylated α-tubulin in the flagella of pf23; tpg1 is significantly less than in the flagella of pf7; tpg1 (Fig 7A). By immunofluorescence, we show that localization of polyglutamylated tubulin at the proximal end of axoneme is reduced in pf7; tpg1, but not in pf23; tpg1, when compared to tpg1 (Fig 7B & 7C). One possibility is that CCDC39 and CCDC40 are required for the activity of one or more tubulin glutamylases at the proximal end of the flagella while the TPG1 is responsible to polyglutamylation of tubulin along the rest of the flagella. There are 10 TTLL proteins found in Chlamydomonas [29] and the flagellar proteome includes only the TTLL9/TPG1 protein [79]. However, a proteomic analysis of flagellar phosphoproteins indicates that at least 3 additional TTLL proteins are found in the flagella [80]. They include TTLL13, a homolog of human tubulin polyglutamylase TTLL6; Cre09.g403108, an ortholog of human tubulin polyglutamylases TTLL4 and TTLL5; and Cre03.g145447, a homolog of human monoglycylase TTLL3. The former two are good candidates to be involved in tubulin glutamylation in the flagella.
Multiple protein kinases affect flagellar length. In Chlamydomonas, three CDK-like kinases (LF2, LF5, and FLS1), one MAP kinase (LF4), and one NIMA-related kinase (CNK2), have been characterized [36, 57, 81–83]. Loss of LF2, LF4, LF5, and CNK2 increase flagellar length. Loss of FLS1, CNK2, and LF4 block flagellar disassembly and loss of CNK2 decreases incorporation of new tubulin at the flagellar tip. The direct targets of these kinases remain to be identified. A recent global phosphoproteomic study revealed that over 180 Chlamydomonas flagellar proteins are phosphorylated [80]. This set includes N-DRC proteins, IFT proteins, outer and inner dynein arm proteins, central pair proteins, radial spoke proteins, and CCDC39.
Our screen for suppressors of the pf7; pf8 double mutant was designed to find swimming cells. However, no restoration of motility was found. The cnk11 mutant alleles led to short stumpy flagella, which still have a motility defect. In addition, we found that the cnk11 mutant alleles rescue the flagellar length defect but not the motility defect of N-DRC mutants as well as the dynein arm deficient mutants (Fig 6A). These results, along with the fact that multiple DRC proteins and axonemal proteins are not restored in pf7; pf8; cnk11-1 and pf3; cnk11-6 mutants (Fig 3A), suggest that the cnk11 mutations partially increase flagellar length via a N-DRC- and dynein protein-independent mechanism. Thus, even though CCDC39 is found to be phosphorylated in the flagella [80], it is unlikely that it is the direct target of CNK11.
During flagellar assembly, a cytoplasmic pool of tubulin subunits are constantly transported to the tip of flagella via IFT [59]. Different from flagellar assembly, flagellar disassembly is not dependent on flagellar length [19]. It is affected by the rates of IFT, disassembly of axonemal microtubules, and disassembly of axoneme-associated protein [81]. Unlike the lf2, lf4, lf5, and cnk2 mutants, both cnk11-1 and tpg1 mutants have normal flagellar length (Figs 1A and 6B and [29]), flagellar assembly (Fig 8A and [29]), and flagellar disassembly (Fig 8C). One difference between cnk11-1 and tpg1 mutants is that cnk11-1 shows a faster than normal tubulin turnover rate at the flagellar tip and tpg1 has a slower than normal rate (Fig 8B). The double mutant shows a turnover rate similar to tpg1 (Fig 8B) and proximal end localized polyglutamylated tubulin similar to tpg1 (Fig 7B & 7C). It is unlikely that TPG1 and TPG2 are the direct targets of CNK11, given that they are not found in the flagellar phosphoproteome [80].
In conclusion, we found that CCDC39 and CCDC40 mutants that have short flagella and fail to assemble the N-DRC and several inner dynein arms. Post-translational modification such as polyglutamylation and phosphorylation can affect flagellar length via IFT-independent and structural protein-independent pathways. These modification, may function similarly to the microtubule stabilizing drug paclitaxel and stabilize the unstable axonemal microtubules found in short flagella mutants. Further analysis of other short flagella mutants, such as shf1, shf2, and shf3 [84], or pf21 [85], are likely to identify more genes involved in flagellar assembly and length control.
The pf7 and pf8 mutant strains were obtained from the Chlamydomonas Resource Center as CC- 568 and CC-560. The strains from the stock center were aflagellate. Different media conditions were tried, but less than 20% of the cells assembled flagella. Both mutants were backcrossed to CC-124 three successive times to determine if reassorting the genome would increase the fraction of flagellated cells. After three backcrosses, strain pf8 2–4 was chosen. After 4 hours in nitrogen-free HSM medium, greater than 10% of cells had ~7 μm flagella. These were used for additional matings and for flagella preparations. After two backcrosses, strain pf7 2–2 was chosen. After 4 hours nitrogen-free medium, greater than 80% of cells had short (<4μm) flagella. Other strains and culture conditions are as reported previously [86]. Treatment of cells with paclitaxel was performed in yellow Lucite boxes to prevent breakdown of the paclitaxel by white light [64, 87].
Chlamydomonas genomic DNA preparation for whole genome sequencing was prepared as described previously [86]. Three micrograms of DNA was submitted to Genome Technology Access Core (Washington University School of Medicine) for library construction, Illumina sequencing, and initial data analysis. For multiplex Illumina sequencing, 7-nucleotide indexes were added to individual DNAs during the library construction before the samples were subjected to sequencing.
Illumina whole genome sequencing reads were aligned onto the Chlamydomonas version 5.3.1 genome assembly, and then aligned to JGI predicted exomes ([42] for details). SAMTools [88] were used for calling of SNPs/short indels from each strain. The SNPs/short indels from individual strains were compared to a SNP/short indel library (http://stormo.wustl.edu/dgranas/form.php) generated from 16 previously sequenced strains (15 in the cases of pf7 and pf8 because they were included as two of 16 original strains used to construct the library) [20, 42]. The unique SNPs/short indels in each strain were analyzed and filtered by SnpEff [89]. Only changes that have Phred quality scores of over 100 and rest within the coding region and splicing sites were retained. Whole genome sequencing reads of pf7 and pf8 can be found under NCBI BioProject Accession Number PRJNA245202. Whole genome sequencing reads of fla12, pf7; pf8; cnk11-1, pf7; pf8; cnk11-2, pf7; pf8; cnk11-3, pf7; pf8; cnk11-4, and pf7; pf8; cnk11-5 can be found under NCBI BioProject Accession Number PRJNA293107.
Revertant analysis was performed as previously reported [86]. Most of the mutant cells fail to oppose gravity and fall to the bottom of the tube. Swimming cells rise to the top of the tube and the upper 10 mL was transferred to fresh medium five times over the course of 13 days. Cultures were plated for individual colonies and one colony with swimming cells was kept from each tube. For the isolation of suppressors of the pf7; pf8 strains, we failed to recover any swimming cells. However, the nature of the pellet changed following the rounds of enrichment. Instead of large clumps of cells, the pellets were smooth and there were single cells. This was used to identify the suppressors.
Two day-old cells were resuspended in nitrogen-free medium for 4 hours before treated with freshly made autolysin and fixed in cold methanol. Cells were stained with anti-acetylated α-tubulin antibody followed by Alexa 594-conjugated goat anti-mouse secondary antibody. ImageJ was used to measure the flagellar length.
Protocols are as described previously [20]. Antibodies used in this study are listed in S4 Table.
Cells were deflagellated by pH shock and the isolated flagella were resuspended in demembranating buffer as described [10]. Half of the resultant axonemes were treated with 0.1 mM ATP at room temperature for 4 minutes. Both ATP-treated and non-treated axonemes were fixed with 2% paraformaldehyde at room temperature for 10 minutes on poly-lysine-coated multi-well slides (Thermo Scientific). The slide was then immersed in cold methanol for 10 minutes at -20°C. The samples were allowed to air dry on the slide before the addition of blocking buffer (5% BSA, 1% fish gelatin). The primary antibodies used were LF5 (1:200 dilution) and acetylated α-tubulin (1:250 dilution) diluted in 20% blocking buffer. The secondary antibodies were Alexa 488-conjugated goat-anti-rabbit IgG (1:500 dilution) and Alexa 594-conjugated goat-anti-mouse IgG (1:500 dilution) diluted in 20% blocking buffer.
Cells were imaged on manufacturer pre-cleaned fused silica chips (6W675-575 20C, Hoya Corporation USA, San Jose, CA), and sandwiched between the fused silica surface and a coverslip (1.8 x 1.8 cm2), resulting in a 25 μm thick water layer that allowed the 10 μm diameter Chlamydomonas cell body to move freely in solution. We used total internal reflection fluorescence (TIRF) microscopy to image the cells. The details of the imaging methods were reported previously [90]. Videos of individual surface-attached flagella were processed into kymographs. For visible IFT tracks in a kymograph, a minimum of 3 consecutive and clearly distinguishable IFT20::GFP intensity profiles were required for a track to be used. For each selected IFT track, the slope of the line through the centroid of the first and last IFT20::GFP intensity profiles in the track was used to determine the IFT velocity.
|
10.1371/journal.ppat.1003376 | A Unique Spumavirus Gag N-terminal Domain with Functional Properties of Orthoretroviral Matrix and Capsid | The Spumaretrovirinae, or foamyviruses (FVs) are complex retroviruses that infect many species of monkey and ape. Although FV infection is apparently benign, trans-species zoonosis is commonplace and has resulted in the isolation of the Prototypic Foamy Virus (PFV) from human sources and the potential for germ-line transmission. Despite little sequence homology, FV and orthoretroviral Gag proteins perform equivalent functions, including genome packaging, virion assembly, trafficking and membrane targeting. In addition, PFV Gag interacts with the FV Envelope (Env) protein to facilitate budding of infectious particles. Presently, there is a paucity of structural information with regards FVs and it is unclear how disparate FV and orthoretroviral Gag molecules share the same function. Therefore, in order to probe the functional overlap of FV and orthoretroviral Gag and learn more about FV egress and replication we have undertaken a structural, biophysical and virological study of PFV-Gag. We present the crystal structure of a dimeric amino terminal domain from PFV, Gag-NtD, both free and in complex with the leader peptide of PFV Env. The structure comprises a head domain together with a coiled coil that forms the dimer interface and despite the shared function it is entirely unrelated to either the capsid or matrix of Gag from other retroviruses. Furthermore, we present structural, biochemical and virological data that reveal the molecular details of the essential Gag-Env interaction and in addition we also examine the specificity of Trim5α restriction of PFV. These data provide the first information with regards to FV structural proteins and suggest a model for convergent evolution of gag genes where structurally unrelated molecules have become functionally equivalent.
| Foamyviruses (FVs) or spuma-retroviruses derive their name from the cytopathic effects they cause in cell culture. By contrast, infection in humans is benign and FVs have entered the human population through zoonosis from apes resulting in the emergence of Prototypic Foamyvirus (PFV). Like all retroviruses FVs contain gag, pol and env structural genes and replicate through reverse-transcription and host genome integration. Gag, the major structural protein, is required for genome packaging, virion assembly, trafficking and egress. However, although functionally equivalent, FV and orthoretroviral Gag share little sequence homology and it is unclear how they perform the same function. Therefore, to understand more about the relationship between FV and orthoretroviral replication we have carried out structural/virological studies of PFV Gag. We present the structure of Gag-NtD, a unique domain found only in FV Gag and show that despite functional equivalence, Gag-NtD is entirely structurally unrelated to orthoretroviral Gag. We also provide the molecular details of an essential interaction between Gag-NtD and the FV Envelope and demonstrate that Gag-NtD contains the determinants of Trim5α restriction. Our findings are discussed in terms of evolutionary convergence of retroviruses and the implications of alternative arrangements of Gag on pattern recognition by viral restriction factors.
| Spuma- or foamy viruses (FVs) are complex retroviruses and constitute the only members of the Spumaretrovirinae subfamily within the Retroviridae family. They have been isolated from a variety of primate hosts [1], [2], [3], [4] as well as from cats [5], [6], [7], cattle [8], horses [9] and sheep [10]. Endogenous FVs have also been described in sloth [11], aye-aye [12] and coelacanth [13]. Prototypic foamy virus (PFV) is a FV isolated from human sources [14], [15]. The PFV genome is highly similar to that of isolates of simian foamy virus from chimpanzee (SFVcpz) and so infection in humans is believed to have arisen through a zoonotic transmission [16], [17], [18]. Nevertheless, even though FVs are endemic within non-human primates and display a broad host range, human-to-human transmission of PFV has never been detected. Moreover, although in cell culture FV infection causes pronounced cytopathic effects [19] infection in humans is apparently asymptomatic [20], [21], [22] making their usage as vectors for gene therapy an attractive proposition [23].
FVs share many similarities with other retroviruses in respect of their genome organisation and life cycle. However, they vary from the Orthoretrovirinae in a number of important ways. These include the timing of reverse transcription that occurs in virus producer cells rather than newly infected cells [24], [25] and the absence of a Gag-Pol fusion protein [26] [27]. In addition, the Gag protein remains largely unprocessed in FVs [28] whereas within the Orthoretrovirinae processing of the Gag polyprotein represents a critical step in viral maturation, producing the internal structural proteins Matrix (MA), Capsid (CA) and Nucleocapsid (NC) found in mature virions. Furthermore, FV Gag lacks the Major Homology Region (MHR) and Cys-His boxes found in orthoretroviral CA and NC, respectively.
Also unique to FVs is a requirement for the interaction of the Gag protein with the viral envelope protein (Env) in order to bud from the producer cell [29], [30], [31]. Nevertheless, despite these profound dissimilarities, the Gag protein contains the cytoplasmic targeting and retention signal (CTRS) [32], [33], [34], essential for both FV and betaretrovirus replication. Moreover, in all retroviral subfamilies Gag carries out the same functional roles including assembly, nucleic acid packaging, transport to and budding through the cytoplasmic membrane of the producer cell as well as trafficking through the cytoplasm of the target cell and uncoating. Similarly, FV Gag also contains the determinants for restriction by Trim5α [35], [36] that in orthoretroviruses are the residues displayed on the assembled CA lattice [37].
To date, high resolution X-ray and/or NMR structures have been reported for MA, CA and NC components of Gag from numerous retroviruses [38], [39], [40], [41], [42], [43], [44], [45], [46], [47]. However, structural information with regard to the Gag of FVs has remained elusive and is vital for any detailed understanding of how FV Gag fulfils its many functions. Here we report the crystal structure an amino terminal domain from the Gag of PFV (PFV-Gag-NtD), provide the molecular details of the interaction of this domain with the N-terminal leader sequence from the PFV Envelope (PFV-Env) and demonstrate that the PFV-Gag-NtD is also the target for Trim5α restriction factors. Our data reveal that the FV Gag is unique and structurally unrelated to the Gag protein of other retroviruses. Nevertheless, the Gag-NtD has functional properties associated with both the MA and CA proteins of the orthoretroviruses. These findings have important implications for the evolution of FVs and the mechanism of virus restriction by Trim5α.
PFV Gag is a 648 polypeptide and the major FV structural protein in the assembled virion. Bioinformatic analysis of the primary sequence and that of related FVs suggested that the N-terminal 179 residues of PFV-Gag comprised a stable domain (PFV-Gag-NtD). This fragment was expressed in E. coli and subsequently the crystal structure determined by SAD methods and refined at a resolution of 2.4 Å. The final Rwork/Rfree are 17.2% and 23.0% respectively. Details of the structure solution and refinement are presented in Table 1. In the crystal, the asymmetric unit comprises a dimer of the protein with residues 9–179 clearly visible in the electron density map for both monomers. The residues preceding Glu9 along with the N-terminal His-tag are not visible and presumably disordered in the crystal. The structure of PFV-Gag-NtD dimer, Figure 1A, comprises a mixed alpha-beta fold dominated by a large central coiled-coil, resembling a two-bladed propeller. The N-terminal region of the protein forms a head domain containing a central 4-stranded β-sheet together with two helices, α1 and α2, that pack against one side of the sheet forming a tight hydrophobic core. The loop between strands β3 and β4 crosses to the opposing side of the sheet where helix α2 leads into a region lacking secondary structure that precedes three further short helices α3, α4 and α5. Helix α5 is immediately followed by α6, a long helix (58 Å) comprising residues Arg140-Ser179 that forms the coiled coil domain making the majority of interactions between the two monomers. The observation of this unusual arrangement prompted us to examine the structural relatedness of the PFV-Gag-NtD with the Gag-derived proteins from other retroviruses and those of hepadnavirus. However, similarity searches undertaken using the DALI [48] and SSM [49] search engines revealed no significant homology between the PFV-Gag-NtD and the retroviral MA or CA. In fact, no homology was detected with any structure deposited in the PDB making the foamy virus Gag-NtD at present unique.
The PFV-Gag-NtD dimer interface buries approximately 1700 Å2 of the monomer surface area. The large central coiled-coil formed by helix α6 comprises the majority of this interface, supplemented by residues in helices α4 and α5 and the adjoining loop. The coiled-coil contains three regions of leucine zipper at residues Leu143/147, Leu160/161 and Leu171/175. Additionally, a highly synergistic hydrogen-bonding network centred on residue Glu154 is located between two of the zipper regions. Here, the Glu154 sidechain forms hydrogen bonds with the sidechains of Glu154*, Gln150* and Tyr127* of the opposing monomer. Gln150 makes further hydrogen bonds with Ser151* that in turn is hydrogen bonded to the mainchain carbonyl of Val130, Figure 1B. The loop between helix α4 and α5 runs alongside this region also making several interactions. In addition, at the amino terminus of the coiled-coil, Arg140 makes bifurcated hydrogen bonds with the backbone hydroxyl of Met138 and the sidechain hydroxyl of Glu144 as well as further hydrogen bonds with the backbone of Ala136 in helix α4 and the side chain of Asp141* in the opposing monomer. This extensive network of intermolecular protein-protein interactions and large molecular interface of 1700 Å2 is nearly twice that of the HIV-1 CA-CtD dimer interface, 920 Å2 [50], [51] and suggests that Gag-NtD of PFV forms a tightly associated dimer. Moreover, sequence alignment with the N-terminal region of Gag from other primate foamy viruses, Figure 1C, reveals strong sequence conservation in loops and secondary structure elements in the head domains together with several buried hydrophobic residues in the coiled coil indicating that a conserved dimeric Gag N-terminal domain is likely a feature of the primate foamy viruses.
Given the unexpected nature of the dimer observed in the crystal structure, the conformation and self-association properties of the Gag-NtD from PFV and from the non-primate feline foamy virus (FFV) were examined using a variety of solution hydrodynamic methods. Initial assessment by Size Exclusion Chromatography coupled Multi-Angle Laser Light Scattering (SEC-MALLS), over a range of protein concentration (12-1.5 mg/ml), yielded invariant solution molecular weights of 40.0 kDa and 34.0 kDa for PFV- and FFV-Gag-NtD respectively, Figure 2A. By comparison, the sequence-derived molecular weights are 22.8 kDa and 19.0 kDa. Given these values together with the lack of a concentration dependency of the molecular weight it is apparent that along with PFV, the Gag-NtD from FFV also forms strong dimers in solution. To confirm the oligomeric state, velocity (SV-AUC) and equilibrium (SE-AUC) analytical ultracentrifugation of PFV- and FFV-Gag-NtD was undertaken. A summary of the experimental parameters, molecular weights derived from the data and statistics relating to the quality of fits are shown in Table 2. Analysis of the sedimentation velocity data for PFV-Gag-NtD revealed no concentration dependency of the sedimentation coefficient (S20,w = 3.08) over the range measured, Figure 2B. Similar data were obtained for FFV-Gag-NtD (S20,w = 2.72) indicating both proteins are single stable species. The molecular weights derived from either C(S) or discrete component analysis were 47 kD and 36 kD respectively, Table 2, consistent with a PFV- and FFV-Gag-NtD dimer. The frictional ratios (f/fo) obtained from the analysis, 1.4–1.5, also suggest both dimers have a similar elongated conformation. Multispeed sedimentation equilibrium studies at varying initial protein concentration were also carried out and typical equilibrium distributions for PFV- and FFV-Gag-NtD from individual multispeed experiments are presented in Figure 2C. Analysis of individual gradient profiles showed no concentration dependency of the molecular weight and so data were fit globally with a single ideal molecular species model, producing weight averaged molecular weights of 44 kDa and 33.7 kDa for PFV- and FFV-Gag-NtD respectively. These data confirm that formation of stable dimeric structures is a common property shared among the Gag proteins of divergent FVs and N-terminal domain mediated dimerisation is likely an important component of FV assembly.
The interaction of foamy virus Gag and Env proteins is a requirement for successful budding and the production of infectious particles [52]. Mutations in either Gag-NtD or the N-terminal leader peptide region of Env (Env-LP) have been shown to block viral egress [31], [53], [54]. To better understand this interaction and shed light on how FV Gags recruit Env, we examined the interaction of the PFV-Gag-NtD with the PFV-Env-LP using SV-AUC. Sedimentation data were recorded for Gag-NtD and for equimolar mixtures of Gag-NtD with either of two Env leader peptides, residues 5–18 or 1–20, Figure 3A. The data were fitted using the C(S) distribution of sedimentation coefficients and the integrated absorbance of the fast moving Gag-NtD 3S component then quantified. In samples containing peptide-protein mixtures a small increase in the apparent sedimentation coefficient of the 3S boundary is apparent, accompanied by an increase in the integrated absorbance, Figure 3B. This shift and absorbance increase results from association of the strongly absorbing Env peptides with the PFV-Gag-NtD (ε280 = 11,400 M−1 cm−1) and simple quantitation of the absorbance change reports the proportion of peptide bound and association constant for the interaction (see methods). In this way an equilibrium association constant of 2.0×104 M−1 for the Gag-NtD interaction with Env residues 5–18 (Env5–18) and 1.3×105 M−1 for the interaction with Env residues 1–20 (Env1–20) was determined. To confirm this observation the interaction of Env1–20 with PFV-Gag-NtD was examined using isothermal titration calorimetry (ITC). The results presented in Figure 3C reveal a 1∶1 stoichiometry where each monomer of the PFV-Gag-NtD binds a single Env peptide with an equilibrium association constant of 1.5×105 M−1 consistent with the SV-AUC experiments.
The structure of the PFV-Gag-NtD bound to the Env1–20 leader peptide was determined by molecular replacement and refined at a resolution of 2.9 Å with a final Rwork/Rfree of 22.6% and 27.1% respectively, Table 1. The asymmetric unit comprises two dimers of the complex with residues 9–179 of the Gag-NtD clearly visible in the electron density map for two of the four monomers and residues 9–170 in the two remaining protomers. Four helical Env peptides are also present, bound at the periphery of each head domain close to α1 and the associated α1-β1 loop of the Gag-NtD monomers, Figure 4A. Largely, the conformation of the Gag-NtD head and stalk domains are the same as in the free structure (RMSD of 0.4 Å between all equivalent Cα atoms) excepting some small differences in the conformation of the β3–β4 loop, Supplementary Figure S1. However, in the bound structure the α1-β1 loop around the highly conserved residue Pro30 undergoes a concerted 2.5 Å shift, Figure 4B and Supplementary Figure S1. Comparison of surface hydrophobicity profiles of the free and bound structures Figure 4B, reveals that this movement opens the Env binding site exposing a deep apolar pocket to accommodate the hydrophobic side chains from the Env peptide. In the complex residues Met1Env to Thr6Env of Env constitute an extended N-terminal region and Leu7Env to Met16Env form the hydrophobic α-helix bound to Gag. Hydrogen bonding between the sidechains of Thr6Env in the N-terminal region and Gln9Env on the amino-terminal turn of the Env helix provides stabilising interactions that maintain the helical conformation of the Env, Figure 4C. Inspection of Gag-Env interface reveals a network of hydrophobic interactions with the apolar and aromatic sidechains of Leu7Env, Trp10Env and Trp13Env on one face of the Env helix packing against the Val14, Leu17, Val18 and Leu21 sidechains on α1 of Gag, Figure 4C. In particular, the side chain of Leu7Env is seated in the apolar pocket in the Gag-NtD were it makes hydrophobic interactions with the aliphatic side chains of both Leu17 and Leu21. Val14 packs against the ring of Trp10Env that also makes a hydrogen bonding interaction between the indole Nε proton and the carbonyl of Leu66 in the β2–β3 loop. This hydrophobic interface is accompanied by a number of polar contacts between the backbone of residues Ala2Env, Pro3Env and Met5Env in the Env N-terminal extended region with the sidechains of Asn63 and Gln59 in the β2–β3 loop and the mainchain of His32 and Pro30 in the α1-β1 loop. The bound conformation is further stabilised by an accompanying helix capping interaction between the Asn29 sidechain and the N-terminal turn of the Env helix.
In order to probe the importance of the interactions in the Gag-Env interface observed in the crystal structure a series of serine and asparagine substitutions were introduced at Val14, Leu17, Val18 and Leu21 to make the Env binding site progressively polar. In addition, in order to examine the contribution of the α1-β1 loop to the Env interaction a conservative Asn29 to Gln substitution was also introduced. The affinity of binding of these Gag-NtD mutants to Env-LP was examined using the sedimentation velocity assay, Figure 5A–E and Supplementary Figure S2. In all cases, the single polar substitutions introduced into the Env binding site reduced the affinity of the Gag-Env interaction. The decrease varied from 5 – 2 fold in the order Leu21>Val14>Leu17≈Val18 identifying these residues as being required for the Gag-Env interaction. Double substitutions decreased the affinity even further with the Val14/Leu21 to serine having the greatest effect, resulting in around a twenty-fold reduction in binding, Figure 5F. Moreover, the triple substitution where Val14, Val18 and Leu21 were all substituted by serine reduced binding to an undetectable level, Supplementary Figure S2. The conservative change Asn29 to glutamine has little effect on Env binding perhaps reflecting the importance of the backbone movement around Pro30 rather than sidechain interactions for Env-binding at this position.
It has been shown previously that mutation of Leu17 in PFV-Gag-NtD gives rise to viral defects and negatively affects viral egress. Substitution by alanine has only minor effects on Env incorporation and particle release but progeny particles show a severe reduction of the infectivity. In contrast, serine substitution results in a loss of viral budding capacity [53]. To assess in vivo effects of serine substitution at Leu17 and at other positions in the Gag-Env interface the Leu17, Val14 and Leu21 to Ser mutations that disrupt the Gag-Env interaction in vitro were introduced and transfected cells assayed for particle production as well as Env/Gag incorporation and viral infectivity, Figure 6. In these in vivo experiments, the greatest effects were seen with Leu17 and Leu17/Leu21 mutant viruses that show greatly reduced levels of Gag released into the supernatant compared to wild type. By contrast, only a small reduction in Gag release was observed in the Val14 virus and in the Leu21 virus the amount of Gag is comparable to wt, Figure 6A. Examination of Env production and processing in the producer cells reveals it is unaffected by any of the mutations, Figure 6B. However, Env incorporation into virions is greatly reduced in both the Leu17 and Leu17/Leu21 particles, moderately reduced in the Val14 virus and that near wt levels are present in Leu21 particles. These results are mirrored when particle release was quantified, Figure 6C, where Leu21 particle production is only slightly reduced, Val14 is reduced around 3-fold, Leu17 around 20-fold and in the double substitution no particles are detectable in the cell supernatant. Where particles were released they were tested for infectivity relative to wt, Figure 6D. Although the Val14 mutant showed greater defects in viral Env and Gag incorporation the viruses were only around 6-fold reduced in infectivity whereas viruses with the Leu21 substitution showed around a 300 fold reduction in infectivity and no infectivity (>100,00-fold reduced) was detectable for the Leu17 mutant. Taking these data together it is apparent that the Leu17 mutation has the least effect on in vitro Env binding but causes very large defects in PFV virion production with little, if any, incorporation of viral proteins into particles. The Leu21 substitution weakens the in vitro Gag-Env interaction more, has little effect on particle production but the resulting viruses are poorly infectious and viruses with the Val14 substitution display intermediate effects having both reduced particle production and reduced infectivity.
Previous experiments have demonstrated that Gag from PFV and the closely related SFVmac contain the target for Trim5α restriction. Moreover, PFV and SFVmac display a differential susceptibility to restriction mediated by the B30.2 domain of Brown capuchin Trim5α (bc-T5α) that is effective only against SFVmac and not PFV [36]. Based on sequence alignment, chimeras were prepared to more precisely map the target of Trim5α restriction in FV Gag. These included PSG-4 and SPG-4, that swap the N-terminal ∼300 residues between PFV and SFVmac Gag and two further chimeras, one where the N-terminal 186 residues of SFVmac Gag was replaced by the N-terminal 195 residues of PFV Gag (PSG-5) and a second where the N-terminal 195 residues of PFV Gag was replaced with the N-terminal 186 residues of SFVmac Gag (SPG-5). The results of bc-T5α restriction assays of parent and chimeric PFV and SFVmac viruses are summarised in Figure 7A, and detailed in Supplementary Figure S3. These data confirm that PFV is resistant to bc-T5α restriction and that SFVmac is susceptible, and that sensitivity maps the N-terminal 300 amino acids of Gag, Figure 7A (PSG-4 and SPG-4). More importantly, these data also reveal that transfer of the N-terminal 186 residues of SFVmac to PFV (SPG-5) now renders the virus susceptible to restriction by bc-T5α. Conversely, transfer of N-terminal 195 residues of PFV to SFVmac (PSG-5) results in reduced sensitivity to bc-T5α restriction demonstrating that at least one determinant of restriction in primate FVs is contained within the Gag-NtD.
Since NtD s of PFV and SFVmac Gag share a high degree of sequence similarity, the conserved residues along with those involved in the dimer interface were mapped onto the PFV-Gag-NtD structure, Figure 7B. Examination of this combined pattern of sequence conservation and surface accessibility reveals a large patch of surface exposed non-conserved residues on the upper surface of the molecule spanning from the β2–β3 loop across the outer surface of α2 and into the α2–α3 loop. The distribution of non-conserved residues over the top surface of the molecule is reminiscent of the distribution of residues that constitute the restriction factor binding sites in the N-terminal domain of the capsid of conventional retroviruses [55]. This suggests that the mode of foamy virus restriction by Trim5α is likely to be the same as in orthoretroviruses. In order to test this notion, mutations were introduced into the RING, B-Box and coiled coil domains of bc-T5α and the restriction of PFV and SFVmac by these impaired factors assayed. These data, summarised in Table 3 and detailed in Supplementary Figure S4, show that disruption of the individual RING and B-Box domains or deletion of the coiled-coil region completely abolishes bc-T5α restriction of SFVmac and does not alter PFV susceptibility. Taken together with data demonstrating that the B30.2 domain of Trim5α mediates the Gag specificity of restriction [36] this demonstrates that FV restriction is reliant on the same functional regions required for orthoretrovirus restriction and likely occurs by the same mechanism.
Based upon both the functional similarities and positioning within PFV Gag it might be expected that the Gag-NtD would display a strong structural similarity with MA of orthoretroviruses. However, following extensive searching of the Protein Database (PDB) no such similarity was apparent and in fact no structures related to PFV-Gag-NtD were found at all. Like FV-Gag-NtD, the orthoretroviral MA protein is required for targeting Gag to the membrane and for viral budding. This is accomplished through a combination of a highly basic region (HBR) and in some subfamilies a myristoyl group located at the N-terminus of MA [56], [57], [58]. However, although the MA functional properties are conserved, neither of these motifs is present in the PFV-Gag-NtD. Further, the structure of MA is highly conserved amongst retroviruses, consisting of a four α-helix globular core and an associated fifth helix [59], [60], [61], [62], [63], [64], [65]. By comparison, our data reveals the PFV-Gag-NtD to be entirely unrelated comprising a mixed α/β protein with head and stalk domains.
The dimeric organisation of FV-Gag-NtD is also not a conserved feature of orthoretroviral MAs. In HIV, myristoyl-MA promotes assembly and budding directly at the plasma membrane (PM) [56] and although it is unclear what the MA oligomerisation state is within immature and mature virions trimeric assemblies have been reported in vitro [61], [63]. In the betaretroviruses that like FVs assemble intracellularly at the pericentriolar region [32], [33], only weak self-association of MA has been demonstrated [66]. By contrast, in the delta-retrovirus HTLV-1 the presence of stable disulphide linked dimers of Gag and MA in both immature particles and mature virions has been observed [67]. Thus, although FV-Gag-NtD and orthoretroviral MA have membrane-targeting roles in the late part of the viral life cycle, the differences in structure and organisation suggests the existence of different evolutionary pathways.
Evidence for this notion also comes from sequence comparisons of the predicted Gag protein from FVs ranging from primate to sloth revealing they all share the same motifs and that they are unrelated to orthoretroviral Gag [11]. This implies there is one evolutionary pathway for the FVs with a single Gag protein and another for the orthoretroviruses in which the Gag precursor protein undergoes significant processing. Moreover, based on the observation of endogenous foamy virus in coelacanths, this divergence occurred more than 400 million years ago [13].
Foamy virus replication also has similarities with that of hepadnaviruses, including reverse transcription in the producer cell and an infectious DNA genome in the virion [24], [25]. As there is no apparent structural homology with orthoretroviral Gag one possibility is that FV Gag may be related to a hepadnavirus structural protein. Inspection of capsid protein of hepadnavirus B (Hep-B) [68] reveals that Hep-B CA is an all-helical protein with a prominent 4-helix bundle making up the interface between CA dimers. This arrangement is reminiscent of the coiled-coil dimer interface of the PFV-Gag, However, in Hep-B the 4-helix bundle forms “spikes” that protrude from the exterior of the capsid shell. Given the arrangement of FV Gag with the N-terminal MA layer found at the greatest radius and the more C-terminal regions of Gag projecting to the virion interior [29] it seems unlikely that FV Gag is related to hepadnavirus CA. This further supports the notion that FV Gag-NtD is the product of convergent evolution that has driven the formation of a unique structure with properties of orthoretroviral MA and CA.
The cytoplasmic targeting and retention signal (CTRS) found in the MA of betaretroviruses and in the Gag-NtD of FVs, promotes assembly in the pericentriolar region of the cell [32], [33], [34]. The consensus sequence in betaretroviruses spans residues Pro43 to Gly60 in MA of the archetypal betaretrovirus Mason-Pfizer monkey virus (MPMV) [69], [70]. Within this sequence the majority of residues, Pro43 to Ile53, constitute the loop that links helix α2 to helix α3 of MA whilst the remainder make up the first two turns of α3 [65]. In FVs the proposed CTRS constitutes residues 43 to 60 of the PFV-Gag-NtD [34] where residues Leu40 to Arg50 form the loop that links β1 to β2 and the remainder make up the β2 strand. Although the betaretroviral and FV CTRSs appear largely dissimilar, one common feature of both is a double aromatic motif G43WWGQ47 in PFV and P43WFPQ47 in MPMV. In both cases the sequences are located in the loop regions of the CTRS and comprise a structural motif consisting of a tight turn and a surface exposed aromatic and glutamine side chain, Figure 8. In MPMV, mutation of the CTRS causes Gag to traffic as a monomer to the plasma membrane where assembly and production of infectious virus still occurs [70]. By comparison, absence of a functional CTRS in FVs completely abrogates assembly and whilst addition of a myristoylation signal to PFV facilitates Gag trafficking to the plasma membrane, infectious particles are not produced [54], [71]. The severest effects on capsid formation and particle production were observed when alanine substitution mutations were introduced at Trp45 or Arg50 in the CTRS of PFV [53]. However, examination of the Gag-NtD structure now reveals that although Trp45 and Arg50 are part of the CTRS both are actually deeply buried in the core of the head domain. Arg50 also forms a number of important hydrogen bonds with neighbouring residues stabilising the interaction of the head domain with helix α5 immediately preceding the coiled-coil. Therefore, in these instances the severe mutational effects associated with alanine substitution can be likely attributed to destabilisation and/or misfolding of the Gag-NtD. However, mutation of the surface exposed Trp44 in the double aromatic motif does allow particle assembly but with a large reduction in both particle export and infectivity (∼105 fold) [53]. In this case, given the exposure of the Trp44 sidechain, Figure 8, the lack of particle egress might be attributed directly to loss of a di-hydrophobic motif dependent CTRS function causing mislocalisation or incorrect trafficking of assembled virions.
FV egress requires interaction between the Gag and Env proteins to ensure correct membrane trafficking and viral budding. It seems likely that FV Gag becomes associated with Env through interaction with Env leader peptide (Env-LP) displayed on the cytosolic side of the ER and Trans-Golgi network (TGN) after core assembly at the pericentriolar region [31], [33]. Env then directs the intracellular transport of the assembled particles to enable mature viruses to bud at the PM or sometimes into intracellular vacuoles. This interaction guarantees Env incorporation into virions and the loss of either interacting domain (Gag-NtD or Env-LP) results in the intracellular stranding of assembled FV capsids [30], [52].
Mutations in the Env binding site of PFV-Gag-NtD have been shown to affect viral assembly, egress and infectivity. Of note is Leu17 that when substituted by serine results in loss of virus production, Figure 6A. However, our in vitro binding data, Figure 5, reveal only modest reductions in affinity (2–5 fold) when single serine substitutions are introduced into the Env binding site suggesting that the Leu17 to serine mutation may have effects prior to Gag-Env association. This notion is further supported by the fact that the mutant displays a phenotype similar to that of the Trp45 and Arg50 alanine mutations that disrupt the CTRS [53]. Examination of virion production and Env incorporation in Val14 and Leu21 serine substitution mutants reveals reduced levels in Val14 serine mutant but near wild type amounts in Leu21 particles. The small effects on virus production observed with the Leu21 and Val14 mutations also correlate well with the modest reductions in KA observed with the single-site mutations. This likely reflects the situation that recruitment of Env by a preassembled FV core rather than by Gag monomers is subject to the avidity effects of having many Gag binding sites arrayed on the core surface. Therefore, even under conditions of reduced binding the cores can still recruit enough Env to bud efficiently. However, whilst the effects on particle number and Gag-Env interaction are small, the Val14 and Leu21 serine mutations result in reduced infectivity, similar to when Leu17 is replaced by alanine [53], suggesting that disruption of the Gag-Env interaction may also be detrimental for post-entry events in the target cells.
In the structure, residues 7–16 of the Env leader peptide comprise the amphipathic α-helix bound in the Env binding site of the Gag-NtD and residues 1–6 provide intramolecular hydrogen bonding that stabilises the helical conformation. The affinity of the interaction, 1.5×105 M−1, is comparable with the value of 0.65×105 M−1 reported for the interaction of residues 1–30 of the FFV Env leader peptide with the equivalent Gag-NtD [29]. Therefore, the hydrophobic interface observed in the structure likely represents the complete interaction between the leader peptide and FV Gag. The apolar character of the Env binding site is largely conserved among primate FVs although there is significant variation in the primary sequences of the α1 helix, Figure 1C. By contrast, the sequences of the N-terminal 13 residues of the Env leader peptide are largely invariant giving rise to the conserved motif [M-A-P-P-M-(T/S/N)-L-(E/Q)-Q-W-Φ-Φ-W] where Φ denotes a residue with a hydrophobic side chain. Our binding data show that removal of the first 4 residues (MAPP) along with Ala19 and His20, not visible in the crystal structure, results in a significant reduction in Gag-Env binding, Figure 3. It has also been demonstrated previously that the N-terminal four residues as well as the conserved tryptophan residues Trp10ENV and Trp13ENV, are essential for PFV egress [31]. Moreover, mutation of the equivalent conserved tryptophans in FFV greatly reduces the Gag-Env interaction in vitro [29]. The necessity for the N-terminal five residues is now apparent from the Gag-Env complex structure as many of the residues in the N-terminal extended region make polar contacts with Gag but also make intramolecular interactions with the Env helix to stabilise the conformation that binds to the Gag. The importance of the tryptophans is also apparent as they form part of the hydrophobic interface with Gag. Given the degree of conservation in the N-terminal of Env it is likely that this mode of interaction is a common feature of the Env-LP interaction with the Gag-NtD in other FVs.
In orthoretroviruses, the viral core is enclosed by a hexameric lattice of CA assembled through combined homotypic and heterotypic interactions mediated by the amino-terminal (CA-NtD) and carboxy-terminal (CA-CtD) domains of CA [39], [51], [72], [73]. In FVs, the structural organisation of the core is less characterised but two regions of FV-Gag required for assembly have been identified. Reminiscent of orthoretroviral CA-NtD and CA-CtD, the first corresponds to the Gag-NtD coiled-coil dimer defined in our structural studies [74] (Figure 1) and the other found in the central region of FV-Gag (Gag-CtD) includes a conserved YXXLGL assembly motif [75]. In all likelihood the interior structural organisation of the FV virion is also formed by combinatorial heterotypic and homotypic protein-protein interactions mediated by these assembly domains, although the requirement for other regions of Gag, not yet identified, cannot be excluded. A further functional similarity of FV Gag-NtD and orthoretroviral CA-NtD is that both appear to be the target of Trim5α, mediated restriction Figure 7, [36], [37] and in orthoretroviruses, it is proposed that underlying hexagonal pattern of the assembled CA is recognised by a complementary hexagonal assembly of Trim5α in order to initiate the restriction process [76], [77]. Presently, the overall arrangement of the Gag protein in an assembled FV is unknown but since the same species dependent Trim5α restriction of PFV and other FVs is apparent [35], [36] the requirement for a lattice structure that arrays FV-Gag-NtD on the exterior of the FV core might also be expected. One possibility is that FV-Gag-NtD dimerisation combined with FV-Gag-CtD interactions generates a higher-order hexagonal Gag assembly targeted by Trim5α factors. However, given the obligate nature of the FV Gag-NtD dimer together with its organisation, dimensions and lack of structural homology with orthoretroviral CA it is difficult to envisage how a hexagonal assembly of equivalent spacing to that of the orthoretroviruses might be present in the FV particle. These observations raise the question of whether Trim5α might target other regular, or even irregular, molecular arrangements in addition to the hexagonal assemblies. Current models rely on a rather rigid overlapping of the orthoretroviral CA and Trim5α supramolecular assemblies. The inclusion of FVs in the cadre of Trim5α targets suggests there is potential flexibility in the pattern recognition receptor activity of Trim5α. Determining how this is accomplished awaits further structural and microscopic studies of the FV virion.
Human HT1080 [78] and 293T [79] cells were maintained in Dulbecco modified Eagle medium supplemented with 10% foetal calf serum and 1% penicillin and streptomycin. Restriction factors were delivered into cells using Moloney MLV (MoMLV)-based vectors produced by transfection of 293T cells. MoMLV-based delivery vectors were made by co-transfection of VSVG, pHIT60, and pLgatewayIRESEYFP containing the restriction gene. FVs were produced by a four-plasmid PFV vector co-transfection system [80], [81] in which pciSFV-1env (providing Env), pcziPol PFV vector (providing Pol), pMD9 (a minimal vector genome with an EGFP marker gene), and a Gag-expressing construct were co-transfected. FV vector supernatants were harvested 48 h post-transfection, aliquoted, and stored at −80°C until further use. Subsequently, individual vector supernatant aliquots were pre-titrated on HT1080 cells using the EGFP marker gene and flow cytometric analysis. For the two-colour restriction assay described below, FV vector supernatants were then used at dilutions that resulted in 3 to 40% EGFP-positive HT1080 cells.
A chimeric TRIM5α with the RBCC domain of human TRIM5α and the PRYSPRY domain of brown capuchin, referred to here as capuchin TRIM5α because the PRYSPRY domain determines restriction specificity, has been described previously [82]. A series of mutants of this factor in RING (C15A/C18A), B-box 2 (C95A/H98A, W115E and E118K/R119K) and coiled-coil (delta 130–231) were prepared by site directed mutagenesis.
Preparation of the PFV and SFV packaging plasmids, PFV pcziGag4 (PGWT) and SFVmac pcziSG (SGWT) respectively, as well as chimeric PFV/SFV Gag packaging constructs, PSG-4 and SPG-4, has been described previously [36]. The novel chimeric constructs PSG-5 and SPG-5 were generated by recombinant overlap PCR starting with PGWT and SGWT. PSG-5 contains amino acids 1–195 of PFV and 187–647 of SFV while SPG-5 encodes amino acids 1–186 of SFV and 196–648 of PFV.
PFV Gag point mutants were generated in context of a the original PFV Gag packaging construct pcziGag4 [80] (L17S/L21S), or a C-terminally HA-tagged variant thereof, pcziPG CLHH (V14S, L17S, L21S).
Restriction was determined by our previously described two-colour fluorescence activated cell sorter (FACS) assay [83]. Briefly, HT1080 cells were transduced with the MLV-based pLgatewayIRESEYFP retroviral vector carrying the restriction gene and an EYFP marker gene 2 days prior to challenging with FVs carrying the EGFP marker. The percentage of YFP positive cells (i.e. restriction factor-positive cells) that were EGFP positive (i.e. FV infected) was then determined by FACS. This was compared to the percentage of FV-infected cells (EGFP positive) in cells that did not express the restriction factor (EYFP negative). A ratio that was less than 0.3 was taken to represent restriction, while a ratio greater than 0.7 indicated the absence of restriction.
Cell culture supernatants containing recombinant viral particles were generated as described previously [84]. Briefly, 293T cells were co-transfected in 10 cm dishes with a Gag expression plasmid (pcziGag4 or PG mutants thereof, as indicated), Env (pcoPE), Pol (pcoPP), and the transfer vector (puc2MD9) at a ratio of 16∶1∶2∶16 using Polyethyleneimine (PEI) reagent and 16 µg DNA total. At 48 h post transfection (p.t.) cell-free viral vector supernatant was harvested using 0.45 µm sterile filters.
For transduction efficiency analysis 2×104 HT1080 cells were plated in 12-well plates 24 h before infection. The target cells were incubated with 1 ml of plain cell-free viral supernatant or serial dilutions thereof for four to six hours. Determination of the percentage of eGFP-expressing cells was performed 72 h after infection by flow cytometry analysis and used for titre determination as previously described [85]. All transduction experiments were repeated at least three times. To compare the infectivity in repetitive experiments the titre obtained for wild type supernatants in individual experiments was set to an arbitrary value of 100%. The other values were then normalized as percentage of the wild type value.
Viral protein expression in transfected cells and particle-associated protein composition was examined by Western blot analysis. Preparation of cell lysates from one transfected 10-cm cell culture dish was performed by incubation with 0.6 ml lysis buffer for 20 min at 4°C followed by centrifugation through a QIAshredder (Qiagen). All protein samples were mixed with equal volumes of 2×PPPC (100 mM Tris-HCl; pH 6.8, 24% glycerol, 8% SDS, 0.2% Bromophenol blue, 2% ß-mercaptoethanol) prior to separation by SDS-PAGE using 7.5% polyacrylamide gels. Viral particles were concentrated from cell-free supernatant of transfected 293T cells by ultracentrifugation through a 20% sucrose cushion at 4°C and 25,000 rpm for 3 h in an SW32 rotor. The viral pellet was resuspended in phosphate-buffered saline (PBS). Immunoblotting using polyclonal antisera specific for PFV Gag [86] or PFV Env leader peptide [75] was performed as previously described [31]. The chemiluminescence signal was digitally recorded using a LAS3000 imager and quantified using ImageGauge in the linear-range of the sample signal intensities as described previously [87].
The DNA sequences coding for PFV-Gag residues 1–179 (PFV-Gag-NtD) and FFV residues 1–154 (FFV-Gag-NtD) were amplified by PCR from template plasmids pcziGag4 and pcDWF003 containing the PFV and FFV Gag genes respectively. PCR products were inserted into a pET47b expression vector (Novagen) using ligation independent cloning in order to produce N-terminal His-tag fusions with 3C protease cleavage sites. The correct sequence of expression constructs was verified by automated DNA sequencing (Beckman Coulter Genomics). His-tagged PFV- and FFV-Gag-NtD were expressed in the E. coli strain Rosetta 2 (DE3) and purified using Ni-NTA affinity (Qiagen) and size exclusion chromatography on Superdex 200 (GE healthcare). Selenium was incorporated into PFV-Gag-NtD by replacement of methionine with seleno-methionine in defined culture medium and by inhibition of methionine biosynthesis just prior to IPTG induction [88]. Verification of the processed N-terminal methionine, correct molecular mass and degree of selenium incorporation was obtained by electrospray ionisation mass-spectrometry. Peptides comprising residues 1–20 and 5–18 from the PFV-Env leader region were purchased HPLC purified from Pepceuticals Ltd.
Size exclusion chromatography coupled multi-angle laser light scattering (SEC-MALLS) was used to determine the molar mass of FFV- and PFV-Gag-Ntd. Samples ranging from 1.5 to 12.0 mgml−1 were applied in a volume of 100 µl to a Superdex 200 10/300 GL column equilibrated in 20 mM Tris-HCl, 150 mM NaCl and 0.5 mM TCEP, pH 8.0, at a flow rate of 0.5 ml/min. The scattered light intensity and the protein concentration of the column eluate were recorded using a DAWN-HELEOS laser photometer and OPTILAB-rEX differential refractometer respectively. The weight-averaged molecular mass of material contained in chromatographic peaks was determined from the combined data from both detectors using the ASTRA software version 6.0.3 (Wyatt Technology Corp., Santa Barbara, CA, USA).
Sedimentation velocity experiments were performed in a Beckman Optima Xl-I analytical ultracentrifuge using conventional aluminium double sector centrepieces and sapphire windows. Solvent density and the protein partial specific volumes were determined as described [89]. Prior to centrifugation, samples were prepared by exhaustive dialysis against the buffer blank solution, 20 mM Tris-HCl, 150 mM NaCl and 0.5 mM TCEP, pH 7.5. Centrifugation was performed at 50,000 rpm and 293 K in an An50-Ti rotor. Interference data were acquired at time intervals of 180 sec at varying sample concentration (0.5–2.5 mg/ml). Data recorded from moving boundaries was analysed using both a discrete species model and in terms of the size distribution functions C(S) using the program SEDFIT [90], [91], [92]. For analysis of Env peptide binding, sedimentation velocity experiments were conducted in 3 mm pathlength centrepieces using equimolar mixtures (75 µM) of PFV-Gag-NtD and Env peptides. In these experiments, radial absorbance scans at 280 nm were also recorded along with the interference data.
Sedimentation equilibrium experiments were performed in a Beckman Optima XL-I analytical ultracentrifuge using charcoal filled Epon six-channel centrepieces in an An-50 Ti rotor. Prior to centrifugation, samples were dialyzed exhaustively against the buffer blank, 20 mM Tris-HCl, pH 7.5, 150 mM NaCl, 0.5 mM TCEP. After centrifugation for 18 h, interference data was collected 2-h intervals until no further change in the profiles was observed. The rotor speed was then increased and the procedure repeated. Data were collected on samples of different concentrations of FFV- and PFV-Gag-NtD (14–100 µM) at three speeds and the program SEDPHAT [93], [94] was used to determine weight-averaged molecular masses by nonlinear fitting of individual multi-speed equilibrium profiles (A versus r) to a single-species ideal solution model. Inspection of these data revealed that the molecular masses showed no significant concentration dependency and so global fitting incorporating the data from multiple speeds and multiple sample concentrations was applied to extract a final weight-averaged molecular mass.
Data were analysed using a general binding expression, Eq. 1. This expression relates the association constant Ka to the fraction of bound peptide θ (θ = [PL]/[Lt]) in terms of the total concentrations of peptide [Lt] and protein [Pt] and is a modification of the formulae employed in [95], [96].(1)In sedimentation velocity experiments θ was determined from the integrated absorbance of the 3S species in the C(S) function that best fits the sedimentation data. As equimolar ratios of peptide and protein were employed ([Lt] = [Pt]) Eq. 1 can be simplified and equilibrium association constants determined from Eq. 2.(2)
ITC was carried out using an ITC-200 calorimeter (MicroCal). Briefly, PFV-Gag-NtD was prepared by were dialysis against 25 mM Na-phosphate pH 6.55, 100 mM NaCl, 0.5 mM TCEP. A typical experiment involved 20 injections of 1 mM Env peptide in the injection syringe into 50 µM PFV-Gag-NtD in the sample cell. Data was analysed using the Origin-based software provided by the manufacturers.
PFV-Gag-NtD was crystallised using hanging drop vapour diffusion. Typically, A 10 mg/ml solution of PFV-Gag-NtD in 150 mM NaCl, 5% glycerol, 10 mM Tris-HCl, pH 8.0 was mixed with an equal volume of crystallisation solution containing 16% PEG 6000 (w/v), 12% ethylene glycol, 0.03 M MgCl2 hexahydrate and suspended over a reservoir of the crystallisation solution. Crystals appeared within 14 days and were transferred into fresh crystallisation solution supplemented with 20% glycerol and flash-frozen in liquid nitrogen. The crystals belong to the space group P21 with one copy of the PFV dimer in the asymmetric unit. Seleno-methionine derived protein was crystallized under the same conditions. Crystals of the PFV-Gag-NtD-Env complex were also grown by vapour diffusion by mixing 500 µM 1∶1 complex in 150 mM NaCl, 5% glycerol, 10 mM Tris-HCl, pH 8.0 with an equal volume a crystallisation solution containing 10% PEG 4000 (w/v), 20% glycerol, 0.03 M MgCl2, 0.03 M CaCl2, 0.1 M Tris-Bicine pH 8.5. Crystals appeared within 2 days and were harvested into fresh crystallisation solution supplemented with 20% glycerol and flash-frozen in liquid nitrogen prior to data collection. Crystals of the complex also belong to the space group P21 but with two copies of the PFV dimer-peptide complex in the asymmetric unit. The structure of PFV-Gag-NtD was solved by single wavelength anomalous diffraction (SAD) using a dataset recorded at 0.9791 Å at 100 K on beamline I03 at the Diamond Light Source (Didcot, UK) using crystals of the seleno-methionine substituted protein. Data was processed using the HKL program package [97] and 13 selenium atoms were located by SAD methods in PHENIX [98]. Further density modification in PHENIX resulted in a figure of merit of 0.79 and a map of sufficient quality for a near complete model to be built using Arp/Warp [99]. The model was completed by iterative rounds of refinement and model building in PHENIX and COOT [100]. TLS groups were included in final round of refinement as determined by TLSMD [101]. The structure was refined to a final Rwork/Rfree of 17.2/23.0 respectively and has good geometry with 98.8% of residues in the preferred region of the Ramachandran plot, only 1.2% in the additionally allowed region and no outliers. Details of crystal parameters and data refinement statistics are presented in Table 1. Data for the PFV-Gag-NtD-Env complex was collected at 100 K on beamline I03 and processed and scaled in space group P21 using XDS/XSCALE [102]. The structure was solved by molecular replacement using Phaser [103] with the Gag-NtD dimer used as a search model to locate the two copies of the complex in the asymmetric unit. The model was completed by iterative rounds of TLS based refinement and model building using Refmac5 [104] and COOT. TLS groups were defined using TLSMD. The structure was refined to a final Rwork/Rfree of 22.6/27.1 in which 98.8% of residues lie within preferred regions of the Ramachandran plot and the remaining 1.2% residues lie within the additionally allowed region. The crystal and refinement parameters are given in Table 1. The coordinates and structure factors of PFV-Gag-NtD and PFV-Gag-NtD-Env complex have been deposited in the Protein Data Bank under accession numbers 4JNH and 4JMR respectively.
|
10.1371/journal.pgen.1004293 | Drosophila Embryogenesis Scales Uniformly across Temperature in Developmentally Diverse Species | Temperature affects both the timing and outcome of animal development, but the detailed effects of temperature on the progress of early development have been poorly characterized. To determine the impact of temperature on the order and timing of events during Drosophila melanogaster embryogenesis, we used time-lapse imaging to track the progress of embryos from shortly after egg laying through hatching at seven precisely maintained temperatures between 17.5°C and 32.5°C. We employed a combination of automated and manual annotation to determine when 36 milestones occurred in each embryo. D. melanogaster embryogenesis takes 33 hours at 17.5°C, and accelerates with increasing temperature to a low of 16 hours at 27.5°C, above which embryogenesis slows slightly. Remarkably, while the total time of embryogenesis varies over two fold, the relative timing of events from cellularization through hatching is constant across temperatures. To further explore the relationship between temperature and embryogenesis, we expanded our analysis to cover ten additional Drosophila species of varying climatic origins. Six of these species, like D. melanogaster, are of tropical origin, and embryogenesis time at different temperatures was similar for them all. D. mojavensis, a sub-tropical fly, develops slower than the tropical species at lower temperatures, while D. virilis, a temperate fly, exhibits slower development at all temperatures. The alpine sister species D. persimilis and D. pseudoobscura develop as rapidly as tropical flies at cooler temperatures, but exhibit diminished acceleration above 22.5°C and have drastically slowed development by 30°C. Despite ranging from 13 hours for D. erecta at 30°C to 46 hours for D. virilis at 17.5°C, the relative timing of events from cellularization through hatching is constant across all species and temperatures examined here, suggesting the existence of a previously unrecognized timer controlling the progress of embryogenesis that has been tuned by natural selection as each species diverges.
| Temperature profoundly impacts the rate of development of “cold-blooded” animals, which proceeds far faster when it is warm. There is, however, no universal relationship. Closely related species can develop at markedly different speeds at the same temperature. This creates a major challenge when comparing development among species, as it is unclear whether they should be compared at the same temperature or under different conditions to maintain the same developmental rate. Facing this challenge while working with flies (Drosophila species), we found there was little data to inform this decision. So, using time-lapse imaging, precise temperature-control, and computational and manual video-analysis, we tracked the complex process of embryogenesis in 11 species at seven different temperatures. There was over a three-fold difference in developmental rate between the fastest species at its fastest temperature and the slowest species at its slowest temperature. However, our finding that the timing of events within development all scaled uniformly across species and temperatures astonished us. This is good news for developmental biologists, since we can induce species to develop nearly identically by growing them at different temperatures. But it also means flies must possess some unknown clock-like molecular mechanism driving embryogenesis forward.
| It has long been known that Drosophila, like most poikilotherms, develops faster at higher temperatures, with embryonic [1], larval [1], [2], and pupal stages [3], [4], as well as total lifespan [5], [6] showing similar logarithmic trends. While genetics, ecology, and evolution of this trait have been investigated for over a century [2], [7]–[17], the effects of temperature on the order and relative timing of developmental events, especially within embryogenesis, are poorly understood.
We became interested in the relationship between species, temperature, and the cadence of embryogenesis for practical reasons. Several years ago, we initiated experiments looking at the genome-wide binding of transcription factors in the embryos of divergent Drosophila species: D. melanogaster, D. pseudoobscura, and D. virilis. With transcription factor binding a highly dynamic process, we tried to match both the conditions (especially temperature, which we believed would affect transcription factor binding) in which embryos were collected and the developmental stages we analyzed. However, our initial attempts to collect D. pseudoobscura embryos at 25°C — the temperature at which we collect D. melanogaster — were unsuccessful, with large numbers of embryos failing to develop, likely a consequence of D. pseudoobscura's alpine origin. While D. virilis lays readily at 25°C, we found that their embryos develop more slowly than D. melanogaster, complicating the collection of developmental stage-matched samples.
Having encountered such challenges with just three species, and planning to expand to many more, we were faced with several important questions. Given that embryogenesis occurs at different rates in different species [8], [18], how should we time collections to get the same mix of stages we get from our standard 2.5–3.5 hour collections in D. melanogaster, or any other stage we study in the future? Is it better to compare embryos collected at the same temperature even if it is not optimal for, or even excludes, some species; or, should we collect embryos from each species at their optimal temperature, if such a thing exists? Should we select a temperature for each species so that they all develop with a similar velocity? Or should we find a set of species that develop at the same speed at a common temperature? And even if we could match the overall rate of development, would heterochronic effects mean that we could not get an identical mix of stages?
We found a woeful lack in the kind of data needed to answer these questions. Powsner precisely measured the effect of temperature on the total duration of embryogenesis in D. melanogaster [1], and Markow made similar measurements for other Drosophila species at a fixed temperature (24°C) [18], but the precise timing of events within embryogenesis had been described only for D. melanogaster at 25°C [19], [20].
The work described here was born to address this deficiency. We used a combination of precise temperature control, time-lapse imaging, and careful annotation to catalog the effects of a wide range of temperatures on embryonic development in 11 Drosophila species from diverse climates. We focused on species with published genome sequences [21] (Table 1), as these are now preferentially used for comparative and evolutionary studies. Of the species we studied D. melanogaster, D. ananassae, D. erecta, D. sechellia, D. simulans, D. willistoni, and D. yakuba are all native to the tropics, though D. melanogaster, D. ananassae, and D. simulans have spread recently to become increasingly cosmopolitan [17]. D. mojavensis is a sub-tropical species, while D. virilis is a temperate species that has become holarctic and D. persimilis and D. pseudoobscura are alpine species (Figure 1A).
We used automated, time-lapse imaging to track the development of embryos held at a constant and precise temperature from early embryogenesis (pre-cellularization) to hatching. We maintained the temperature at 0.1°C using thermoelectric Peltier heat pumps. Different sets of embryos were analyzed at temperatures ranging from 17.5°C to 32.5°C, in 2.5°C increments. Images were taken every one to five minutes, depending on the total time of development. A minimum of four embryos from each species were imaged for each temperature, for a total of 77 conditions. In total, time-lapse image series were collected and analyzed from over 1000 individual embryos.
We encountered, and solved, several challenges in designing the experimental setup, including providing the embryos with sufficient oxygen [22], [23] and humidity. We found that glass slides were problematic due to a lack of oxygenation and led to a 28% increase in developmental time, so we instead employed an oxygen-permeable tissue culture membrane, mounted on a copper plate to maintain thermal conduction. At higher temperatures, we found that the embryos dehydrated, so humidifiers were used to increase ambient humidity. Detailed photos of the apparatus and descriptions can be found in Figure S1.
We used a series of simple computational transformations (implemented in Matlab) to orient each embryo, correct for shifting focus, and adjust the brightness and contrast of the images, creating a time-lapse movie for each embryo. We manually examined images from 60 time-lapse series in D. melanogaster and identified 36 distinct developmental stages [19], [20] that could be recognized in our movies (Table 2, http://www.youtube.com/watch?v=dYSrXK3o86I and http://www.youtube.com/watch?v=QKVmRy3dDR0 or “D. melanogaster with labelled stages” and “D. melanogaster with labelled stages at reduced framerate” in DOI:10.5061/dryad.s0p50”). Due to the volume of images collected, we implemented a semi-automated system to annotate our entire movie collection. Briefly, images from matching stages in manually annotated D. melanogaster movies were averaged to generate composite reference images for each stage (Figure 2). We then used a Matlab script to find the image-matrix correlation between each of these composite reference images to the images in each time-lapse to estimate the timing of each morphological stage via the local correlation maximum (Figure S2A).
Of the 36 events, the eight most unambiguous events (Figure S3), identifiable regardless of embryo orientation, were selected for refinement and further analysis (pole bud appears, membrane reaches yolk, pole cell invagination, amnioproctodeal invagination, amnioserosa exposed, clypeolabrum retracts, heart-shaped midgut, and trachea fill) (Figure S2B,C). Using a Python-scripted graphical user interface, each of the eight events in every movie was manually examined and the algorithm prediction adjusted when necessary. Timing of hatching was excluded from these nine primary events because it was highly variable, likely due to the assay conditions following dechorionation, and suitable only as an indication of successful development, not as a reliable and reproducible time point. The “membrane reaches yolk stage” was used throughout as a zero point due to the precision with which the stage could be identified in all species and from all orientations.
Links to representative time-lapse videos are provided in Table 3.
As expected, the total time of embryogenesis of D. melanogaster had a very strong dependence on temperature (Figure 3, http://www.youtube.com/watch?v=-yrs4DcFFF0 or “D. melanogaster at 7 temperatures” in DOI:10.5061/dryad.s0p50). From 17.5°C to 27.5°C, there was a two-fold acceleration in developmental rate, matching the previously observed doubling of total lifespan with a 10°C change in temperature [6]. The velocity of embryogenesis at 30°C is roughly the same as at 27.5°C, and is appreciably slower at 32.5°C, likely due to heat stress. At 35°C, successful development becomes extremely rare.
To examine how these temperature-induced shifts in the total time of embryogenesis were reflected in the relative timing of individual events, we rescaled the time series data for each embryo so that the time from our most reliable early landmark (the end of cellularization) to our most reliable late landmark (trachea filling) was identical, and examined where each of the remaining landmarks fell (Figure 3C). We were surprised to find that D. melanogaster exhibited no major changes in its proportional developmental time under any of the non-stressful temperature conditions tested. Therefore, at least as far as most visually evident morphological features go, embryogenesis scales uniformly across a two-fold range of total time.
When the embryos were under heat stress (30°C), we observed a very slight contraction in the proportion of time between early development (pole bud appears) to the end of cellularization (membrane reaches yolk), and a slight contraction between the end of cellularization and mid-germ band retraction (amnioserosa exposure).
In each of the ten additional Drosophila species we examined we observed all of the 36 developmental landmarks we identified in D. melanogaster in the same temporal order (Figure 4A). However, there was marked interspecies variation in both the total time of embryogenesis at a given temperature (Figure 4B–E, Table 3) and the way embryogenesis time varied with temperature (Figure 5).
When we examined the 10 remaining species, we found not only that the relative timing of events was constant across temperature within a species, as observed in D. melanogaster, but that landmarks occurred at the same relative time between species at all non-stressful temperatures (Figures 6, Table 4).
Between 17.5°C and 27.5°C the total developmental time for all species can be approximated relatively accurately by an exponential regression (). For all species we find that temperature T can be related to developmental time , agreeing with a long history of temperature-dependent rate modeling [24]:and developmental rate v:The parameters of these relations for each species, which includes two independent coefficients, are included in Table 5. Also included in Table 5 is the , an empirical description of biological rate change from a 10°C temperature change, for the 17.5°C to 27.5°C interval. At higher temperatures, heat stress appears to counter the logarithmic trend and lengthens developmental time. Since the temperature responses are highly reproducible, the developmental time for each species can be modeled and predictions made for future experiments (Figure S4).
Seven of the eleven species we examined were of tropical origin, with only two alpine, one subtropical and one temperature species. At mid-range temperatures (22.5°C–27.5°C), the tropical species developed the fastest, followed by the subtropical D. mojavensis, the alpine D. pseudoobscura and D. persimilis, and the temperate D. virilis (Figure 5), in accord with [18].
Some tropical species have expanded into temperature zones and a variety of wild strains have been collected from a variety of climates. We examined nine additional strains of D. melanogaster collected along the eastern United States [25], [26]. Though collected along a tropical to temperate cline and there was some variation between strains, no trends were seen (Figure S5A,B).
The tropical species all showed highly similar responses to temperature, even though they originate from different continents (Africa, Asia and South America) and are not closely related (five of the species are in the melanogaster subgroup, but D. ananassae and D. willistoni are highly diverged from both D. melanogaster and each other). Though they possess similar temperature-responses, these species possess significantly different and independent temperature response curves () and the differences are large enough to be relevant for precise developmental experiments. These cross-species differences tend to be, but are not necessarily, larger than those seen between D. melanogaster strains (Figure S5C). The embryogenesis rate for these species increases rapidly with temperature () before slowing down at and above 30°C (Figure S6A–F, http://www.youtube.com/watch?v=vy6L4fmWkso or “D. ananassae at 7 temperatures” in DOI:10.5061/dryad.s0p50). The two closely related alpine species (D. pseudoobscura and D. persimilis) match the embryogenesis rate of the tropical species at 17.5°C, but accelerate far less rapidly with increasing temperature (), especially at 25°C and above (Figure S6I,J, http://www.youtube.com/watch?v=sYi-FUXpv4Q or “D. pseudoobscura at 6 temperatures” in DOI:10.5061/dryad.s0p50). These species also show a sharp increase in embryogenesis rate and low viability above 27.5°C, consistent with their cooler habitat. The subtropical D. mojavensis (Figure S6H, http://www.youtube.com/watch?v=XWMs4oUx_mU or “D. mojavensis at 6 temperatures” in DOI:10.5061/dryad.s0p50) and temperate D. virilis (Figure S6G, http://www.youtube.com/watch?v=eyr4ckDb0kM or “D. virilis at 6 temperatures” in DOI:10.5061/dryad.s0p50) both develop very slowly at low temperature, but accelerate rapidly as temperature increases ( of and respectively). D. virilis remains the slowest species up to 30°C, while D. mojavensis is as fast as the tropical species at high temperatures. These species are both members of the virilis-repleta radiation and it remains to be seen if this growth response is characteristic of the group as a whole, independent of climate.
Under heat-stress, the proportionality of development is disrupted in some embryos (Figure S7A). The effect is not uniform, as some embryos developed proportionally under heat-stress and others exhibited significant aberrations, largely focused in post-germband shortening stages. This can be most clearly seen in individuals of D. ananassae, D. mojavensis, D. persimilis, and D. pseudoobscura. We did not identify any particular stage as causing this delay, but rather it appears to reflect a uniform slowing of development.
Early heat shock significantly disrupts development enough to noticeably affect morphology in yolk contraction, cellularization, and gastrulation (Figure S7B). Syncytial animals are the most sensitive to heat-shock (Figure S7C). In D. melanogaster and several other species we observed a slight contraction of proportional developmental time between early development (pole bud appears) and the end of cellularization (membrane reaches yolk) under heat-stress (30°C, Figure S7D). While all later stages following cellularization maintain their proportionality even at very high temperatures, the pre-cellularization stages take proportionally less and less time. This indicates that at higher temperatures, some pre-cellularization kinetics scale independently of later stages, possibly leading to mortality as the temperature becomes more extreme.
We have addressed the lack of good data on the progress of embryogenesis in different species and at different temperatures with a carefully collected and annoted series of time-lapse movies in 11 species at seven temperatures that span most of the viable range for Drosophila species. From a practical standpoint, the predictable response of each species to temperature, and the uniform scaling of events between species and temperature, provides a relatively simple answer to the question that motivated this study - to determine how to obtain matched samples for genomic studies: simply choose the range of stages to collect in one strain or species, and scale the collection and aging times appropriately. The fact that development scales uniformly over non-extreme temperatures would seem to give some leeway in the choice of temperature, so long as heat-stress is avoided, though it remains unclear how molecular processes are affected by temperature.
In carrying out this survey, we were surprised to find that the relative timing of landmark events in Drosophila embryogenesis is constant across greater than three-fold changes in total time, spanning 15°C and over 100 million years of independent evolution. And the fact that the same holds true for 34 developmental landmarks at two temperatures in the zebrafish Danio rerio [27], (the only other species for which we were able to locate similar data), suggests that this phenomenon may have some generality. But why is this so?
Drosophila development involves a diverse set of cellular processes including proliferation, growth, apoptosis, migration, polarization, differentiation, and tissue formation. One might expect (we certainly did) these different processes to scale independently with temperature, much as different chemical reactions do, and as a result, different stages of embryogenesis or parts of the developing embryo would scale differentially with temperature. But this is not the case. The simplest explanation for this observation is that a single shared mechanism controls timing across embryogenesis throughout the genus Drosophila. But what could such a mechanism be? One possibility is that there is an actual clock — some molecule or set of molecules whose abundance or activity progresses in a clocklike manner across embryogenesis and is read out to trigger the myriad different processes that occur in the transition from a fertilized egg to a larvae. However there is no direct evidence that such a clock exists (although we note that there is a pulse of ecdysone during embryogenesis with possible morphological functions [28], [29]). A more likely explanation is that there is a common rate limiting process throughout embryogenesis. Our data are largely silent on what this could be, but we know from other experiments that it is cell, or at least locally, autonomous [30]–[32] and would have to limit processes like migration that do not require cell division (we also note that cell division has been excluded as a possibility in zebrafish [32]). However, energy production, yolk utilization, transcription or protein synthesis are reasonable possibilities.
Although there are very few comparisons of the relative timing of events during development, it has long been noted that various measurements of developmental timing scale exponentially with [1], [5], [6], [24], [33], but no good explanation for this phenomenon has been uncovered. Perhaps development is more generally limited by something that scales exponentially with , like metabolic rate, which, we note, has been implicated numerous times in lifespan, which is, in some ways, a measure of developmental timing.
Gillooly and co-workers, noting the there was a relationship between metabolic rate, temperature and animal size, have proposed a model that incorporates mass into the Arrhenius equation to explain the relationship between these factors in species from across the tree of life [34], [35]. We, however, do not find that mass can explain the differences in temperature-dependence between species. Even closely-related species, with nearly 2-fold differences in their mass (e.g. D. melanogaster, D. simulans, D. sechellia, D. yakuba, and D. erecta), have significant divergence in their proportionality coefficients that do not converge at all when correcting for differences in mass through the one quarter power scaling proposed by Gillooly, et al. This suggests that some other factor is responsible for the differences, as has been argued by other groups [18], [36], [37]. The relationship between climate and temperature response raises the possibility that whatever this factor is has been subject to selection to tune the temperature response to each species' climate. However, without additional data this is purely a hypothesis.
Although a common rate-limiting step is simplest explanation for uniform scaling, it is certainly not the only one. It is possible that different rate limiting steps or other processes control developmental velocity at different times or in different parts of the embryo, and that they scale identically with temperature either coincidentally, or as the result of selection (it is important to remember that, as per Arrhenius, one does not expect different reactions to scale identically with temperature). If this is the result of selection, what is the selection pressure? Evolutionary developmental biologists, perhaps most notably Stephen J. Gould, have long written about how changes in either the absolute or relative timing of different events during development have had significant effects on morphology throughout animal evolution [38]–[41]. Perhaps this is also true for fly embryogenesis, but that any such changes in morphology are selectively disadvantageous and have been strongly selected against. It is also likely that many developing fly embryos experience significant changes in temperature while developing, so there may be strong selection to maintain uniform development across temperature to ensure normal progression while the temperature is changing.
Finally, we note that there are limits to this uniformity. At extreme temperatures, especially high ones, things no longer scale uniformly, likely reflecting the differential negative effects of high temperature at different stages of embryogenesis as well as the differential ability of the embryo to compensate for them. There are also clearly checkpoints in place that, while not triggered during normal embryogenesis, are important in extreme or unusual circumstances. Most strikingly, when Lucchetta et al. and Niemuth et al. examined embryos developing in chambers that allowed for independent temperature control of the anterior and posterior portions of the embryo, the two parts of the embryo developed at different velocities for much of embryogenesis [30], [31]. They found that embryos are robust to asynchrony in timing across the embryo, though there are critical periods that, once passed, do not permit re-synchronization of development [30], hinting at some specific checkpoints or feedback.
The clustering of developmental timing and its temperature response with climate — especially amongst tropical species from different continents and parts of the Drosophila tree — suggests that this is an adaptive, or in some cases permissive, phenotype, although with only 11 species and poor coverage of non-tropical species this has to remain highly speculative. There are necessarily additional components to the temperature response, as significant variation exists within the tropical species and between D. melanogaster strains. The virilis-repleta radiation, which includes both D. virilis and D. mojavensis may have a climate-independent adaptation that leads to slowed development at cooler temperatures, a feature that is hard to rationalize. The poor response of the alpine D. pseudoobscura and D. persimilis to high temperature is consistent with their cool climate. Nevertheless, little is known about when and where most of these species lay their eggs and their natural microclimates.
The clustering of developmental responses in species by their native climates rather than their climates of collection suggests that if climate adaptation is a contributing factor, the response arises slowly or rarely. The tested strains of D. melanogaster were collected in temperate, subtropical, and tropical climates and the D. simulans strain was collected in a sub-tropical climate. Nevertheless, both species performed qualitatively like other tropical species and unlike native species collected nearby. This suggests that temperature responses are neither rapidly evolving (with D. melanogaster being present in the temperate United States for over 130 years [42]) nor primed for change in tropical species.
Drosophila strains were reared and maintained on standard fly media at 25°C, except for D. persimilis and D. pseudoobscura which were reared and maintained at 22°C. D. melanogaster lines were raised at 18°C and 22°C for several years and their temperature response profiles were observed, verifying that transferring embryos from the ambient growth temperature for a line to the experimental temperature did not lead to heat-shock responses and had relatively little impact on the temperature response (Figure S8A,B). Egg-lays were performed in medium cages on 10 cm molasses plates for 1 hour at 25°C after pre-clearing for all species except D. persimilis, which layed at 22°C. Comparisons to D. melanogaster raised and laying at 22°C confirmed that growth at lower temperatures does not account for all of the differences between the tropical and alpine species (Figure S8C).To encourage egg-lay, cornmeal food media was added to plates for D. sechellia and pickled cactus was added to plates for D. mojavensis. Embryos were collected and dechorionated with fresh 50% bleach solution (3% hypochlorite final) for 45 to 90 seconds (based on the species) in preparation for imaging. Dechorionation timing was selected as the time it took for 90% of the eggs to be successfully dechorionated. This prevented excess bleaching, as many species, such as D. mojavensis, are more sensitive than D. melanogaster. Strains used were D. melanogaster, OreR, DGRP R303, DGRP R324, DGRP R379, DGRP R380, DGRP R437, DGRP R705, Schmidt Ln6-3, Schmidt 12BME10-24, and Schmidt 13FSP11-5; D. pseudoobscura, 14011-0121.94, MV2-25; D. virilis, 15010-1051.87, McAllister V46; D. yakuba, 14021-0261.01, Begun Tai18E2; D. persimilis, 14011-0111.49,(Machado) MSH3; D. simulans, 14021-0251.195, (Begun) simw501; D. erecta, 14021-0224.01, (TSC); D. mojavensis wrigleyi, 15081-1352.22, (Reed) CI 12 IB-4 g8; D. sechellia, 14021-0248.25, (Jones) Robertson 3C; D. willistoni, 14030-0811.24, Powell Gd-H4-1; D. ananassae, 14024-0371.13, Matsuda (AABBg1).
Embryos were placed on oxygen-permeable film (lumox, Greiner Bio-one), affixed with dried heptane glue and then covered with Halocarbon 700 oil (Sigma) [43]. The lumox film was suspended on a copper plate that was temperature-regulated with two peltier plates controlled by an H-bridge temperature controller (McShane Inc., 5R7-570) with a thermistor feedback, accurate to 0.1°C. Time-lapse imaging with bright field transmitted light was performed on a Leica M205 FA dissecting microscope with a Leica DFC310 FX camera using the Leica Advanced Imaging Software (LAS AF) platform. Greyscale images were saved from pre-cellularization to hatch. Images were saved every one to five minutes, depending on the temperature. A humidifier was used to mitigate fluctuations in ambient humidity, though fluctuations did not affect developmental rate. Due to fluctuations in ambient temperature and humidity, the focal plane through the halocarbon oil varied significantly. Therefore, z-stacks were generated for each time-lapse and the most in-focus plane at each time was computationally determined for each image using an algorithm (implemented in Matlab) through image autocorrelation [44], [45]. Time-lapse videos available from Dryad Digital Repository: doi:10.5061/dryad.s0p50.
A subset of time-lapses in D. melanogaster were analyzed to obtain a series of representative images for each of the 36 morphological events, selected as all events defined by [19], [46] that were reproducibly identifiable under our conditions, described. These images were sorted based on embryo orientation and superimposed to generate composite reference images. Images from each time-lapse to be analyzed were manually screened to determine the time when the membrane reaches the yolk, the time of trachea filling, and the orientation of the embryo (Figure S3. This information was fed into a Matlab script, along with the time-lapse images and the set of 34 composite reference images, to estimate the time of 34 morphological events during embryogenesis via image correlation. The same D. melanogaster reference images were used for all species for consistency. A correlation score was generated for each frame of the time-lapse. The running score was then smoothed (Savitzky-Golay smoothing filter) and the expected time window was analyzed for local maxima. The error in event calling for the computer is very large (greater than what we see for the overall spread across individuals of a single species at a given temperature), necessitating manual verification or correction of events. Many of these errors are due to aberrations in the image that confuse the computer but would not confuse a person. This results in a few bad images having a very negative effect of the overall accuracy of the computer analysis, but permits a significant improvement with just a little user input. The error in manual calls is very small compared to the variation between individuals. Computer-aided estimates were individually verified or corrected using a python GUI for all included data.
Statistical significance of event timing was determined by t-test with Bonferonni multiple testing corrections. Median correction to remove outliers was used in determining the mean and standard deviation of each developmental event. Least-squares fitting was used to determine the linear approximation of log-corrected developmental time for each species. Python and Matlab scripts used in the data analysis are available at github.com/sgkuntz/TimeLapseCode.git.
|
10.1371/journal.pgen.1008326 | Methyl-CpG-binding domain 9 (MBD9) is required for H2A.Z incorporation into chromatin at a subset of H2A.Z-enriched regions in the Arabidopsis genome | The SWR1 chromatin remodeling complex, which deposits the histone variant H2A.Z into nucleosomes, has been well characterized in yeast and animals, but its composition in plants has remained uncertain. We used the conserved SWR1 subunit ACTIN RELATED PROTEIN 6 (ARP6) as bait in tandem affinity purification experiments to isolate associated proteins from Arabidopsis thaliana. We identified all 11 subunits found in yeast SWR1 and the homologous mammalian SRCAP complexes, demonstrating that this complex is conserved in plants. We also identified several additional proteins not previously associated with SWR1, including Methyl-CpG-BINDING DOMAIN 9 (MBD9) and three members of the Alfin1-like protein family, all of which have been shown to bind modified histone tails. Since mbd9 mutant plants were phenotypically similar to arp6 mutants, we explored a potential role for MBD9 in H2A.Z deposition. We found that MBD9 is required for proper H2A.Z incorporation at thousands of discrete sites, which represent a subset of the genomic regions normally enriched with H2A.Z. We also discovered that MBD9 preferentially interacts with acetylated histone H4 peptides, as well as those carrying mono- or dimethylated H3 lysine 4, or dimethylated H3 arginine 2 or 8. Considering that MBD9-dependent H2A.Z sites show a distinct histone modification profile, we propose that MBD9 recognizes particular nucleosome modifications via its PHD- and Bromo-domains and thereby guides SWR1 to these sites for H2A.Z deposition. Our data establish the SWR1 complex as being conserved across eukaryotes and suggest that MBD9 may be involved in targeting the complex to specific genomic sites through nucleosomal interactions. The finding that MBD9 does not appear to be a core subunit of the Arabidopsis SWR1 complex, along with the synergistic phenotype of arp6;mbd9 double mutants, suggests that MBD9 also has important roles beyond H2A.Z deposition.
| The histone H2A variant, H2A.Z, is found in all known eukaryotes and plays important roles in transcriptional regulation. H2A.Z is selectively incorporated into nucleosomes within many genes by the activity of a conserved ATP-dependent chromatin remodeling complex in yeast, insects, and mammals. Whether this complex exists in the same form in plants, and how the complex is targeted to specific genomic locations have remained open questions. In this study we demonstrate that plants do indeed utilize a complex analogous to those of fungi and animals to deposit H2A.Z, and we also identify several new proteins that interact with this complex. We found that one such interactor, Methyl-CpG-BINDING DOMAIN 9 (MBD9), is required for H2A.Z incorporation at thousands of genomic sites that share a distinct histone modification profile. The histone binding properties of MBD9 suggest that it may guide H2A.Z deposition to specific sites by interacting with modified nucleosomes and with the H2A.Z deposition complex. We hypothesize that this represents a general paradigm for the targeting of H2A.Z to specific sites.
| Nucleosomes, the fundamental units of chromatin that consist of ~147 bp of DNA wrapped around a histone octamer, efficiently condense large eukaryotic DNA molecules inside the nucleus. At the same time, nucleosomes present a physical barrier that restricts the access of DNA-binding proteins to regulatory sequences. This physical constraint imposed by nucleosomes on DNA can be modulated to expose or occlude regulatory DNA sequences, and is thereby used as a mechanism to control processes such as transcription that rely on sequence-specific DNA binding proteins. Thus, enzymatic complexes that can remodel chromatin structure by manipulating the position and/or composition of nucleosomes are essential for proper transcriptional regulation and the execution of key developmental programs.
All chromatin-remodeling complexes (CRCs) contain a DNA-dependent ATPase catalytic subunit that belongs to the SNF2 family of DNA helicases, along with one or more associated subunits [1, 2]. There are four major subfamilies of CRCs: SWI/SNF, INO80, ISWI, and CHD, all of which use the energy of ATP to either slide, evict, or displace nucleosomes, or to replace the canonical histones within nucleosomes with histone variants. One member of the INO80 CRC subfamily is the SWI2/SNF2-related 1 (SWR1) chromatin remodeler, a multisubunit protein complex required for incorporation of the H2A variant, H2A.Z, into chromatin [3, 4]. H2A.Z is a highly conserved histone variant found in all eukaryotes that plays important roles in regulating a variety of cellular processes, including transcriptional activation and repression, maintenance of genome stability and DNA repair, telomere silencing, and prevention of heterochromatin spreading [5–12]. Although loss of H2A.Z is not lethal in yeast [2, 13], H2A.Z is essential for viability in other organisms such as Tetrahymena [14], Drosophila [15, 16], and mice [17]. Interestingly, H2A.Z-deficient plants are viable but display many developmental abnormalities such as early flowering, reduced plant size, altered leaf morphology, and reduced fertility [18–20].
The SWR1 complex that mediates H2A.Z incorporation into chromatin was first described in yeast and is composed of 13 subunits, including Swr1, the catalytic and scaffolding subunit of the complex [3, 4, 21, 22]. In mammals, the functional and structural homolog of yeast SWR1 complex is the SRCAP (SNF2-related CREB-binding protein activator protein) complex. This complex is composed of 11 of the same subunits found in yeast SWR1 and is likewise able to exchange H2A/H2B dimers for H2A.Z/H2B dimers in nucleosomes [23–26]. Intriguingly, higher eukaryotes possess an additional multi-subunit complex, 60 kDa Tat-interactive protein (TIP60), that has histone acetyltransferase (HAT) activity and can also mediate the deposition of H2A.Z into nucleosomes [27–30]. Furthermore, the yeast Nucleosome Acetyltransferase of histone 4 (NuA4) complex, which is structurally related to SWR1 through the sharing of four subunits [31–35], was shown to regulate the incorporation of H2A.Z into chromatin cooperatively with the SWR1 complex [34].
Many homologs of yeast SWR1 and animal SRCAP complex subunits have been identified in Arabidopsis thaliana including ACTIN RELATED PROTEIN 6 (ARP6), SWR1 COMPLEX SUBUNIT 2 (SWC2), SWC6, PHOTOPERIOD-INDEPENDENT EARLY FLOWERING 1 (PIE1), and three H2A.Z paralogs: HTA8, HTA9, and HTA11. Numerous genetic and biochemical experiments suggest that the SWR1 complex is conserved in Arabidopsis. For example, it has been recently shown that Arabidopsis SWC4 protein directly interacts with SWC6 and YAF9a, two known components of the SWR1 complex [36]. Additionally, protein interaction experiments have demonstrated that PIE1 interacts directly with ARP6, SWC6, HTA8, HTA9, and HTA11 [18, 20, 37, 38], suggesting that PIE1 may serve as the catalytic and scaffolding subunit of an Arabidopsis SWR1-like complex. Furthermore, functional characterizations of PIE1, ARP6 and SWC6 have revealed that mutations in these genes have similar pleiotropic effects on Arabidopsis development, including a loss of apical dominance, curly and serrated rosette leaves, early flowering due to reduced expression of FLOWERING LOCUS C (FLC), altered petal number, and reduced fertility [18, 37–43]. Interestingly, genetic experiments revealed that the pie1 null phenotypes are more severe than those of arp6, swc6, hta9;hta11, or hta8;hta9;hta11 (h2a.z near-null) plants [18, 19, 37–43]. The more dramatic phenotypes in pie1 plants suggest that PIE1 has additional functions outside of H2A.Z deposition by SWR1, as previously proposed [5, 6, 18, 19]. A recent report also showed that mutant plants null for pie1 and h2az exhibited early developmental arrest, dying shortly after germination [19], further supporting the notion that PIE1 has H2A.Z-independent functions in Arabidopsis. On the other hand, genetic analyses of pie1;swc6 double mutant plants revealed that they had a phenotype indistinguishable from pie1 single mutants [38], and arp6;swc6 plants displayed the same defects as either arp6 or swc6 single mutant plants [18, 37]. These results further support the idea that PIE1, ARP6, and SWC6 act in the same genetic pathway and/or are the components of the same protein complex, but that PIE1 has additional functions.
Despite the strong genetic and biochemical evidence that Arabidopsis contains many conserved subunits homologous to the components of the yeast SWR1 complex and mammalian SRCAP, the plant SWR1 complex has not been successfully isolated and characterized. Recently, Bieluszewski and colleagues used Arabidopsis SWC4 and ARP4 proteins, two subunits shared between yeast SWR1 and NuA4, as well as mammalian SRCAP and TIP60 complexes, as baits to affinity purify their interacting partners from Arabidopsis cell suspension cultures [44]. These studies identified most of the subunits normally found in the SWR1 and NuA4 complexes, but it was not possible to determine whether this collection of proteins represented a single large complex or multiple complexes. Overall, it is not yet clear whether plants possess separate SWR1 and NuA4 complexes, SWR1 and TIP60-like complexes, or all three complexes [44].
The main goal of our study was to purify the Arabidopsis SWR1 complex in order to identify all of its components. To achieve this, we used the ARP6 protein, a subunit unique to SWR1 in other organisms, as bait in Tandem Affinity Purification (TAP) experiments. We performed three independent TAP experiments to isolate and identify ARP6-associated proteins. We identified all 11 conserved subunits found in yeast SWR1 and mammalian SRCAP complexes, demonstrating that Arabidopsis contains a bona-fide functional and structural homolog of these complexes. In addition, we identified several unexpected proteins that associated with ARP6, including the plant homeodomain (PHD)- and Bromo domain-containing protein Methyl CpG-BINDING DOMAIN 9 (MBD9), and the PHD domain-containing proteins ALFIN-LIKE 5 (AL5), AL6, and AL7. Association of these proteins with the SWR1 complex is consistent with the results of a recent related study that used ARP6-MYC or -FLAG tag epitope purification followed by mass spectrometry to identify ARP6-interacting proteins (Potok et al. 2019, bioRxiv 10.1101/657296). Genetic analyses revealed that mbd9 mutants showed phenotypic similarities to arp6 mutants, so we further explored a possible role for MBD9 in regulating H2A.Z incorporation into chromatin. We found that MBD9 is required for H2A.Z incorporation at a subset of the sites that normally harbor H2A.Z nucleosomes, and that these MBD9-dependent H2A.Z sites have distinct chromatin features. Furthermore, MBD9 strongly interacted with acetylated histone H4 peptides, H3 peptides mono- or dimethylated at lysine 4, as well as H3 dimethylated at arginines 2 or 8. We also found that MBD9 is not a core subunit of the Arabidopsis SWR1 complex and that double mutant arp6;mbd9 plants exhibited much more severe phenotypes than single arp6 or mbd9 mutants. These results collectively suggest that MBD9 targets the SWR1 complex to a subset of genomic loci but also has important functions beyond H2A.Z deposition.
To isolate the Arabidopsis SWR1 complex, we decided to use ARP6 protein as bait in Tandem Affinity Purification (TAP) experiments because ARP6 is exclusively found in the SWR1 complex in other organisms and is not shared by any other known CRCs [4, 18, 34]. For our purification experiments we used a GSrhino TAP-tag, which consists of two protein G domains, a tandem repeat of rhinovirus 3C protease cleavage site, and the streptavidin-binding peptide. This tag has been successfully used to purify several plant nuclear complexes, including SWI/SNF type CRCs [45]. Furthermore, the use of this tag provides high yield of purified proteins and specificity of purification. In addition, a list of 760 proteins that nonspecifically bind to this tag or the associated purification beads has been assembled from data on 543 GS-based TAP experiments [45]. We fused the GSrhino TAP-tag to either the N-terminal end (N-TAP-ARP6) or the C-terminal end (C-TAP-ARP6) of the genomic ARP6 coding sequence, containing the endogenous ARP6 promoter, and introduced the constructs into arp6-1 mutant plants to test for the ability of each transgene to complement a null arp6 allele. Using western blotting, we first detected the presence of the 67 kDa ARP6-TAP-tag fusion protein specifically in plants homozygous for the transgene and not in arp6-1 or wild-type (WT) plants (Fig 1A). Next, we assessed the ability of the transgenes to rescue the morphological defects of arp6-1 plants. When grown in parallel, the transgenic plants appeared almost indistinguishable from WT plants, with more compacted, non-serrated rosette leaves as compared to arp6-1 mutants (Fig 1B). The N-TAP-ARP6 and, to a lesser degree, the C-TAP-ARP6 transgenes are able to rescue the early flowering phenotype of arp6-1 plants (Fig 1C) as evident by the significantly higher number of rosette leaves at the time of flowering in wild type and transgenic plants compared to arp6-1 plants (Fig 1D). Finally, all transgenic plants showed full complementation for the loss of apical dominance and fertility defects of arp6-1 mutant plants (Fig 1E and 1F). Overall, we conclude that the N-TAP-ARP6 and C-TAP-ARP6 transgenes are fully functional and thus suitable for affinity purification.
Since the N-TAP-ARP6 and C-TAP-ARP6 transgenes were fully functional, we proceeded with the affinity purification experiments using two independent N-TAP-ARP6 transgenic lines (N-TAP 11–4 and N-TAP 1–2) and one C-TAP-ARP6 line (C-TAP 10–2). We followed the protocol described by Van Leene and colleagues to purify and elute the ARP6-TAP-tag interacting proteins and identified the eluted proteins by liquid chromatography coupled with tandem mass spectrometry (LC-MS/MS) [45]. All eluted proteins detected in our three TAP-tag experiments are listed in S1 Table. Using the database of non-specific binders of the GSrhino TAP-tag, we eliminated many proteins from this list as false positives and compiled a list of ARP6-interacting proteins. Among these proteins, we identified ARP6, SWC2, SWC4, SWC6, PIE1, RuvB1, RuvB2, ACTIN1, ARP4, YAF9a, and H2A.Z proteins as Arabidopsis homologs of all 11 conserved subunits found in both yeast SWR1 and mammalian SRCAP complexes (Table 1). While we were able to detect HTA9 or HTA11 in all of our TAP-tag experiments, we never detected HTA8, the third member of the H2A.Z family. This finding is perhaps not surprising considering the fact that HTA8 has the lowest expression in all tissues when compared to the other H2A.Z family members in Arabidopsis (S1A Fig). We also did not detect the YAF9b protein (Table 1) even though Arabidopsis YAF9a and YAF9b have been previously shown to act redundantly and are required for proper FLC expression [44, 46, 47]. In addition to H2A.Z, we identified three H2B histones in our TAP experiments: HTB2, HTB4, and HTB9 (Table 1). Since H2A.Z histones are deposited into nucleosomes as H2A.Z/H2B dimers, we sought to investigate whether specific H2A.Z proteins might have preferential H2B partners. If this is true, we would expect to see synchronized expression of specific H2A.Z/H2B pairs in various Arabidopsis tissues. Using publicly available microarray expression data [48] for the two H2A.Z and three H2B histones that we identified, we observed that HTA11 and HTB2 had highly similar expression profiles across tissues (S1B Fig), while the expression of HTA9 matched very well with HTB4 expression and, to a slightly lesser degree, with HTB9 expression (S1C and S1D Fig). These results indicate that the Arabidopsis H2A.Z histones may have preferential H2B partners when deposited as dimers into nucleosomes.
Interestingly, in addition to known subunits of the SWR1 complex, we also reproducibly identified several other nuclear proteins as being associated with the SWR1 complex (Table 1). These include MBD9, a protein with a methyl-CpG-binding domain and various chromatin-binding domains [49–51], TRA1A and TRA1B, two subunits in the SPT module of an Arabidopsis SAGA complex [52] that are also homologs of the yeast NuA4 subunit Tra1 and the mammalian TIP60 subunit TRRAP [34, 44], and three members of a plant-specific Alfin1-like family (AL5, AL6, and AL7) best known for their regulation of abiotic stress responses in Arabidopsis and ability to bind di- and tri-methylated lysine 4 of histone H3 (H3K4me2/3) [53, 54].
MBD9, a methyl-CpG-binding domain-containing protein, was identified in all three TAP-tag experiments as an ARP6-interacting partner (Table 1). Previous studies have shown that mbd9 mutants flowered significantly earlier than WT plants, due to reduced FLC expression, and produced more inflorescence branches when compared to WT plants [50], which are phenotypes also found in arp6 mutants [40]. We discovered that, in addition to the above-mentioned defects, mbd9 plants have serrated rosette leaves and a significantly increased number of flowers with extra petals (S2 Fig), which are phenotypes also associated with the loss of ARP6 [18, 39, 40]. Furthermore, examination of the previously reported MBD9 enrichment pattern at the FLC locus revealed that MBD9 occupied the two FLC regions previously shown to also have the highest H2A.Z enrichment in WT plants [41, 51]. Given that arp6 and mbd9 plants have similar phenotypes, that H2A.Z and MBD9 appear to occupy the same sites at the FLC locus, and that MBD9 co-purified with ARP6 in our TAP-tag experiments, we investigated whether MBD9 plays a role in the incorporation of H2A.Z into chromatin. We performed three biological replicates of chromatin immunoprecipitation coupled with high-throughput sequencing (ChIP-seq) using an H2A.Z antibody [41] on WT, arp6-1, and mbd9-1 seedlings. At least 8 million high-quality, nuclear reads were aligned to the TAIR10 genome for each replicate, resulting in thousands of reproducible H2A.Z peaks for each genotype (S3A Fig). Spearman correlation analysis indicated a high degree of overlap among H2A.Z ChIP replicates for a given genotype with the exception of mbd9-1 replicate 3, which correlated moderately with other two mbd9 replicates (S3B Fig).
The average H2A.Z enrichment profile (the average of all three scaled replicates) across all gene bodies in WT plants showed the highest enrichment of H2A.Z just after the transcription start site, with decreasing enrichment toward the 3’ end, as expected (Fig 2A). The pattern of H2A.Z enrichment across genes in arp6-1 showed a similar profile but with extremely reduced enrichment, while mbd9-1 plants had an intermediate level of H2A.Z enrichment between WT and arp6-1 (Fig 2A). To analyze all of the regions normally enriched for H2A.Z, we identified peaks of H2A.Z enrichment that were present in at least two of the three H2A.Z ChIP-seq replicates in WT plants, and we then examined H2A.Z levels at these same sites in arp6-1 and mbd9-1 mutants. As observed for gene bodies, the 7039 sites reproducibly enriched for H2A.Z in WT were nearly depleted of H2A.Z in arp6-1, while there was an intermediate H2A.Z depletion in mbd9-1 plants (Fig 2B). Western blotting for H2A.Z on acid-extracted nuclear proteins isolated from WT, arp6-1, and mbd9-1 showed a decrease in H2A.Z levels in both arp6-1 and mbd9-1, suggesting that the lack of chromatin incorporation may lead to H2A.Z degradation (S4A and S4B Fig). To test whether the observed reduction of H2A.Z incorporation might be due to either SWR1 components or H2A.Z genes being misregulated in mbd9 plants, we performed qRT-PCR experiments to measure the expression of PIE1, ARP6, MBD9, SWC4, SWC6, and YAF9a, as well as HTA8, HTA9, and HTA11 genes, in WT, arp6-1, and mbd9-1 plants. We found that the expression of these genes was not substantially altered and, therefore, unlikely to account for the observed depletion of H2A.Z levels in mbd9-1 plants (S5 Fig). Using RNA-seq, a related study has also demonstrated that the expression of SWR1 subunit genes was largely unaffected in mbd9-3 plants (Potok et al. 2019, bioRxiv 10.1101/657296). Overall, these results indicate that MBD9 is indeed required for proper H2A.Z incorporation into chromatin.
To further confirm our results with respect to the role of MBD9 in H2A.Z deposition, we performed ChIP-qPCR experiments using WT, arp6-1, mbd9-1, and two additional mbd9 T-DNA alleles, mbd9-2 and mbd9-3 [50]. We first assayed H2A.Z abundance at two distinct regions of the FLC gene: the first and last exon (regions 2 and 9, respectively, as described in [41]). Regions 2 and 9 are the sites on the FLC gene where H2A.Z is most highly enriched in WT plants, and that enrichment is lost in arp6-1 mutant plants (S6A Fig, [41]). We found that in plants homozygous for any of the three mbd9 alleles, the amount of H2A.Z at FLC regions 2 and 9 was reduced at least 2-fold when compared to WT plants (S6A Fig), indicating that MBD9 contributes to H2A.Z deposition at the FLC gene. We also measured the H2A.Z abundance at ASK11 and At4, two phosphate starvation response genes previously shown to have H2A.Z deposited in their chromatin [55]. We discovered that in mbd9 plants these genes were depleted of H2A.Z to similar levels as in arp6-1 plants when compared to WT (S6B Fig). Taken together, our results indicate that MBD9 is required for proper H2A.Z levels at multiple Arabidopsis genomic loci and is, therefore, functionally related to the SWR1 complex.
As MBD9 was previously reported to have histone acetyltransferase (HAT) activity in vitro and was found to associate with acetylated H4 [51], we also examined the global levels of histone H4 N-terminal acetylation (H4Ac) in WT, mbd9-1, and arp6-1 plants. At least 5 million high-quality, nuclear reads were aligned to the TAIR10 genome for each replicate, resulting in thousands of reproducible, H4Ac peaks for each genotype (S3A Fig). Spearman correlation analysis indicated a high degree of overlap among H4Ac ChIP replicates for a given genotype with the exception of mbd9-1 replicate 3, which correlated moderately with other two mbd9-1 replicates due to having the lowest levels of H4Ac among the replicates (S3C and S7 Figs). If MBD9 is responsible for global acetylation of H4, as previously reported [51], we would expect to see a dramatic reduction of acetylated H4 in mbd9-1 mutants compared to WT. However, we found that the average genome-wide distribution of acetylated H4 was only modestly reduced in mbd9-1 plants, while arp6-1 plants had H4Ac levels similar to WT when examined across all gene bodies or at all sites enriched for acetylated H4 in WT plants (Fig 2C and 2D). Western blot analysis using H4Ac antibodies on acid-extracted nuclear proteins from WT, arp6-1, and mbd9-1 plants was consistent with the ChIP-seq findings that mbd9-1 plants have moderately reduced levels of H4Ac when compared to WT (S4A and S4C Fig). Collectively, our ChIP-seq results demonstrate a major role for MBD9 in maintaining proper H2A.Z levels. MBD9 may also have a minor role in the acetylation of H4, but this is likely to be indirect given that the protein does not contain an identifiable acetyl CoA-binding or acetyltransferase domain.
The intermediate loss of H2A.Z in mbd9-1 plants compared with arp6-1 plants (Fig 2A and 2B) suggested that MBD9 may be required for incorporation of H2A.Z at only a subset of H2A.Z-enriched sites or it may be required for complete H2A.Z deposition at all sites. To identify genomic regions that require MBD9 for H2A.Z incorporation into chromatin, we quantified the normalized H2A.Z ChIP-seq read abundance from WT, arp6-1, and mbd9-1 mutant plants across all of the H2A.Z-enriched regions that were reproducibly identified in the WT replicates. We performed DESeq analysis [56] to quantitatively compare WT to mbd9-1 and WT to arp6-1 H2A.Z levels at each site (S2 Table). H2A.Z levels were significantly depleted in arp6-1 at nearly all of the H2A.Z sites, as expected for a mutation that disrupts the SWR1 complex (Fig 3A). In contrast, out of the 7039 H2A.Z-enriched sites, we identified only 1391 sites that had reduced H2A.Z by at least 1.5-fold (log2 fold change of at least 0.6 with an adjusted p value ≤ 0.05) in mbd9-1 compared to WT (Fig 3B).
To further examine the nature of the H2A.Z deposition defect in mbd9 mutants, we visualized H2A.Z enrichment and distribution across the 1391 sites that lose H2A.Z in mbd9-1, which we refer to as MBD9-dependent H2A.Z sites. For comparison, we selected a similarly sized set of MBD9-independent H2A.Z sites (1505 sites with an average fold difference of less than 1.19 between WT and mbd9-1, which is an absolute log2 fold change of less than 0.25). This analysis revealed a drastic reduction in H2A.Z occupancy at each of the MBD9-dependent H2A.Z sites when comparing WT and mbd9-1, but with maintenance of the same overall pattern of occupancy (Fig 3C). The level of H2A.Z in mbd9-1 at MBD9-dependent sites was depleted to levels similar to arp6-1. In contrast, the MBD9-independent H2A.Z sites showed equivalent profiles and occupancy levels between WT and mbd9-1, but still showed reduced H2A.Z in arp6-1 (Fig 3C). To further validate these results, we chose three MBD9-dependent and three MBD9-independent H2A.Z sites and performed qPCR experiments using the same H2A.Z ChIP material from WT and mbd9-1 plants previously utilized for ChIP-qPCR analysis of FLC, ASK11, and At4 genes. We found that all three MBD9-dependent loci showed statistically significant depletion of H2A.Z in mbd9-1 when compared to WT (S8A and S8B Fig), while none of the three MBD9-independent sites had significantly different levels of H2A.Z between mbd9-1 and WT plants (S8C and S8D Fig). Thus, we conclude that MBD9 is required for proper H2A.Z deposition at a subset of the sites that this histone variant normally occupies.
In order to understand why MBD9 is required for H2A.Z deposition at certain sites and not others, we first analyzed the genomic distribution of MBD9-dependent H2A.Z sites compared to the MBD9-independent H2A.Z sites. We found that the two sets are distributed similarly across the genome, with more than 80% of each set of coordinates localizing within genic regions (S9A Fig). Both MBD9-dependent and MBD9-independent H2A.Z sites were found primarily at the 5’ end of protein-coding gene (PCG) bodies, with the MBD9-dependent sites found slightly more upstream of the MBD9-independent sites (S9B and S9C Fig). PCGs nearest to the MBD9-dependent sites had reduced H2A.Z levels throughout the entire gene body in mbd9-1 and arp6-1 plants, whereas PCGs nearest to the MBD9-independent sites had similar gene body H2A.Z levels in WT and mbd9-1, but still had reduced H2A.Z levels in arp6-1 (S9C Fig). For the 1322 PCGs found nearest to the MBD9-dependent H2A.Z sites (S2 Table) there was no significantly overrepresented Gene Ontology (GO) terms identified using either of the two different GO analysis tools (see methods), indicating that MBD9-mediated deposition of H2A.Z is not associated with functionally-related gene sets or a particular cellular pathway.
We also examined various histone modification profiles at the two types of sites using publicly available ChIP-seq data from WT plants in order to discern any differences between H2A.Z sites that require MBD9 and those that do not. Interestingly, we found that in WT plants the average level of histone H3 lysine 9 acetylation (H3K9Ac) is higher at MBD9-dependent H2A.Z sites than it is at the sites that do not require MBD9 (Fig 4A). However, no significant differences were found in the average enrichment of histone H3 lysine 18 acetylation or H3 lysine 27 trimethylation (H3K18Ac and H3K27me3, respectively) between the two types of loci (Fig 4B and 4C). On the other hand, we did observe less abundant enrichment of dimethylated lysine 9 of histone H3 (H3K9me2) at MBD9-dependent H2A.Z sites (Fig 4D). The same pattern was observed for H3 trimethylation at lysine 4 or lysine 36 (H3K4me3 and H3K36me3, respectively), with consistently lower levels of each mark at MBD9-dependent H2A.Z sites (Fig 4E and 4F).
To better understand the differences in chromatin profiles between the two sets of loci, ten different histone marks (H2A.Z, H4Ac, H3K27me3, H2AK121Ub, H3K9me2, H3K9Ac, H3K18Ac, H3K27Ac, H3K4me3, and H3K36me3) were plotted around a 2 kb window centered on the H2A.Z peak of either the MBD9-dependent or MBD9-independent sites (S10A Fig). The heatmaps were clustered into 4 k-means clusters and three similar sets of chromatin states were observed in both MBD9-dependent and MBD9-independent loci. These include euchromatic regions (S10B Fig), H3K27me3-rich sites likely regulated by PRC2 (S10C Fig), and sites that have H2A.Z but are relatively depleted of the other examined histone modifications. The observed average differences in specific histone marks between the two sets of loci (Fig 4) appear to be driven by increased levels of H3K9Ac at MBD9-dependent euchromatic regions (S10B Fig), decreased levels of H3K9me2 and H3K4me3 in MBD9-dependent PRC2-regulated regions (S10C Fig), as well as reductions in H3K9me2 and H3K36me3 at the MBD9-dependent regions relatively depleted of most histone modifications (S10D Fig).
Given the predominant gene-body localization of H2A.Z, the differences in histone modification levels between MBD9-dependent and -independent sites could simply reflect differences in expression levels of the underlying genes. However, using publically available RNA-seq data, we found that genes nearest to the sites in each category span a wide range of expression levels and are not significantly different from one another in terms of steady-state transcript levels (unpaired t-test, p < 0.05, S11 Fig). Thus, MBD9 may specifically recognize or be repelled by specific chromatin features, which could help guide the SWR1 complex to specific DNA sites.
To examine which histone marks MBD9 may directly recognize, we performed a histone peptide array assay using the full-length MBD9 protein (S12 Fig). We found that MBD9 most strongly interacted with histone H4 peptides that were acetylated at lysine (K) residues 12, 16, and 20, or acetylated at K12 and K16 while being di- or trimethylated at K20 (Fig 5). The protein also bound to H4 peptides containing both K16Ac and K20Ac, or K12Ac and K16Ac along with K20me1. This affinity of MBD9 for acetylated H4 is consistent with a previous report [51].
Interestingly, MBD9 interacted with H3 peptides that were mono- or dimethylated at K4, but did not appear to bind the H3 peptide that was trimethylated at K4. This lack of affinity for H3K4me3 could explain the relative depletion of this mark at MBD9-dependent H2A.Z sites (Fig 4E). Other H3 peptides that showed significant binding of MBD9 include those dimethylated at arginines 2 and/or 8 (Fig 5). Collectively, these results suggest that MBD9 recognizes distinct combinations of histone modifications that are likely to influence its chromatin binding characteristics.
To determine whether the MBD9 protein, with an estimated molecular mass of 240 kDa, is an integral component of the Arabidopsis SWR1 complex (Fig 6A) we performed size-exclusion chromatography (SEC) experiments on protein extracts from WT and mbd9-1 plants, followed by western blotting for ARP6. This allowed us to define the native size of the complex and to determine whether this size changes in the absence of MBD9, as would be expected if this protein were a stoichiometric component of the SWR1 complex (Fig 6B), as previously demonstrated for the PIE1 subunit [41]. Using an ARP6 monoclonal antibody [40], we detected ARP6 protein in its native form as a part of a multi-subunit complex with a molecular mass of ~800 kDa (Fig 6C). When the SEC experiments were performed on mbd9-1 extracts, the ARP6 peak did not significantly shift and the estimated molecular mass of the native ARP6 complex in mbd9-1 plants was ~775 kDa (Fig 6C) in two biological replicates. These results strongly suggest that MBD9 is not a core component of the ARP6-containing SWR1 complex, but most likely interacts with it in a more transient manner. Alternatively, MBD9 may be a stable component of a minor subset of SWR1 complexes.
Collectively, we discovered that MBD9 is required for proper H2A.Z deposition at many sites, but is not stably associated with the SWR1 complex. To investigate genetic interactions between MBD9 and ARP6, we generated arp6-1;mbd9-1 double mutant plants. We have shown that single arp6-1 and mbd9-1 mutant plants have similar phenotypic defects (S2 Fig) and that both ARP6 and MBD9 regulate H2A.Z incorporation into chromatin (Fig 2A and 2B). If these two proteins are subunits of the same complex or function exclusively in the same genetic pathway then double mutant plants should be phenotypically indistinguishable from single mutants, as previously shown for arp6;swc6 plants [18, 37]. Instead, we observed that the double mutants displayed much more severe defects (dwarf stature, deformed leaves, and drastically reduced fertility) than the individual single mutants throughout development (Fig 7). Importantly, these phenotypes in the double mutant plants reverted back to those of each single mutant by introducing either the genomic ARP6 or MBD9 constructs into the double mutants (S13 Fig), indicating that these defects were truly the result of simultaneous loss of ARP6 and MBD9 functions. To investigate whether the more severe phenotype in arp6-1;mbd9-1 plants is caused by further reduction of H2A.Z incorporation into chromatin, we performed a ChIP-qPCR experiment using H2A.Z antibody on WT, arp6-1, mbd9-1, and arp6-1;mbd9-1 plants at FLC regions 2 and 9. We found that double mutant plants had similar levels of H2A.Z depletion at FLC when compared to arp6-1 (S14 Fig). Collectively, our findings support the idea that MBD9 is not a core subunit of the SWR1 complex and suggest that this protein has additional functions outside of H2A.Z incorporation [57].
Previous studies provided important evidence suggesting that Arabidopsis contains a SWR1-like complex that mediates incorporation of H2AZ into chromatin [6, 18, 20, 36–44]. Using the SWR1-specific subunit ARP6 as bait, we successfully purified the Arabidopsis SWR1 complex and identified all 11 conserved subunits that are also found in the yeast SWR1 and mammalian SRCAP complexes. Recently, a similar study also used ARP6 as bait to isolate the Arabidopsis SWR1 complex and identified the same proteins as our study, with the addition of AL4 (Potok et al. 2019, bioRxiv 10.1101/657296). These and our results suggest that the function and structure of the canonical SWR1 complexes that incorporate histone H2A.Z into nucleosomes have been well preserved over evolutionary timescales and may be found in all eukaryotes.
TIP60 in animals is a single multifunctional complex that combines the subunits and functions of yeast SWR1 and NuA4 complexes [34]. It appears that this merger evolved as a result of the fusion of the two major scaffolding proteins, the Swr1 ATPase of the yeast SWR1 complex and the Eaf1 protein of the yeast NuA4 complex, into a single p400-like protein. This is based on the fact that p400 contains HSA, ATPase, and SANT domains, which are found separately in the yeast Swr1 and Eaf1 proteins [34, 44]. Intriguingly, Arabidopsis PIE1 (yeast Swr1 homolog) also contains HSA, ATPase, and SANT domains, implying that PIE1 may also be an ortholog of p400. In fact, Bieuszewski and colleagues originally hypothesized that PIE1 is a scaffolding component of an Arabidopsis TIP60-like complex [34, 44]. Interestingly, TRA1A and TRA1B, two proteins that co-purified with ARP6 in our TAP experiments, were recently characterized as subunits of the SPT module of an Arabidopsis SAGA complex [52] and are homologs of the yeast NuA4 subunit Tra1 and the mammalian TIP60 subunit TRRAP, further suggesting an intimate functional relationship among Arabidopsis SWR1 and NuA4/HAT complexes (Table 1). Taken together, it is plausible that plants possess both the canonical SWR1 complex and an independent NuA4-like complex, as in yeast, and may also contain a TIP60-like complex, which is found only in animals. Future purification experiments using Arabidopsis PIE1 as bait are crucial to address the question of whether plants have two distinct PIE1-containing complexes (SWR1 and a TIP60-like), and which subunits are shared between the two complexes. The existence of two different PIE1 complexes could also explain why the phenotype of pie1 mutant plants is distinct from that of h2a.z mutants or mutations in other SWR1 components [18, 19, 37–43].
Based on multiple studies in many model organisms, we now have a good understanding of how the SWR1 complex incorporates H2A.Z into nucleosomes [4, 58, 59]. However, several aspects of SWR1 biology are still poorly understood, including precisely how the complex is recruited to specific chromatin regions to deposit H2A.Z. In yeast, it has been shown that NuA4-mediated acetylation of specific nucleosomal sites is important for SWR1 targeting to chromatin and H2A.Z incorporation [31–35, 60]. In addition, it has been proposed that Bdf1, a bromodomain-containing subunit of the yeast SWR1 complex, recruits the complex to chromatin by recognizing acetylated H4 tails [31, 61]. Supporting this notion, the loss of Bdf1 results in global reduction of H2A.Z in chromatin [62].
In plants, little is known about the mechanisms that target the SWR1 complex to specific chromatin loci. Recent results from the Jarillo group suggest that binding of the SWC4 subunit to AT-rich DNA elements in promoters of certain genes can recruit the Arabidopsis SWR1 complex to these chromatin regions to deposit H2A.Z [36]. However, only a subset of H2A.Z-enriched genes contain AT-rich elements in their promoters, which strongly suggests that additional mechanisms of SWR1 recruitment exist in plants. The same group has recently demonstrated that the YAF9a subunit, by interacting with acetylated histones, can recruit the SWR1 complex to a subset of SWC4 target genes [47].
What role may MBD9 play in SWR1 recruitment to specific chromatin loci? In addition to a methyl-CpG-binding (MBD) domain, which is in fact thought to bind unmethylated DNA ([63], Potok et al. 2019, bioRxiv 10.1101/657296), MBD9 contains an acetyl lysine-binding bromodomain [64], and two plant homeodomains (PHD) which may recognize methylated lysine and arginine residues [65, 66]. We showed that in mbd9 mutant plants the level of H2A.Z incorporation is significantly reduced at a subset of H2A.Z-enriched regions (Fig 3) and that these MBD9-dependent H2A.Z loci have distinct histone modification profiles relative to MBD9-independent H2A.Z loci (Figs 4 and S10). Specifically, H2A.Z sites that are dependent on MBD9 had higher levels of H3K9Ac and lower levels of H3K4me3, H3K36me3, and H3K9me2. Using the histone peptide array assay, we demonstrated that MBD9 strongly interacts with acetylated H4, suggesting that MBD9 may recognize specific H4 acetylations in the context of other modifications. The peptide array data also revealed that MBD9 has high affinity for both symmetrically and asymmetrically di-methylated arginines in the H3 N-terminus (Fig 5). The significance of this interaction is not yet clear as there are no currently available ChIP-seq data for these modifications in Arabidopsis, and the effects of methylated arginines on chromatin architecture are only poorly understood in plants [67]. The histone modification preferences of MBD9 defined by peptide array may at least partially explain some of the differences in chromatin profiles between H2A.Z sites that require MBD9 and those that do not. For example, the H3K4me3 depletion seen in a subset of MBD9-dependent sites (S10C and S10D Fig) may be explained by the relative preference of MBD9 for binding H3K4me1 or H3K4me2, but not H3K4me3. However, the binding preferences of MBD9 appear to be complex, and there are likely one or more features that more closely define the MBD9-dependent sites (e.g. methylation of H3R2 and/or R8).
MBD9 was recently found to preferentially localize to nucleosome-depleted regions (NDRs) directly upstream of H2A.Z nucleosomes (Potok et al. 2019, bioRxiv 10.1101/657296). In yeast, SWR1 also localizes to NDRs and it is proposed that while NDR localization serves as a general recruiting mechanism for SWR1, it is the interaction between SWR1 components and the nucleosomes flanking these NDRs that define the actual sites of H2A.Z deposition [68, 69]. Collectively, the current data point to a model whereby MBD9 recognizes nucleosomes with specific modification patterns, and perhaps specific DNA sequences, and interacts with SWR1 to effect its localization to specific genes.
Although MBD9 was co-purified in all of our ARP6 TAP-tag experiments (Table 1), MBD9 appears not to be a core component of the SWR1 complex (Fig 5). Two possible conclusions about MBD9’s interaction with the SWR1 complex can be made based on these results. MBD9 may interact only transiently with components of the SWR1 complex, and is therefore detected in TAP-tag experiments as previously observed for transcription factors and cofactors that recruit Arabidopsis SWI/SNF and PRC2 complexes to specific chromatin sites [70–73]. Alternatively, MBD9 could be more tightly associated with only a subset of all SWR1 complexes in Arabidopsis. In that case, our size exclusion chromatography peak of SWR1 most likely would not show a significant mass reduction in mbd9-1 plants when compared to WT because the loss of MBD9 would affect only a minor fraction of SWR1 complexes.
With regard to the synergistic phenotype of arp6;mbd9 mutants, this could result from a further reduction of H2A.Z deposition in the double mutant compared to arp6, or it may be a manifestation of functions of MBD9 beyond H2A.Z deposition. These two possibilities are, of course, not mutually exclusive. While we did not observe additional H2A.Z reduction at FLC in arp6;mbd9 compared to arp6, a related study found that in arp6;mbd9 double mutant plants the level of H2A.Z incorporation into chromatin is indeed further reduced genome-wide (Potok et al. 2019, bioRxiv 10.1101/657296). However, this study also found that MBD9 strongly interacts with the ISWI family of CRCs, suggesting that MBD9 has additional nucleosome remodeling functions outside of H2A.Z deposition. While further experiments are needed to determine the precise nature of MBD9’s interaction with the SWR1 complex, it is clear that MBD9 is functionally associated with the SWR1 complex and is integral to the deposition of H2A.Z at a subset of loci.
An important point to consider is whether the specific H2A.Z loss from chromatin in mbd9 mutants reflects a lack of targeting of SWR1 to MBD9-dependent H2A.Z sites, or whether the complex is targeted properly but H2A.Z is not retained in nucleosomes after deposition in mbd9. It is indeed formally possible that MBD9 plays a role in H2A.Z retention in chromatin. While an absolute resolution to this issue awaits further experimentation, the most parsimonious explanation for the present observations is that MBD9 interacts with the SWR1 enzyme complex to influence where its reaction occurs, rather than acting on the product of the reaction (an H2A.Z-containing nucleosome).
The presence of H2A.Z in chromatin has been linked to both gene activation and gene repression, but how H2A.Z affects transcription in this context-dependent manner is not clear [74, 75]. In addition, how the chromatin remodelers that deposit H2A.Z are recruited to specific chromatin loci is poorly understood. Our isolation of the Arabidopsis SWR1 complex identified unexpected proteins that co-purified with this complex, including MBD9 and three members of the plant-specific Alfin family, each of which contain known modified histone-binding domains. Based on our results and data from other studies, we propose that these SWR1-associated proteins are involved in the recruitment of the complex to chromatin to incorporate H2A.Z at specific loci. In this view, the H2A.Z landscape reflects the collective effects of inherent SWR1 targeting as well as SWR1’s association with a variety of adaptor proteins, such as MBD9. With the identification of these SWR1-associated proteins, we can now start to address important mechanistic questions about the activity of SWR1 in plants and how MBD9, and perhaps Alfin1-like family proteins, may modulate SWR1 functions.
Arabidopsis thaliana of the Columbia (Col-0) ecotype was used as the wild type reference, and all mutant seeds are of the Col-0 ecotype. The arp6-1 (SAIL_599_G03), and mbd9-1 (SALK_054659), mbd9-2 (SALK_121881) and mbd9-3 (SALK_039302) alleles were described previously [40, 50]. Seedlings were grown in either soil, half-strength Murashige and Skoog (MS) liquid media [76], or on half-strength MS media agar plates, in growth chambers at 20°C under a 16 hour light/8 hour dark cycle. Plasmids containing the N-TAP-ARP6, C-TAP-ARP6, and gMBD9 constructs, driven under endogenous ARP6 and MBD9 promoters, respectively, were introduced into Agrobacterium tumefaciens GV3101 strain by electroporation. Plants were transformed with these constructs via the floral dip method [77]. Primary transgenic plants were selected on half-strength MS media agar plates containing 50 mg/L hygromycin and 100 mg/L timentin, and then transferred to soil. Two to three grams of sterilized WT seeds and T3 seeds homozygous for N-TAP-ARP6 and C-TAP-ARP6 constructs were germinated for 6 days in flasks containing 600 ml of half-strength MS media with constant shaking on rotating platform (80–90 rpm). After 6 days, the germinated seedlings were filtered to remove the excess liquid, and 50 grams of seedling tissue was frozen in liquid nitrogen and stored at -80°C.
To construct the ARP6-TAP-tag, we fused genomic ARP6 sequence to the tandem affinity purification (TAP) GSrhino tag, recently developed for efficient affinity purifications of protein complexes in plants [45]. Gateway–compatible plasmids containing either a C-terminal TAP-tag (pEN-R2-GS_rhino-L3, [45]) or an N-terminal TAP-tag (pEN-L1-NGS_rhino-L2, [45]) were used to produce the C-TAP-ARP6 and the N-TAP-ARP6 constructs, respectively. To generate the C-TAP-ARP6 construct, a total of six primers were used in three overlapping PCR reactions to produce a ~4.7 kb attB PCR fragment. This PCR product contained ~4.1 kb of the genomic ARP6 sequence (from -2040 bp upstream of the start codon to +2083 bp downstream from the start codon), ~600 bp of the TAP-tag sequence fused to the C-terminal end of the ARP6 gene, and attB adapters at the 5’ and 3’ ends of the PCR product for Gateway cloning. This PCR fragment was subcloned into pDONR221 gateway plasmid via BP recombination reaction using BP clonase II enzyme (Invitrogen). The construct was verified by sequencing and further sub-cloned into the destination gateway plasmid pMDC99 [78] using the LR clonase II enzyme in LR recombination reaction (Invitrogen). Similarly, the attB N-TAP-ARP6 construct was first produced using six PCR primers in overlapping PCR reactions containing the same genomic ARP6 DNA fragment as in the C-TAP-ARP6, with the TAP-tag fused at the N-terminal end of the ARP6 gene. This PCR fragment was then sub-cloned into pDONR221 via BP reaction, verified by sequencing, and finally sub-cloned into the pMDC99 destination plasmid via LR reaction.
To generate gMBD9 construct, which was used to transform arp6-1;mbd9-1 double mutant plants, we first PCR-amplified 11,311 bp of genomic MBD9 sequence (from –1936 bp upstream of the start codon to + 9372 downstream from the start codon) using gMBD9 sequence-specific primers with attB adapters at their 5’ ends. The attB PCR product was then sub-cloned into pDONR221 gateway plasmid via BP recombination reaction (Invitrogen), verified by sequencing, and finally sub-cloned into the destination gateway plasmid pMDC99 [78] using LR recombination reaction (Invitrogen).
To clone the N-terminal Myc-MBD9 construct into pT7CFEChis vector (Thermo Scientific), we designed two sets of primers: first set was used to PCR-produce an N-terminal Myc-tag fused in frame with the full length MBD9 cDNA, while second primer pair was used to PCR-amplify the full pT7CFEChis vector. The two PCR products were then used in a Gibson Assembly reaction (New England Biolabs) to clone Myc-MBD9 at an NdeI restriction cloning site in the pT7CFEChis plasmid to achieve a maximum expression of a fusion protein in this system. The cloning was verified by sequencing.
To produce arp6-1;mbd9-1 double mutant plants, pollen from arp6-1 plants was used to manually pollinate mbd9-1 plants. Since ARP6 and MBD9 genes are both on chromosome 3, we were only able to identify the F2 plants that were homozygous for one T-DNA allele and heterozygous for the other. We used F3 seeds from arp6-1/arp6-1;mbd9-1/+ plants to identify the double mutant plants.
The proteins for western blot detection shown in Fig 1A were extracted from ~100 mg of 6-day old whole transgenic seedlings homozygous for the arp6-1 allele and either the C-TAP-ARP6 or N-TAP-ARP6 constructs by first making a crude nuclei preparation using Nuclei Purification Buffer [79]. The nuclei pellets were then resuspended in 2 volumes of 1x Laemmli’s sample buffer (125 mM Tris-HCl pH 6.8, 4% SDS, 30% glycerol, and 1% β-mercaptoethanol) prior to heating and loading on a gel. For ARP6 detection on fractions from SEC experiments (see below), the eluted proteins were isolated by adding 20 μl of the StrataClean resin (Agilent) to 1 ml of each SEC fraction, incubating for 20 minutes at room temperature (RT) on a rotating platform, and then spinning down for 2 minutes at 5,000g at RT. The pelleted proteins were resuspended in 20 μl of 1x Laemmli’s sample buffer. The total proteins for western blot detection shown in S4 Fig were isolated from ~150 mg of young leaves using acid extraction protocol (see below) and the pelleted proteins were resuspended in 50 μl of 1x Laemmli’s sample buffer. The proteins were then separated on 4–20% Novex WedgeWell tris-glycine gel (Invitrogen) and transferred to Amersham nitrocellulose blotting membrane (GE Healthcare). After blocking overnight in PBST buffer (137 mM NaCl, 2.7 mM KCl, 10 mM Na2HPO4, 2 mM KH2PO4, and 0.05% tween20) containing 5% non-fat dry milk, the blots were incubated with primary antibody (1:100 dilution for monoclonal mouse ARP6 antibody [40], 1:2,000 dilution for peroxidase anti-peroxidase (PAP) soluble complex antibody (Sigma-Aldrich) that detects the TAP-tag, 1:1000 dilution for H2A.Z antibody [41], and 1:500 dilution for H4Ac antibody (Millipore 06–866)) in blocking solution for 1 hour at RT. The blots were washed 3 times for 5 minutes in PBST. The ARP6 blot was then incubated with the anti-mouse horseradish peroxidase (HRP)-conjugated secondary antibody (1:2,000 dilution, GE Healthcare), while H2A.Z and H4Ac blots were incubated with anti-rabbit HRP-conjugated secondary antibody (1:3,000 dilution, GE healthcare). The ARP6, H2A.Z, and H4Ac blots were washed 3 more times for 5 minutes in PBST, and all blots were then incubated with ECL detection reagents for 2 minutes (Thermo Scientific). The ARP6 blots and PAP blot were exposed to Amersham Hyperfilm ECL (GE Healthcare) to detect protein bands, while H2A.Z and H4Ac blots were scanned for chemiluminescence signal using ChemiDoc MP imaging system instrument (BioLab).
Approximately 150 mg of young leaves from WT, mbd9-1, and arp6-1 plants were ground to a fine powder, homogenized in 3 ml of histone extraction buffer (0.25 M sucrose, 1 mM CaCl2, 15 mM NaCl, 60 mM KCl, 5 mM MgCl2, 15 mM PIPES pH 7.0, 0.5% Triton X-100 with protease inhibitors cocktail (Roche) and 10 mM sodium butyrate), filtered through the 70 micron strainer, and incubated for 15 minutes at 4°C on a nutator. After centrifugation at 4,500g for 20 minutes at 4°C, the pellets were resuspended in 500 μl of 0.1 M H2SO4 and incubated overnight at 4°C on a nutator. After centrifugation for 10 minutes at 17,000g, total proteins were precipitated from the supernatant with concentrated trichloroacetic acid to a final concentration of 30%. The protein pellets were washed twice with an acetone solution containing 0.1% HCl and then once with acetone. The protein pellet was then air-dried and resuspended in 50 μl of 1x Laemmli’s sample buffer.
Myc-MBD9 fusion protein, cloned into the pT7CFEChis plasmid, was expressed using the 1-step human coupled in-vitro translation (IVT) kit (Thermo Scientific) following manufacturer’s recommendations. In parallel, a GFP protein (pT7CFEcHis plasmid carrying GFP; included in the kit) was expressed and served as a positive control for IVT reaction. The expression of GFP was confirmed by visualizing GFP on a fluorescence microscope (Olympus) using 2 μl of an IVT reaction, while the expression of Myc-MBD9 was confirmed by western blotting using 4 μl of an IVT reaction and an anti-myc antibody to detect the 245 kDa protein (S12 Fig).
Two MODified peptide array slides (Active Motif) were first briefly washed in PBST buffer and then blocked in 5% milk-PBST buffer for 1 hour at RT. After blocking, the slides were washed three times for 5 minutes in PBST. One slide was then incubated with 16 μl of the IVT Myc-MBD9 protein, while the other slide was incubated with 16 μl of the IVT GFP protein, both diluted in 8 ml of the binding buffer (50 mM HEPES pH 7.5, 100 mM NaCl, 1 μM DTT, and 50% glycerol) overnight at 4°C on a nutator. Next day, the slides were washed three times for 5 minutes in PBST buffer and then incubated with a primary anti-myc antibody (1:2,000 dilution, Abcam) for 1 hour at RT. The slides were washed again three times for 5 minutes in PBST and then incubated with anti-mouse HRP-conjugated secondary antibody (1:5,000 dilution, GE Healthcare) for 30 minutes at RT. The slides were incubated with ECL detection reagents for 2 minutes (Thermo Scientific) and scanned with the ChemiDoc MP imaging system (BioRad) to detect the spot signals. Two biological replicates of histone peptide array assays were performed for each protein.
The chemiluminescence signal intensities of each spot on histone peptide array slides were quantified using The Array Analyze Software available from the manufacturer (Active Motif) and were assigned the values ranging from 0 (minimum, no signal) to 1 (maximum, strongest signal). The strongest signal for each array was observed at spot “P21” where the myc-tag peptide is spotted given that an anti-myc antibody was used to detect Myc-MBD9 binding. Since MBD9 is not expected to interact with the H2B peptide (spot “P4” on the array) due to differences in the N-terminal amino acid sequences between the human H2B and Arabidopsis H2B [80], the average signal intensity of the “P4” spot was subtracted from the average intensity values for each spot for each array. The average background intensities of the “P4” spots from two arrays (S12A and S12B Fig) were 0.1145 and 0.2605, respectively. After the background subtraction, only the normalized average spot intensities with values higher than 0.3 were presented in Fig 5, along with the spot intensities for H3K4me3 and H3K27me3.
The ARP6-TAP-containing protein complex was purified as described in [45], with the following modifications: 1) instead of using a kitchen blender in a stainless steel wine cooler, 50 grams of frozen seedlings were ground with a mortar and pestle in liquid nitrogen, and 2) all washing steps of the IgG-Sepharose and streptavidin-Sepharose Poly-Prep columns were performed using a peristaltic pump at a flow rate of 1 ml/min at 4°C.
On-bead digestion of the TAP-purified proteins was performed as previously reported [81, 82]. Residual wash buffer was removed and 200 μl of 50 mM NH4HCO3 was added to each sample. Samples were reduced with 1 mM dithiothreitol for 30 minutes and alkylated with 5mM iodoacetamide in the dark for an additional 30 minutes. Both steps were performed at room temperature. Digestion was started with the addition of 1 μg of lysyl endopeptidase (Wako) for 2 hours and further digested overnight with 1:50 (w/w) trypsin (Promega) at room temperature. Resulting peptides were acidified with 25 μl of 10% formic acid (FA) and 1% trifluoroacetic acid (TFA), desalted with a Sep-Pak C18 column (Waters), and dried under vacuum.
Liquid chromatography coupled to tandem mass spectrometry (LC-MS/MS) on an Orbitrap Fusion mass spectrometer (ThermoFisher Scientific, San Jose, CA) was performed at the Emory Integrated Proteomics Core (EIPC) [81, 82]. The dried samples were resuspended in 10 μl of loading buffer (0.1% formic acid, 0.03% trifluoroacetic acid, 1% acetonitrile). Peptide mixtures (2 μl) were loaded onto a 25 cm x 75 μm internal diameter fused silica column (New Objective, Woburn, MA) self-packed with 1.9 μm C18 resin (Dr. Maisch, Germany). Separation was carried out over a 140-minute gradient by a Dionex Ultimate 3000 RSLCnano at a flow rate of 300 nl/min. The gradient ranged from 3% to 99% buffer B (buffer A: 0.1% formic acid in water, buffer B: 0.1% formic in ACN). The spectrometer was operated in top speed mode with 3 second cycles. Full MS scans were collected in profile mode at 120,000 resolution at m/z 200 with an automatic gain control (AGC) of 200,000 and a maximum ion injection time of 50 ms. The full mass range was set from 400–1600 m/z. Tandem MS/MS scans were collected in the ion trap after higher-energy collisional dissociation (HCD). The precursor ions were isolated with a 0.7 m/z window and fragmented with 32% collision energy. The product ions were collected with the AGC set for 10,000 and the maximum injection time set to 35 ms. Previously sequenced precursor ions within +/- 10 ppm were excluded from sequencing for 20s using the dynamic exclusion parameters and only precursors with charge states between 2+ and 6+ were allowed.
All raw data files were processed using the Proteome Discoverer 2.1 data analysis suite (Thermo Scientific, San Jose, CA). The database was downloaded from Uniprot and consists of 33,388 Arabidopsis thaliana target sequences. An additional sequence was added for the TAP-tagged bait protein. Peptide matches were restricted to fully tryptic cleavage and precursor mass tolerances of +/- 20 ppm and product mass tolerances of +/- 0.6 Daltons. Dynamic modifications were set for methionine oxidation (+15.99492 Da) and protein N-terminal acetylation (+42.03670). A maximum of 3 dynamic modifications were allowed per peptide and a static modification of +57.021465 Da was set for carbamidomethyl cysteine. The Percolator node within Proteome Discoverer was used to filter the peptide spectral match (PSM) false discovery rate to 1%.
For each sample, 1.5 grams of 6-days-old seedlings (with roots removed) were used for ChIP-seq and ChIP-qPCR experiments. The ChIP-seq experiments were performed in biological triplicates on WT, arp6-1, and mbd9-1 seedling tissues as described previously [83], with the following modifications: 1) after the centrifugation of the nuclei in extraction buffer 3, the pellets were resuspended in 210 μl of nuclei lysis buffer, 2) after aliquoting out 10 μl for input, the fragmented chromatin from each sample was split into two equal volumes (100 μl each) and diluted with 1ml of ChIP-dilution buffer, and 3) one diluted half was incubated with 1.5 μg of the affinity-purified polyclonal H2A.Z antibody [41] while the other half was incubated with 5 μl of H4Ac polyclonal antibody serum (Millipore cat.# 06–866). The ChIP-qPCR experiments were performed in duplicates on WT, arp6-1, mbd9-1, mbd9-2, mbd9-3, and arp6-1;mbd9-1 seedling tissues as described previously [83], with the following modifications: 1) after the centrifugation of the nuclei in extraction buffer 3, the pellets were resuspended in 105 μl of nuclei lysis buffer, and 2) after sonication using a Diagenode Bioruptor, 5 μl of the fragmented chromatin was used for input, while 100 μl was diluted with 1 ml of the ChIP dilution buffer. The whole solution was then used for incubation with 1.5 μg of the affinity-purified polyclonal H2A.Z antibody. The ChIP and input DNA samples from the ChIP-qPCR experiment were analyzed by real-time PCR using the ACT2 (At3g18780) 3’ untranslated region sequence as the endogenous control, and with primers that span the genomic regions of FLC (At5g10140), ASK11 (At4g34210), and At4 (At5g03545) genes (S4 Fig), and of AT2G34202, AT1G62760, AT5G28410, AT4G15960, AT1G76740, and AT3G21240 genes (S5 Fig). The sequences of the primers used in S6 Fig were previously described [41, 55], while the sequences of the primers used in S8 Fig are listed in S3 Table.
Libraries were prepared starting with 500 pg of ChIP or input DNA samples using the Swift Accel-NGS 2S Plus DNA library kit (Swift Biosciences) according to the manufacturer’s instructions. All libraries were pooled and sequenced using single-end 50 nt reads on an Illumina NextSeq 500 instrument. Reads were mapped to the Arabidopsis thaliana genome (TAIR10) using the Bowtie2 package [84]. Quality filtering and sorting of the mapped reads, as well as removal of the reads that mapped to the organellar genomes, was done as previously described [85] using Samtools 1.3.1 [86]. The filtered and sorted BAM files were converted to bigwig format as previously described [85] using deepTools 2.0 software [87]. Correlation plots for the different H2A.Z or H4Ac samples were computed and plotted using the “multiBamSummary” and “plotCorrelation” functions in the deepTools package.
For visualization, for a given antibody, BAM files of each genotype were scaled to the same number of reads. This was done using the “samtools view -c” and “samtools view -s” commands to count the number of reads in a BAM file and to scale down the global read amounts in a BAM file, respectively. Three scaled, replicate BAM files of each genotype for H2A.Z were combined and converted to a single bigwig file for each genotype. The same was done for each genotype of H4Ac. Average plots and Heatmaps displaying ChIP-seq data were generated using the SeqPlots app [88].
Peak calling on ChIP-seq data was done by employing the “Findpeaks” function of the HOMER package [89] using the input ChIP-seq files as reference and the “-region” option to identify sites of read density enrichment. Called peaks were processed using Bedtools [90] to identify peaks called in at least one other replicate for a given genotype. This was done by keeping any peaks that overlapped by at least 50% between biological replicates. The retained peaks were concatenated and then merged together if they overlapped by at least 200 base pairs.
The amount of H2A.Z reads in WT, mbd9-1, and arp6-1 plants present in H2A.Z-enriched peaks reproducibly identified in WT plants was quantified using HTSeq’s htseq-count script [91]. Three replicates of counted reads for all three genotypes were then processed using DESeq2 [92]. MBD9-dependent and MBD9-independent peaks were determined from the DESeq2 results comparing wild type and mbd9-1 counted peaks. Peaks that had a log2 fold change of 0.6 or more and an adjusted p-value less than or equal to 0.05 were designated as MBD9-dependent H2A.Z sites. Peaks with an absolute log2 fold change less than 0.25 were designated as MBD9-independent H2A.Z sites.
The PAVIS web tool [93] was used to determine the genomic distribution of H2A.Z ChIP-seq peaks. PAVIS annotates each peak to a genomic feature using the center of the peak. The “upstream” regions were defined as 2,000 bp upstream of the annotated transcription start site, and “downstream” regions were defined as 1,000 bp downstream of the transcription end site.
Genes nearest to the MBD9-dependent and MBD9-independent H2A.Z sites were identified using the “TSS” function of the PeakAnnotator 1.4 program [94] as previously described [85].
Gene ontology (GO) analysis was carried out on gene lists from S2 Table using two different GO web tools: 1) the AgriGO GO Analysis Toolkit, with default parameters [95, 96], and 2) Gene Ontology enrichment analysis [97]. GO terms that had a false discovery rate (FDR) of 0.05 or less were considered significant.
Publicly available RNA-seq FPKM values for genes nearest to the MBD9-dependent and MBD9-independent H2A.Z sites were plotted using ggplot2 [98]. Unpaired t-tests were used to determine whether FPKM values were significantly different between the two sets of genes, with P values less than or equal to 0.05 considered as statistically significant.
We performed student t-tests to calculate the significance of fold changes in H2A.Z levels at three MBD9-dependent and three MBD9-independent sites between WT and mbd9-1 plants (S8 Fig). Based on our ChIP-seq results, we only expect a reduction in H2A.Z levels at MBD9-dependent sites in mbd9-1 plants compared to WT. Therefore, we performed a one-tail t-test to calculate the significance of H2A.Z depletion at MBD9-dependent sites (S8A Fig). At MBD9-independent sites, however, we expect that H2A.Z levels may vary in either direction in mbd9-1 plants versus WT. Thus, for MBD9-independent sites we performed a two-tail t-test analysis of qPCR results. Additionally, when the two-tailed t-test is performed on MBD9-dependent ChIP-qPCR results two out of three probed sites (AT2G34202 and AT1G62760) still show statistically significant depletion of H2A.Z levels.
Real-time PCR was performed on the Applied Biosystems StepOnePlus real-time PCR system using SYBR Green as a detection reagent. The 2-ΔΔCt method [99] of relative quantification was used to calculate the fold enrichment. The results presented for ChIP-qPCR experiments are averaged relative quantities from two biological replicates ± SD.
Total RNA was isolated from 6-day old seedlings (with roots removed) using the RNeasy plant mini kit (Qiagen). 2 μg of total RNA was converted into cDNA with Super-Script III first strand synthesis kit (Invitrogen). The cDNAs were used as templates for real-time PCR and ran on StepOnePlus real-time PCR system (Applied Biosystems) using SYBR Green as a detection reagent. The 2-ΔΔCt method [99] of relative quantification was used to calculate the fold enrichment. The PP2A gene (AT1G13320) was used as the endogenous control [100]. The primer sequences used to measure relative expression levels of PIE1, ARP6, MBD9, SWC4, SWC6, YAF9a, HTA8, HTA9, and HTA11 in WT, arp6-1, and mbd9-1 plants are listed in S3 Table.
SEC was performed on the HiPrep 16/60 Sephacryl S-400 HR column (GE Healthcare) equilibrated with SEC buffer (the same extraction buffer as described in [45], but without NP-40 detergent). A mixture of protein standards ranging from 669 to 44 kDa (GE Healthcare), resuspended in the SEC buffer, were run on the column to produce a calibration curve of molecular weights versus elution volumes. The slope equation of the calibration curve was then used to calculate the molecular weight of the peak ARP6 SEC fractions. Total protein extracts were isolated from 1 gram of the WT and mbd9-1 seedling tissues (without roots) using the same extraction buffer that was used for the ARP6-TAP-tag protein complex purification [45]. For each run, between 1.8 and 2.0 ml of the protein extract was loaded onto the column and 1 ml fractions were collected. For each sample, two biological replicates of the SEC experiments were performed and gave nearly identical results.
Raw data from ChIP-seq experiments performed on young WT seedlings using antibodies against H3K4me3 (GSM2544796, [101]), H3K36me3 (GSM2544797, [101]), H3K9me2 (GSM2366607, [102]), H3K9Ac (GSM2388452, [103]), H3K18Ac (GSM2096925, [104]), H3K27Ac (GSM2096920, [104]), H3K27me3 (GSM2498437, [105]), and H2AK121ub (GSM2367138, [105]), were processed and analyzed with the same procedures as for our ChIP-seq data (see above) and used to generate the average plots presented in Fig 4. The FPKM values from two different RNA-seq experiments (GSM2752981 and GSM2367133, respectively, [105, 106]) were used to compare expression levels in WT plants of genes nearest to the MBD9-dependent and MBD9-independent H2A.Z sites.
All ChIP-seq data generated in this study have been deposited to the NCBI GEO database under accession number GSE117391.
|
10.1371/journal.pntd.0004810 | Relationship between Community Drug Administration Strategy and Changes in Trachoma Prevalence, 2007 to 2013 | Australia is the only high income country with persisting endemic trachoma. A national control program involving mass drug administration with oral azithromycin, in place since 2006, has some characteristics which differ from programs in low income settings, particularly in regard to the use of a wider range of treatment strategies, and more regular assessments of community prevalence. We aimed to examine the association between treatment strategies and trachoma prevalence.
Through the national surveillance program, annual data from 2007–2013 were collected on trachoma prevalence and treatment with oral azithromycin in children aged 5–9 years from three Australian regions with endemic trachoma. Communities were classified for each year according to one of four trachoma treatment strategies implemented (no treatment, active cases only, household and community-wide). We estimated the change in trachoma prevalence between sequential pairs of years and across multiple years according to treatment strategy using random-effects meta-analyses.
Over the study period, 182 unique remote Aboriginal communities had 881 annual records of both trachoma prevalence and treatment. From the analysis of pairs of years, the greatest annual fall in trachoma prevalence was in communities implementing community-wide strategies, with yearly absolute reductions ranging from -8% (95%CI -17% to 1%) to -31% (-26% to -37%); these communities also had the highest baseline trachoma prevalence (15.4%-43.9%). Restricting analyses to communities with moderate trachoma prevalence (5–19%) at initial measurement, and comparing community trachoma prevalence from the first to the last year of available data for the community, both community-wide and more targeted treatment strategies were associated with similar absolute reductions (-11% [-8% to -13%] and -7% [-5% to -10%] respectively). Results were similar stratified by region.
Consistent with previous research, community-wide administration of azithromycin reduces trachoma prevalence. Our observation that less intensive treatment with a ‘household’ strategy in moderate prevalence communities (5-<20%) is associated with similar reductions in prevalence over time, will require confirmation in other settings if it is to be used as a basis for changes in control strategies.
| Australia is the only high income country with persisting endemic trachoma and a national control program has been in place since 2006. The program involves annual screening of children for trachoma in communities designated to be at high risk of disease and treatment of those affected with the antibiotic azithromycin. Depending on the level of trachoma detected in children, antibiotic treatment is also given to households and other community members. We used data collected annually from 2007 to 2013 to examine what effect the extent of azithromycin treatment had on subsequent levels of trachoma in children aged 5–9 years. We found that in communities with high levels of trachoma, when all community members received azithromycin (community-wide treatment), the greatest reduction in trachoma level was achieved. However in communities with moderate levels of trachoma, using either community-wide treatment or more targeted (household) treatment resulted in equivalent reductions in trachoma. This observation needs to be confirmed in other studies before changes to current recommendations regarding trachoma control strategies are considered.
| Trachoma, caused by serotypes of Chlamydia trachomatis is a major cause of blindness globally.[1] In 1997 The Alliance for the Global Elimination of Blinding Trachoma by 2020 (GET 2020) initiative was launched. Supported by the World Health Organization (WHO), the alliance promotes its goal of elimination through the SAFE strategy, with its key components of surgery to correct trichiasis (S), antibiotic treatment (A), facial cleanliness (F) and environmental improvements (E).[1] Randomised controlled trials have shown that antibiotics, either topical or oral, are effective for treatment.[2] There is a more limited body of trial evidence that has been used to support the strategy of mass drug administration (MDA), or whole community treatment, which is one of the main components of the SAFE strategy in many countries. There have been few comparisons of alternative community treatment strategies, and relatively limited follow up studies of long term trends following implementation of prevention programs.[3,4] Evidence of effective treatment strategies across a range of prevalence settings will become increasingly important as more countries approach the goal of trachoma elimination.
Australia is the only high-income country with endemic trachoma.[5] The disease occurs primarily in remote Aboriginal communities in three Australian jurisdictions, the Northern Territory (NT), South Australia (SA), and Western Australia (WA), although it has also been identified in Queensland and New South Wales.[6,7] In 2013 overall prevalence among children aged 5–9 years in endemic areas was estimated to be 4% with substantial variation between communities; an estimated 50% of communities had no clinically detectable trachoma and 8% had hyperendemic levels (>20%).[8]
Since 2006 the Australian Government has funded control programs based on regular mapping of trachoma prevalence in endemic areas. Trachoma management has been based on guidelines first endorsed in 2006 [9] and revised in 2014.[10] Unlike the WHO guidelines, the 2006 Australian guidelines recommended screening every community considered at risk annually, regardless of trachoma prevalence, as well as a tiered approach to antibiotic treatment depending on trachoma prevalence (see Table 1). Australia therefore has an opportunity to examine the impact of different treatment strategies, in more detail than has been possible in other trachoma endemic settings, where only MDA has been used, and prevalence is generally monitored at much longer intervals. We report here an analysis based on routinely collected trachoma prevalence data over seven years in Australia’s endemic areas. These data have the potential to inform trachoma control programs both in Australia and internationally.
Since 2006 when the National Trachoma Management Program was initiated, screening for trachoma and management has been consistently undertaken in three trachoma-endemic Australian jurisdictions, the NT, SA and WA. At program initiation, each jurisdiction identified Aboriginal communities considered to be at high risk of endemic trachoma from historical prevalence data and local knowledge, and over time, additional communities have been added to those considered at risk. In each designated community, regular screening rounds were undertaken over short time periods (generally several days), involving external teams working with local health staff. In most communities, 5–9 year olds were the focus of screening, as they were mostly in school, although children aged under 5 or 10–14 years were also screened if present at the time. Between 2010 and 2013 the estimated proportion of children resident in communities aged 5–9 that were screened for trachoma ranged from 57–71%.[6,8,11] Of those screened the WHO trachoma grading criteria[12] were used to diagnose and classify trachoma. Data from each community screened were collected on standardised data collection forms and included the numbers of children screened (in age groups 1–4, 5–9, 10–14 years), with active trachoma and with clean faces. The treatment strategy undertaken in the community and, from 2011, the numbers of household members and other community members treated, were also recorded. Data from screened communities have been presented in annual reports by the National Trachoma Surveillance and Reporting Unit and form the basis for analyses reported here.[6]
De-identified community-based data were obtained for each year from 2007, when comprehensive data collection began, through to 2013. As the majority of trachoma screening in communities was undertaken through primary school programs targeting 5–9 year olds, we restricted analyses of prevalence to this age group. The unit for analysis was a single episode of screening in 5–9 year olds within a single community. Community trachoma prevalence was estimated by dividing the number of 5–9 year olds with active trachoma during a screening round by the number screened. The treatment strategy adopted for a community in a given year was classified into one of four categories according to what was reported in the national database: no treatment; “active”cases only treated; “household” treatment under which active cases and their households members were treated; and “community-wide” treatment which covered both whole-of-community treatment (also known as “mass drug administration”) and a strategy under which active cases, household members and all children aged <15 years in the community were treated. For all strategies, the treatment administered for those over 6 months of age was a single weight-based dose (20mg/kg) of oral azithromycin [9].
Descriptive analyses by calendar year, examining all communities with eligible screening episodes were conducted. From 2011 onwards treatment coverage in communities for which household or community-wide treatment strategies were recorded was calculated by summing the population aged 0–14 years recorded as treated with azithromycin, and dividing by the total estimated population aged 0–14 years according to both census[13] and local health worker community population estimates.
To compare treatment strategies, we undertook two analyses. First we identified all communities for which data on trachoma prevalence in 5–9 year olds were available for pairs of consecutive calendar years. For each such pair of years, we estimated the change between the years in community prevalence, by simple difference. We then grouped communities by the first year of the consecutive pair and by the treatment strategy recorded in that year, and calculated a combined estimate of change for each treatment strategy using a random effects meta-analysis.[14] Second, for each community with at least two years of screening data, regardless of whether they were consecutive, we compared the change in prevalence from the first and final year of available data according to broad categories of treatment strategy (never treated, any non-community-wide treatment, treated at least once with community-wide), using the same meta-analytic method. As the treatment strategy used was strongly influenced by trachoma prevalence,[9] we conducted sensitivity analyses restricting communities to those with moderate prevalence (≥5% to <20%) at the start of the interval. We also stratified results by the two jurisdictions contributing the majority of data (NT and WA), and by community size (<250 versus ≥250 people) based on 2011 Australian census estimates.[13] Finally we conducted post-hoc analyses only including data collected for years 2007 to 2010 with the goal of differentiating secular trends in trachoma prevalence from effects of treatment. We used RevMan 5.5 software to estimate absolute differences in trachoma prevalence and SAS (version 9.3) for estimates of trachoma prevalence (function was unavailable in RevMan). The DerSimonian and Laird random effects model was used to obtain pooled estimates of risk difference, using the Mantel-Haenszel method to estimate the variation between studies. We estimated the combined prevalence using an exact likelihood approach.[15]
Administrative approvals for the data collection and analyses reported here were provided by the health departments of the three jurisdictions involved. Ethical approval was by the University of New South Wales Human Research Ethics Committee (ref 9-14-042).
We identified 914 screening episodes from 215 unique remote Aboriginal communities with children aged 5–9 years screened at least once between 2007 and 2013. Of the 215 communities, the majority were in the NT (n = 90; 42%) and WA (n = 99; 46%). There were 33 communities screened only once, 46 had 2–3 episodes, 59 had 4–5 episodes, and 77 had 6–7. The communities screened less frequently were more likely to have been screened for the first time more recently, with the median year of screening for communities with 3 or fewer years of screening data being 2012 compared to 2010 for those with four or more years of data.
Table 2 shows the number of communities screened each year, the proportion of communities screened from each of the three jurisdictions, the median number of children screened, the trachoma prevalence in 5–9 year olds and treatment strategies used. Biannual treatment (a second dose of antibiotics administered in the same year) was recorded following 1% of screening episodes. As biannual treatment was not unique to a particular treatment strategy, and numbers were small, we did not include this as a separate treatment classification. In general, the number of communities screened increased until 2013 when the NT adopted the revised guideline for screening[10] which recommends that screening in communities with high trachoma prevalence takes place every 3 years rather than annually. The median number of 5–9 year old children screened per community remained relatively stable over the seven years numbering about 20 (IQR 10–38). From 2008 to 2013, the proportion of communities with no trachoma detected increased (from 22.8% to 60.3%) while the proportion of communities with trachoma prevalence above 5% decreased (from 67.5% to 27.6%). In communities with trachoma detected, the median prevalence also decreased, from 23.1% to 8.9%. There was an increase in the proportion of communities not treated from 25.7% in 2008 to 56.0% in 2013, while from 2009 there was a fall in the number of communities treating ‘active’ cases only.
For 2011 and 2012 (Table 3), using local estimates of the population size, treatment coverage among 120 communities that reported having used a “household” treatment strategy was 11.9% (95%CI 11.4–12.5%) while for the 33 communities using a “community-wide” treatment strategy, treatment coverage was 75.0% (95%CI 73.5–76.4%). The estimates were similar when Census population estimates[13] were used.
After excluding the 33 communities with only a single year of screening data available, there remained 881 records from 182 unique communities; 77 (42.3%) were from the NT, 21 (11.5%) from SA and 84 (46.2%) from WA. Of 121 communities that applied a treatment strategy in more than one of the years observed, 89 were recorded as having changed strategies over the time period, 30 communities used only household treatments, two used only community-wide treatments, and none applied the “active” case only strategy more than once.
Fig 1 shows the estimated change in trachoma prevalence between pairs of successive years, according to the treatment strategy used in the initial year of the pair. Communities recorded as receiving no treatment are separated according to whether they had trachoma detected or not in the initial year of the pair. In the earlier years of the program (2007–2010) for communities without trachoma detected and not treated, there was a significant increase in prevalence between pairs of years (e.g. absolute risk increase of 10% [95%CI 3% to 16%] from 2007 to 2008) but after 2010 there was no substantial change. The number of communities that were not recorded as having been treated despite trachoma being detected decreased over time. In these communities, trachoma prevalence between years did not change significantly between pairs. For all categories of treated communities (active case only, household, or community-wide) there was a reduction in estimated trachoma prevalence between the pairs of years; in most years this was not statistically significant. The largest absolute reductions in trachoma prevalence were in communities that were recorded as having received community-wide treatment, with point estimates ranging from -8% to -31%; the reductions were only statistically significant for the years 2007–2011. These communities receiving community-wide treatment also had the highest prevalence in the earlier of the paired years (range for combined estimates 15.4%-43.9%).
The majority of communities (n = 176) had annual records of trachoma screening that included at least two years of the eligible period (2007 to 2013), including 68 with data for 7 consecutive years, 47 with 6 years, 25 with 5 years, 12 with 4 and 24 with less than 4. Based on the treatment strategies used between the first and final year of data recorded, communities were grouped into three categories (Table 4): those never treated; those treated but never with community-wide strategies (i.e. only active case or household treatment); and those treated at least once using a community-wide strategy.
Fig 2 shows the estimated change in trachoma prevalence between the first and final years of data, by the three classifications of communities in Table 4. For communities never recorded as receiving azithromycin for trachoma, the estimate of prevalence in the first year of screening was 0.1% and the estimated absolute reduction over time 0% (-3% to 2%). For communities that received only active case and household strategies, the prevalence in the first year of screening was 5.8% and the reduction over time -4% (-2% to -6%). Among communities treated at least once with community-wide strategies, initial prevalence was 23.9%, and the reduction -21% (-16% to -26%). These patterns were similar when communities from the NT or WA were considered separately.
When we restricted analyses to communities with moderate trachoma prevalence (5-<20%) in their first year of screening (Fig 2) we found that there was only a small difference between those that had received community-wide treatment and those that had not, with reductions of -11% (-8% to -13%) and -7% (-5% to -10%) respectively, from a similar initial prevalence (11.5% and 10.1% respectively). Restricting analyses to communities with at least four years of screening data did not change the findings, and we found no differences in treatment effects when we compared smaller (N<250 people) to larger sized communities (N≥250 people).
Analyses of data restricted to 2007–2010, the period during which there were substantial increases in trachoma prevalence in previously trachoma-free communities that were untreated (see Fig 1), are shown in Table 5. In the communities that were never treated, and in those treated but not with a community wide strategy, there was no significant fall in trachoma prevalence, while communities with at least one community-wide treatment had a 14% (95%CI 9 to 20%) decline in trachoma prevalence. When we further restricted analyses to communities with moderate trachoma prevalence (5-<20%) we found that those with at least one community-wide treatment had a significant reduction in prevalence but those treated with more targeted strategies did not.
In this investigation of the relationship between different community treatment strategies for trachoma control and long-term changes in trachoma prevalence, we found that in high prevalence communities, community-wide administration of azithromycin, or MDA, was associated with a substantially reduced trachoma prevalence after one year or more. In settings with moderate trachoma prevalence (5-<20%), more limited strategies were equally effective in the longer term. As discussed in more detail below, this finding may have particular relevance for countries moving towards elimination, but with localised areas of moderate prevalence remaining.
Observational studies have shown that a single MDA of azithromycin in communities with endemic trachoma results in substantial reductions at one year in trachoma prevalence in both hyperendemic (>20% prevalence)[16,17] and moderately endemic (5-<20% prevalence)[3,18] communities. Recent trials have compared annual versus biannual mass azithromycin administration in high prevalence communities, but the trials have not consistently found that larger or more sustained reductions can be achieved with more frequent treatment.[19,20] There are few reports comparing different treatment strategies in moderate prevalence settings. One study found that targeted (household) treatment may be as effective as mass treatment of all children but only had follow-up for 6 months.[21] Another suggested a single mass drug administration may be effective in sustaining a reduction in trachoma prevalence over many years[3] and another suggested that treatment that was not community-wide led to re-infections.[22] Our findings regarding ‘community-wide’ treatment (Fig 1) over one year concur with the observational studies of mass drug administration showing that this approach is effective in substantially reducing trachoma prevalence in high prevalence settings. Our main analyses also suggest that in more moderate prevalence settings, targeted treatment strategies (mostly ‘household’ strategies, whereby active cases and all members of their household were treated with azithromycin), were also associated with reduced trachoma over a year and for longer periods (Figs 1 and 2).
In the paired-year analyses (Fig 1), among communities that had no trachoma detected at the start of the observation period, and were consequently not treated, we found that there were annual increases in prevalence between 2007 and 2010. However from 2011 onwards, prevalence remained at zero, i.e. no change. Given the mobility of people between Aboriginal communities,[23] this observation may be evidence that antibiotic treatment programs in communities with trachoma detected can have a “herd” effect, in that transmission to trachoma-free communities is prevented. It is also possible that this resulted from other components that are delivered as part of the SAFE strategy, such as promotion of facial cleanliness and environmental improvements.
As the paired-year analysis indicated no overall increase in trachoma prevalence in trachoma-free communities from 2010 onwards, we undertook analyses involving multiple years with the goal of distinguishing effects of treatment from temporal changes in trachoma. In these sensitivity analyses, only communities receiving community-wide treatment were found to have a reduction in trachoma prevalence (see Table 5). It is therefore possible that the trachoma reduction in moderate prevalence settings observed in our primary analyses may in fact have been a result of overall declines in trachoma burden rather than a result of targeted treatment.
This was an observational study using routinely collected surveillance data.[5] While diagnosis was undertaken by specialised teams of health care workers following standard international guidelines, there may have been diagnostic error, to an extent that cannot be measured. Our analysis of impact was also limited by the absence of detailed data for all years on the level of treatment coverage achieved in each community. However for 2011 and 2012, data were available for the majority of communities and this indicated that coverage was substantially different between those communities reporting “household” compared to those reporting “community-wide” treatment. There may also be factors that differed between communities or changes over time that were not measured but were associated with treatment strategy and therefore could have affected the summary estimate of difference in trachoma prevalence observed. For example, we did not include in our analyses other factors that may contribute to changes in trachoma prevalence.[1] Facial cleanliness, and facial cleanliness promotion (‘F’ in the SAFE strategy) was reported in the communities screened, but not considered to be sufficiently standardised or validated to use in the analyses presented here.[24] Data on environmental factors (‘E’ in the SAFE strategy) such as improved housing conditions, or the availability of swimming pools, were limited and inconsistent.[25] Despite the absence of information on facial cleanliness and environmental factors, we do not have evidence to suggest that they linked to treatment status and thus had any potential to bias our results.
The strengths of our study are the use of annual trachoma screening data from all communities in the three jurisdictions with known endemic trachoma leading to a more comprehensive picture of not only the effects of different treatment strategies on single communities, but also programmatic effects on all communities in a real-world setting. We also had observations for the majority of communities over a significant period of time (at least four years) and were able to observe the effects of a targeted treatment strategy in a moderate prevalence setting.
In summary, our study supports current evidence that recommends mass azithromycin administration to reduce trachoma prevalence in high prevalence settings. We also found that less intensive treatment with a “household” strategy in moderate prevalence communities (5-<20%) may be associated with reductions in prevalence similar to mass drug administration. This finding may have implications for countries that are moving to lower levels of endemic trachoma and wish to reduce the amount of azithromycin being used. The strategy does however have the requirement that individual examination must take place, to determine which households have affected members. If a targeted approach is to be considered, trials and health economic analyses are required to determine which option may be more cost-effective in particular programmatic and community contexts.[26] Finally our results also suggest that trachoma program implementation can reduce trachoma prevalence in communities not specifically targeted (“herd effects”) and thereby contribute to reducing trachoma transmission.
|
10.1371/journal.pntd.0002578 | Knockdown of Asparagine Synthetase A Renders Trypanosoma brucei Auxotrophic to Asparagine | Asparagine synthetase (AS) catalyzes the ATP-dependent conversion of aspartate into asparagine using ammonia or glutamine as nitrogen source. There are two distinct types of AS, asparagine synthetase A (AS-A), known as strictly ammonia-dependent, and asparagine synthetase B (AS-B), which can use either ammonia or glutamine. The absence of AS-A in humans, and its presence in trypanosomes, suggested AS-A as a potential drug target that deserved further investigation. We report the presence of functional AS-A in Trypanosoma cruzi (TcAS-A) and Trypanosoma brucei (TbAS-A): the purified enzymes convert L-aspartate into L-asparagine in the presence of ATP, ammonia and Mg2+. TcAS-A and TbAS-A use preferentially ammonia as a nitrogen donor, but surprisingly, can also use glutamine, a characteristic so far never described for any AS-A. TbAS-A knockdown by RNAi didn't affect in vitro growth of bloodstream forms of the parasite. However, growth was significantly impaired when TbAS-A knockdown parasites were cultured in medium with reduced levels of asparagine. As expected, mice infections with induced and non-induced T. brucei RNAi clones were similar to those from wild-type parasites. However, when induced T. brucei RNAi clones were injected in mice undergoing asparaginase treatment, which depletes blood asparagine, the mice exhibited lower parasitemia and a prolonged survival in comparison to similarly-treated mice infected with control parasites. Our results show that TbAS-A can be important under in vivo conditions when asparagine is limiting, but is unlikely to be suitable as a drug target.
| The amino acid asparagine is important not only for protein biosynthesis, but also for nitrogen homeostasis. Asparagine synthetase catalyzes the synthesis of this amino acid. There are two forms of asparagine synthetase, A and B. The presence of type A in trypanosomes, and its absence in humans, makes this protein a potential drug target. Trypanosomes are responsible for serious parasitic diseases that rely on limited drug therapeutic options for control. In our study we present a functional characterization of trypanosomes asparagine synthetase A. We describe that Trypanosoma brucei and Trypanosoma cruzi type A enzymes are able to use either ammonia or glutamine as a nitrogen donor, within the conversion of aspartate into asparagine. Furthermore, we show that asparagine synthetase A knockdown renders Trypanosoma brucei auxotrophic to asparagine. Overall, this study demonstrates that interfering with asparagine metabolism represents a way to control parasite growth and infectivity.
| Asparagine is a naturally occurring non-essential amino acid found in many proteins. Due to its high nitrogen/carbon ratio, asparagine is likely to be linked to nitrogen homeostasis and protein biosynthesis [1]. AS is the protein involved in asparagine biosynthesis. There are two distinct types of AS, AS-A and AS-B, encoded by asnA and asnB genes, respectively. AS-A encoding genes have been reported in archaea [2], [3], prokaryotes [4]–[7], and in the protozoan parasite Leishmania [8]. The AS-B encoding gene is present in prokaryotes [5], [9] and also in eukaryotes, including mammalian cells [10], [11], yeasts [12], algae [13], and higher plants [14]. Both types catalyze the ATP-dependent conversion of aspartate into asparagine. While AS-B can use both ammonia and glutamine (reaction B) as amide nitrogen donors [5], [15]–[20], Escherichia coli (E. coli) AS-A was reported to be dependent strictly on ammonia (reaction A) [21], [22].
AS-A and AS-B share no sequence or structural similarities. Their three-dimensional structures provided important information concerning their distinct catalytic mechanisms [2], [23]–[25]. AS-A exists as a dimer where each monomer has a core of eight β-strands flanked by α-helices, resembling the catalytic domain of class II aminoacyl-tRNA synthetases such as aspartyl-tRNA synthetase [24]. AS-A synthesizes asparagine in two steps: the β-carboxylate group of aspartate is first activated by ATP to form an aminoacyl-AMP, followed by amidation by a nucleophilic attack with an ammonium ion [2]. The AS-B enzyme also forms a dimer, but each monomer contains two distinct domains, each of which contains a catalytic site. The N-terminal site catalyzes the conversion of glutamine into glutamic acid and ammonia, while aspartate reacts with ATP in the C-terminal site, generating the intermediate β-aspartyl-AMP [26], [27]. Similarly to other glutamine dependent amidotransferases, ammonia released in the N-terminal domain of the enzyme travels through an intramolecular tunnel connecting the active sites, and reacts with the reactive acyladenylate intermediate to produce asparagine [28].
An open reading frame encoding a putative AS-A is present in the genome of the protozoan parasites, Trypanosoma cruzi (T. cruzi) and Trypanosoma brucei (T. brucei) [29]–[31]. T. cruzi and T. brucei are transmitted to a mammalian host through an invertebrate vector, and are responsible for Chagas disease and African sleeping sickness, respectively. Disease control is dependent on drug therapy, but treatment options are limited, both by high toxicity and recent emergence of drug resistance [32]–[34]. Vaccines for T. brucei infections are unlikely to be developed not only because of extensive antigenic variation [35], but also because infections compromise host humoral immune competence [36].
Trypanosome AS-A might be a drug target due to the absence of a homologue in humans [8]. AS-A is important in other microorganisms. For example, asnA is an essential gene in Haemophilus influenzae (DEG10050178) [37], and is strongly up-regulated in Pasteurella multocida during host infection [38], and when Klebsiella aerogenes is grown in an amino acid-limited but ammonia rich environment [5]. We therefore undertook biochemical and genetic studies of AS-A in trypanosomes to ascertain its biological role and evaluate its potentiality as drug target.
All experiments involving animals were carried out in accordance with the IBMC.INEB Animal Ethics Committees and the Portuguese National Authorities for Animal Health guidelines, according to the statements on the directive 2010/63/EU of the European Parliament and of the Council. IL, JT and ACS have an accreditation for animal research given by the Portuguese Veterinary Direction (Ministerial Directive 1005/92).
Procyclic and bloodstream forms of Trypanosoma brucei brucei Lister 427 were used. Procyclic forms were grown in MEM-Pros medium supplemented with 7.5 µg/ml hemin, 10% fetal calf serum (FCS) and 100 IU/mL of penicillin/streptomycin at 27°C, with cell densities between 5×105 cells/ml to 1–2×107 cells/ml. Bloodstream forms were grown in complete HMI-9 medium (supplemented with 10% FCS and 100 IU/mL of penicillin/streptomycin) [39] in vented tissue culture flasks; these cultures were diluted when cultures reached the cell density of 2×106/ml and incubated in a humidified atmosphere of 5% CO2, at 37°C. Bloodstream RNAi cell cultures were supplemented with 7.5 µg/ml hygromycin and 0.2 µg/ml phleomycin.
T. brucei asparagine synthetase A (TbASA) and T. cruzi asparagine synthetase (TcASA) genes were obtained by performing PCR on genomic DNA from Trypanosoma brucei brucei TREU927 and Trypanosoma cruzi CL Brener Non-Esmeraldo-like. Fragments of the open reading frames of TbASA (Tb927.7.1110; chromosome Tb927_07_v4; 28861 to 289067) and TcASA (Tc00.1047053503625.10; chromosome TcChr29-P; 687159–688206) were PCR-amplified using a Taq DNA polymerase with proofreading activity (Roche). The sequences of the primers were as follows: sense primer 5′ - CTAATTACATATGGGCGACGACGGTTATTC - 3′ and antisense primer 5′ - CCCAAGCGAATTCTTACAACAAATTGTGC - 3′, sense primer 5′ - CAAT TTGCATATGACATCGGGAGATCC - 3′ and antisense primer 5′ - CCCAAGCAAGCTTTCACAGCAAGGG - 3′, respectively. PCR conditions were as follows: initial denaturation (2 min at 94°C), 35 cycles of denaturation (30 s at 94°C), annealing (30 s at 45°C) and elongation (2 min at 68°C) for TbAS-A, and annealing (30 s at 50°C) and elongation (2 min at 68°C) for TcAS-A, and a final extension step (10 min at 68°C). The PCR products were isolated from a 1% agarose gel, purified by the Qiaex II protocol (Qiagen), and cloned into a pGEM-T Easy vector (Promega) and sent to Eurofins MWG (Germany) for sequencing.
The TbASA and TcASA genes were subcloned into pET28a(+) expression vector (Novagen). The recombinant 6-His-tagged proteins were expressed in E. coli BL21DE3 by induction of log-phase cultures with 0.5 mM IPTG (NZYTech) for 3 h at 37°C (TcAS-A) and at 18°C, overnight (O/N) (TbAS-A). Bacteria were harvested and resuspended in buffer A [0.5 M NaCl (Sigma), 20 mM Tris.HCl (Sigma), pH 7.6]. The sample was sonicated and centrifuged to obtain the bacterial crude extract. The recombinant proteins were purified using Ni2+ resin (ProBond) and washing and elution with increasing levels (25 mM to 1 M) of imidazole (Sigma). The presence and purity of the recombinant protein in the several fractions was determined by SDS-PAGE and Coomassie staining. Dialysis was performed against PBS [137 mM NaCl (Sigma), 2.7 mM KCl (Sigma), 10 mM Na2HPO4.2H2O (Riedel-de Haën), 2 mM KH2PO4 (Riedel-de Haën) pH 7.4].
To generate rat and rabbit polyclonal antibodies against TbAS-A, each animal was first immunized with 150 µg of recombinant TbAS-A protein. After 2 weeks, 4 boosts with 100 µg of recombinant TbAS-A were given weekly. The collected blood samples were centrifuged to obtain the serum.
Extracts were obtained in RIPA buffer [(20 mM Tris-HCl (Sigma) (pH 7.5), 150 mM NaCl (Sigma), 1 mM Na2EDTA (Sigma), 1 mM EGTA (Sigma), 1% Nonidet P-40 (Sigma), 1% sodium deoxycholate (Sigma), 2.5 mM sodium pyrophosphate (Sigma), 1 mM β-glycerophosphate (Sigma), 1 mM Na3VO4 (Sigma)], with freshly-added complete protease inhibitor cocktail (Roche Applied Science). The total protein amount was quantified using Biorad Commercial Kit (Reagents A, B and S) and the samples were then kept at −80°C. For analysis of parasites from mice, trypanosomes were purified from mouse blood using a DE-52 (Whatman) column [40].
For Western blotting, 2 µg of recombinant TbAS-A and TcAS-A proteins, 10 µg of total soluble cell extract, or 1×107 parasites, were resolved in SDS/PAGE and transferred on to a nitrocellulose Hy-bond ECL membrane (Amersham Biosciences). The membrane was blocked in 5% (w/v) non-fat dried skimmed milk in PBS/0.1% Tween-20 (blocking solution), followed by incubation with an anti-His-tag rabbit antibody (MicroMol-413) (1∶5000) or a combination of an anti-TbAS-A rabbit antibody (1∶1000) with an anti-aldolase rabbit antibody (1∶5000) in blocking solution at 4°C O/N, respectively. Blots were washed with PBS/0.1% Tween-20 (3×15 min). Horseradish peroxidase-conjugated goat anti-rabbit IgG (Amersham) (1∶5000 for 1 h, at room temperature) in blocking buffer was used as the secondary antibody. The membranes were developed using SuperSignal WestPico Chemiluminescent Substrate (Pierce). ImageJ software (version 1.43u) was used for protein bands semi-quantification.
AS activity was assessed by quantification of asparagine formation [41]. The reactions were performed in a total of 150 µl of enzyme assay mixture in 85 mM Tris-HCl (Sigma) containing aspartate (Sigma), ammonia (Sigma), ATP (Sigma) and 8.4 mM Mg2+ (Sigma). Following incubation for set times at 37°C, enzymatic reactions were terminated by boiling 4 min, and then centrifuged at maximum speed for 30 s. 100 µl of the reaction mixture supernatant was added to 900 µl of ninhydrin 0.05% in ethanol. The resulting mixtures were boiled at 100°C for 5 min, then centrifuged for 30 s and maintained on ice. 300 µl of clear supernatant fluids were transferred to 96-well plates, and the absorbance at 340 nm determined [41]. Based on reaction linearity studies, 7.5 µg of enzyme and 15 min incubation at 37°C were selected as final conditions. To determine Kms, the concentrations of substrates were varied in the following ranges: 1.25–20 mM (aspartate), 0.78–50 mM (ammonia) and 0.62–10 mM (ATP), while the remaining substrates concentrations were in excess ([aspartate] >20 mM, [ATP] >10 mM, and [ammonia] >50 mM). Km for glutamine was determined using a concentration range of 1.5625–25 mM, while ATP and aspartate were maintained in excess. Measurements were performed in triplicate, and the initial rate was analyzed to obtain values of Vmax and Km by curve fitting using GraphPad Prism (5.0 version).
Using a query based on L-cysteine-S-sulfinic acid inhibitor [42], the ZINC database was screened using the program ROCS (version 2.3.1) to find compounds that have good shape similarity (measured by 3D Tanimoto) and similar functional group overlap to the query molecule. L-cysteine-S-sulfate (Sigma; PubChem Substance ID 24892471) was used under the following conditions: 2.5 mM aspartate, 1.25 mM ATP, 12.5 mM ammonia, and 8.4 mM Mg2+. The characterization of the mechanism of inhibition consisted in the determination of Km and Vmax for each substrate, in the presence of four inhibitor concentrations (0.025, 0.050, 0.1 and 0.2 mM). The following substrate concentration ranges 1.25–10 and 1.25–50 mM were used for aspartate and ammonia, respectively, while to determine Km for ATP, a range from 0.625 to 10 mM (TbAS-A) or from 0.3125 to 5 mM (TcAS-A) was assayed. Ki was determined by “Km app Method”[43].
EcAS-A, TbAS-A and TcAS-A protein alignments were performed using ClustalW [44]. Aline, Version 011208 [45], was used for editing protein sequence alignments and preparing Fig. 1. TbAS-A and TcAS-A homology models were obtained with SWISS-MODEL, using EcAS-A crystal structure (Protein Data Bank (PDB) accession code 12AS [24]) as a template (percentage of sequence identity of 56% and 57%, respectively) [46]–[48]. The 3D structures were rendered in PyMOL (The PyMOL Molecular Graphics System, Version 1.3, Schrödinger, LLC).
The “stuffer strategy” was used to generate RNAi-mediated AS-A depletion. First, the TbASA fragment (amplified with a sense oligo with a BglII – SphI linker 5′ - GAGAAGATCTGCATGCGCGACGACGGTTATTCGTCATAC - 3′, and an antisense oligo with a EcoRI – SalI 5′ – CGGAATTCGTCGACACTCCGTTTTTCGGATTGCGGC – 3′) was cloned twice in opposite direction on either sides of a ‘stuffer’ fragment of the pHD1144 vector (also digested with SphI and SalI) (Fig. S1A). The resulting [(target)-stuffer-(reverse-complement target)] construct obtained through HindIII and BglII digestion, which generates a stem-loop RNA, was cloned into pHD1145 (also digested with HindIII and BglII) (Fig. S1B). The final construct was linearized with NotI and 10 µg of DNA was transfected into 2×107/ml bloodstream form cell line carrying pHD1313 plasmid (contains two copies of the tet repressor and a phleomycin resistance cassette) by electroporation using Amaxa Basic Parasite Nucleofector Kit (Lonza). Transcription occurs on induction with tetracycline (100 ng/ml), hence producing mRNA homologs to the target the gene. Stable individual clones were selected 5 to 7 days after transfection with 7.5 µg/ml of hygromycin.
To analyse growth, T. brucei RNAi cell line and cells expressing the tet repressor only (wt), were seeded at 2×105 cells/ml of complete HMI9 medium, in the presence and absence of tetracycline. Cell growth was monitored microscopically on a haemocytometer (Marienfeld) and the culture diluted back to 2×105 cells/ml daily. The same protocol was repeated in complete HMI9 medium with basal IMDM without asparagine, complete HMI9 medium with basal IMDM without asparagine supplemented with 6.7×104 nM of asparagine (levels found in human plasma [49]), and complete HMI9 medium with basal IMDM without asparagine supplemented with 1.67×105 nM of asparagine (levels found in normal medium).
Wild-type and transgenic bloodstream T. brucei parasites were cultured in the absence of selecting drugs (hygromycin and phleomycin) for 24 h, then tetracycline was added. After a further 48 h, parasites were inoculated in mice. For each experiment, 4 groups of BALB/c mice (6–8 weeks old, n = 4) (Harlan Laboratories, United Kingdom) were infected by intraperitoneal injection of 104 T. brucei bloodstream forms. 2 groups were injected with wt strain (with or without tetracycline) and the other 2 groups were injected with RNAi cell line (with or without tetracycline). 48 h prior infection the 2 RNAi induced groups were given doxycycline (treated with 1 mg/ml doxycycline hyclate and 5% sucrose containing water). The 2 non-induced groups were given standard water. To evaluate the virulence of RNAi induced parasites in mice with reduced plasmatic levels of asparagine, animals were treated with 50 IU of E. coli L- asparaginase (ProSpec-Tany TechnoGene) 48 h before injection and every 48 h. According to the literature, L-asparagine could not be detected in the blood 48 h following an intravenous injection of E. coli L-asparaginase, at a dose of 50 IU/mouse [50]. Mice were monitored every day for general appearance and behaviour. Parasitemia was monitored daily from the fifth day post-infection, using tail-vein blood, in a haematocytometer under a microscope. Animals with a parasitemia greater than 108 parasites/ml were euthanized, as previous studies had established that these levels were consistently lethal within the next 24 h.
T. brucei bloodstream forms from log-phase cultures, with or without RNAi, were fixed in μ-Chamber 12 well (Ibidi) for 15 min, at room temperature, in PBS containing 3% p-formaldehyde, washed twice with PBS, and then permeabilized in PBS containing 0.1% of Triton X-100. Fixed cells were incubated for 60 min in PBS containing 10% FCS at room temperature (RT), in a humidified atmosphere, then washed twice with PBS/2% FCS. Cells were then incubated with primary rat or rabbit polyclonal antibody against TbAS-A (1∶100 and 1∶5000 respectively, both diluted in blocking solution) overnight at 4°C, followed by two washes with PBS/2% FCS. Subsequently, cells were incubated with Alexa Fluor 647 conjugated goat anti-rat or Alexa Fluor 488 conjugated goat anti-rabbit secondary antibodies (Molecular probes from Life technologies) (1∶500 diluted in blocking solution) for 1 h at RT in an humidified atmosphere, then washed twice with PBS. Next, the slides were stained and mounted with Vectashield-DAPI (Vector Laboratories, Inc.). Images were captured using fluorescence microscope AxioImager Z1 and software Axiovision 4.7 (Carl Zeiss, Germany). Pseudo-coloring of images was carried out using ImageJ software (version 1.43u).
In case of TbAS-A immunolocalization, T. brucei wt bloodstream forms cells were co-stained using rat anti-TbAS-A antibody (1∶100 diluted in blocking solution), rabbit anti-aldolase antibody (1∶5000 diluted in blocking solution), anti-BiP antibody (kindly provided by Dr. Jay Bangs, 1∶500 diluted in blocking solution), anti-enolase antibody (kindly provided by Dr. Paul Michels, 1∶5000 diluted in blocking solution) or anti-GRASP antibody (kindly provided by Dr. Graham Warren, 1∶200 diluted in blocking solution). Alexa Fluor 647 conjugated goat anti-rat (1∶500) and Alexa Fluor 488 conjugated goat anti-rabbit (1∶500) were used as secondary antibodies. Staining with MitoTracker Orange (Invitrogen) followed by Alexa Fluor 488 conjugated goat anti-rabbit (1∶500), as a secondary antibody. The labelling of parasites with MitoTracker was done by adding 250 nM to the cell culture medium (without FCS) for 30 minutes, prior to washing, fixing and staining using the protocol described above. Images were captured using the confocal microscope Leica TCS SP5II and LAS 2.6 software (Leica Microsystems, Germany). Again, image analysis was done using ImageJ version 1.43U software.
For each sample condition, 1.0×107 bloodstream cells were washed once with cold trypanosome homogenization buffer (THB), containing 25 mM Tris (Sigma), 1 mM EDTA (Sigma) and 10% sucrose (Sigma), pH = 7.8. Just before cell lysis, leupeptin (Sigma) (final concentration of 2 µg/ml) and different digitonin (Calbiochem) quantities (final concentrations of 5, 12.5, 25, 50, 100, 150 and 200 ug/ml) were added to 500 µl of cold THB, for cell pellet resuspension. Untreated cells (0 µg/ml of digitonin) and those completely permeabilized (total release, the result of incubation in 0.5% Triton X-100) were used as controls. Each sample was incubated 60 min on ice, and then centrifuged at 2000 rpm, 4°C, for 10 min. Supernatants were transferred to new chilled tubes and 500 µl of cold THB was added to each pellet and then mixed. All fractions were analysed through Western blot as described above.
T. brucei bloodstream forms were analyzed by flow cytometry for DNA content following RNAi induction. Cells were collected by centrifugation and washed twice with PBS containing 2% FCS. Each 2×106 cells were resuspended in 1 mL of PBS/2% FCS and 3 mL of cold absolute ethanol was added while vortexing. Cells were fixed for 1 hour at 4°C and then washed twice in PBS. 1 mL of staining solution [3.8 mM sodium citrate dehydrate (Sigma), 50 µg/mL propidium iodide (Sigma), 0.5 µg/µL RNAse A (Sigma) in PBS] was added to the cell pellets and vortex. Samples were analysed by FACS (Becton Dickinson) after a incubation at 4°C for 30 min. Data was analyzed by FlowJo software (Ashland, OR).
One-way ANOVA and two-tailed Student's test were used for statistical analysis. Statistical analysis was performed using GraphPad Prism Software (version 5.0), and p values≤0.01 were considered to be statistically significant. Asterisks indicate statistically significant differences (*** p≤0.001, ** p≤0.01).
One open reading frame that code for a putative AS-A is present in the genomes of T. cruzi CL Brener Non-Esmeraldo-like and T. brucei TREU927 (http://tritrypdb.org) [29]–[31]. A protein multiple sequence alignment, performed using ClustalW [44], of AS-A from T. brucei (Tb927.7.1110, NCBI-GeneID:3658321), T. cruzi (Tc00.1047053503625.10, NCBI-GeneID:3534325) and E. coli (NCBI-GeneID:948258) is shown in Figure 1. The amino acid residues known to be involved in the active-site formation in E. coli [24] are highly conserved within the three sequences (Fig. 1). Protein alignments demonstrated 58% similarity for EcAS-A versus TbAS-A, 60% for EcAS-A versus TcAS-A, and 63% for TbAS-A versus TcAS-A.
Like EcAS-A, TbAS-A and TcAS are predicted to be dimeric, as seen from superimposed homology models with the EcAS-A crystal structure [24] (Fig. 2A). The only structurally divergent region (area marked by dashed rectangle) (Fig. 1, 2B), is present in both TbAS-A (from residues Q232 to S250) and TcAS-A (from residues D232 to S247), but absent in EcAS-A. This region is distant from the enzyme active site and the dimer interface and its functional and structural significance are unknown. The amino acids involved in asparagine binding are all strictly conserved, while in the AMP binding pocket, the majority of the residues are conserved, except for three residues (Fig. 1). In EcAS-A, E103 (D106 and E106 in TbAS-A and TcAS-A, respectively) and L109 (I112 and T112 in TbAS-A and TcAS-A, respectively) (Fig. 1) are not directly involved in polar interactions with the nucleotide base, but form part of the outer wall of the binding site [24]. The main chain of L249 in EcAS-A (V271 and L268 in TbAS-A and TcAS-A, respectively) is directly involved in hydrogen bonds with ribose from AMP, however the different side chains of leucine and valine do not affect the shape of AMP binding site.
TbAS-A and TcAS-A coding sequences were cloned into the bacterial expression vector pET28a. Histidine-tagged fusion proteins were purified under non-denaturing conditions (Fig. 3A, B). As expected, the recombinant proteins were recognized by anti-His Tag monoclonal antibody (Fig. 3A, B). Rabbit polyclonal antibodies produced against recombinant TbAS-A recognized the protein in total extracts from two different parasite developmental stages, bloodstream forms (mammalian host parasite stage) and procyclic forms (insect vector parasite stage) (Fig. 3C).
The capacities of TbAS-A and TcAS-A to produce asparagine from aspartate in the presence of ATP, ammonia and Mg2+ were determined using a specific quantitative colorimetric assay for L-asparagine [41]. The pH optimum was 7.6, with detectable activity from 6.0 to 9.0 (data not shown). Mg2+ was an essential co-factor for TbAS-A and TcAS-A (data not shown), as previously described for EcAS-A [22]. We included 8.4 mM Mg2+ in the final reaction mixture. Lower concentrations (2, 4 and 6 mM) gave lower activity while increased concentrations (up to 16 mM) resulted in no substantial activity improvement (data not shown). TbAS-A and TcAS-A showed similar Kms for aspartate and ammonia (p>0.01), while TcAS-A showed higher Km for ATP than TbAS-A (p = 0.0042) (Table 1). ATP is the substrate required for the generation of the β-aspartyl adenylate intermediate, which reacts with ammonia, releasing asparagine. In its absence, the reaction did not occur (data not shown). To our surprise, both TbAS-A and TcAS-A could also use glutamine as a nitrogen donor (Table 2). TbAS-A showed higher Km for this nitrogen donor than TcAS-A, however not statistically significant (p>0.01). Both enzymes present higher Km values for ammonia than for glutamine, but these differences were not statistically significant (p>0.01) (Table 1 and 2). TbAS-A had a higher Vmax than TcAS-A, for both ammonia (p<0.0001) and glutamine (p = 0.0043) dependent-activities (Table 1 and 2). TbAS-A had similar catalytic rates for both glutamine and ammonia-dependent activities (p>0.01), whereas TcAS-A presented a slightly higher, but not statistically significant, rate for glutamine-dependent activity (p>0.01) (Table 1 and 2). The high conservation of the active sites and the small amino acid differences identified in the homology models do not allow an accurate structural interpretation of the small differences observed. Indeed, these might have been due to slight differences in the proportion of protein that was correctly folded.
L-cysteine-S-sulfate, considered a putative AS-A inhibitor from a virtual screening, inhibited both enzymes, with IC50s of 126 and 100 µM for TbAS-A and TcAS-A, respectively (Fig. 4A). For both enzymes, the kinetic characteristics suggested competition with ATP binding (Fig. 4C, D). No changes in the Kms and Vmaxs for aspartate and ammonia were observed (p>0.01) (Fig. 4E, F, G, H), suggesting the inhibition is exclusively due to ATP binding interference (p≤0.01) . Ki values of 137 and 128.9 µM were determined for TbAS-A and TcAS-A, respectively (Fig. 4B).
The subcellular localization of TbAS-A was determined by immunofluorescence and digitonin fractionation in bloodstream forms. As expected, induction of RNAi resulted in a decrease in the fluorescence intensity (Fig. S2A, B, C). TbAS-A is in the cytosol, as revealed by colocalization with the cytosolic enzyme enolase [51] (Fig. 5A) and no colocalization with aldolase, BiP, GRASP or mitotracker (Fig. S3), markers for glycosomes [52], endoplasmic reticulum [53], Golgi and mitochondria compartments [54], respectively. Controls performed with rat or rabbit pre-immune sera and secondary antibody alone, showed no detectable signal (data not shown). Digitonin fractionation also resulted in similar profiles for AS-A and enolase (cytosolic marker) and no similarity to aldolase (glycosomes marker) (Fig. 5B).
To study the biological role of AS-A in T. brucei bloodstream forms, cells were stably transfected with an RNA interference plasmid construct. RNAi against asparagine synthetase A was induced in normal medium (complete HMI-9) by adding tetracycline. No difference was observed in cell proliferation between induced and non-induced cells (Fig. 6A), although AS-A protein was reduced to ≈13% of the normal level within 48 hours (Fig. 6B). When, however, the AS-A-depleted cells were grown in HMI-9 medium with only the asparagine from the fetal calf serum, growth was impaired, with an increase in the proportion of cells in G0/G1 (Fig. 6C and 7B, C). Presumably the asparagine from the serum allowed this slower growth. Levels of asparagine usually found in human serum (6.7×104 nM) [49], which are somewhat lower than in normal medium (1.67×105 nM; IMDM - Iscove's modified Dulbecco's basal medium from Gibco Invitrogen), were sufficient to overcome this defect (Fig. 6E, G). In complete HMI-9 medium, with only asparagine from the fetal calf serum, the growth defects of induced RNAi clones are abrogated at day 5 post-induction (Fig. 6C, 7D), and the percentage of cells in GO/G1 and S phases of the cell cycle return to the ones found in non-induced cells (Fig. 7A), suggesting the appearance of RNAi revertants, as is also visible on the Western blot (Fig. 6D). Similar reversion to evade lethal RNAi in trypanosomes has been seen many times before [55]. In the presence of asparagine, low AS-A levels were maintained (Fig. 6B, F and H).
Normal T. brucei parasites also showed a statistically significant slower growth under conditions of asparagine limitation (p≤0.01; data not shown) (compare Fig. 6C, with A, E, G). It is therefore possible that even when the parasite has AS, it also requires external asparagine for optimal in vitro growth.
To test whether AS-A is important for parasite infection in a disease model, two groups of mice (n = 4) were inoculated with the parental cell line, and other two groups with RNAi cells. Two mice groups were fed with water containing doxycycline to induce down-regulation of TbAS-A, while the remaining mice were kept as non-induced controls. Within six days of inoculation, all mice from the different groups developed high levels of parasitemia (Fig. 8A), and all had to be euthanized after seven or eight days post-infection (Fig. 8B). The results confirm that the asparagine in mouse blood is sufficient to compensate for the ≈87% downregulation of AS-A (Fig. 8C, D).
To assess the contribution of blood L-asparagine in vivo, mice were treated with L-asparaginase [50]. L-asparaginase treatment did not affect growth of normal parasites in mice (Fig. 9A) and consequently did not extend animal survival (Fig. 9B). However L-asparaginase treatment in mice infected with TbAS-A RNAi-induced parasites caused a decrease in the parasitemia (Fig. 9D), thus leading to an increase of mice survival (Fig. 9E). Even so, the infection resulted in death. As happened in vitro, RNAi revertants appeared during the course of infection in asparaginase-treated, but not untreated, mice (Fig. 9F). Parasites extracts from wt infected mice were used as controls (Fig. 9C).
In this study we demonstrated that trypanosomes AS-A use both ammonia or glutamine as nitrogen donors for the ATP dependent conversion of aspartate into asparagine. Such hybrid activity was only previously demonstrated for type B enzymes, which prefer glutamine to ammonia [15]–[19]. The small differences in Km of TbAS-A and TcAS-A for ammonia and glutamine (1.5 and 2 fold, respectively) are lower than the difference found in most AS-B enzymes, with the exception of the human enzyme, which has similar affinities for both [16]–[20]. Purified E. coli AS-A used only ammonia as the nitrogen source, and results from Klebsiella aerogenes also suggested that AS-A preferentially uses ammonia as substrate [5], [21], [22]. The conclusions for AS-A enzymes of these two Gram-negative organisms relied on both biochemical and genetic analysis, but given technical limitations at the time, and the fact that background enzyme activity was seen in the absence of both ammonia and glutamine, some re-examination in bacteria would be worthwhile. Moreover the overall Km values of trypanosomes AS-A for aspartate, are 6 up to 20 fold higher than the ones found in the literature for prokaryotic asparagine synthetase type A [5], [21], [22]. Trypanosoma AS-A structures were not yet been solved and our protein homology models are not completely enlightening, nevertheless we can speculate that such differences may result in the fact that parasite enzymes were expressed and purified as recombinant proteins in bacteria and not purified directly from trypanosomes extracts. As a consequence, differences in protein post-transcriptional processing and/or changes in protein conformation cannot be excluded.
Our results suggest that bloodstream-form parasites rely on two major sources of asparagine to ensure normal proliferation: uptake from the extracellular medium and biosynthesis by AS-A. Bloodstream form proliferation, either in vitro or in vivo, was only significantly affected when both asparagine sources were compromised. Also consistent with this idea, in the published RNAi screen, a very slight (possibly insignificant) growth disadvantage was seen in bloodstream forms depleted of AS-A [56]. In the same way, our in vitro results are corroborated with previous studies, as mammalian cells with low expression of AS are similarly susceptible to asparagine depletion [57]–[59], and asparaginase isolated from E. coli and Erwinia carotovora act as potent anti-leukemic agents [60].
In the trypanosomes genome there is a second open reading frame (Tb927.3.4060) coding for a putative AS domain. However this is apparently not a classical AS, despite the presence of a good Pfam AS domain (PF0073) at the C-terminus. A BLASTp search using the T. brucei sequence revealed a variety of proteins of unknown function that aligned not only across the AS domain, but also in the N-terminal region, which contains N-terminal aminohydrolase domains. Best matches originate from extremely diverse eukaryotes including a plant, an alga, a member of the fungi and an amoeba. BLASTp against the Saccharomyces cerevisiae predicted proteome yielded YML096W, and the reciprocal BLASTp on the trypanosome genome indeed gave Tb927.3.4060 as best match. The function of YML096W is not known, and in a trypanosome RNAi screen no growth defect was seen for Tb927.3.4060 [56].
The capacity of trypanosomes to grow using asparagine from the extracellular environment, and the lack of growth defect when the levels of AS-A are reduced, show that only a combination therapy using both a TbAS-A inhibitor and an extracellular asparagine depletor (e.g. L-asparaginase) or an asparagine transport blocker could inhibit parasite growth. This is not appropriate for African sleeping sickness treatment. A combination that absolutely required simultaneous activities of two different drugs would be wide open to resistance development, and drug combination including an intravenously-introduced enzyme is likely to be both too expensive and logistically inappropriate for treatment of African trypanosomiasis. Moreover, L-asparaginase treatment in cancer results in serious adverse events [61]–[63]. We therefore conclude that AS-A is not a good candidate as a sleeping sickness drug target. Its role in Trypanosoma cruzi, however, remains to be established.
|
10.1371/journal.ppat.1006538 | A SIV molecular clone that targets the CNS and induces neuroAIDS in rhesus macaques | Despite effective control of plasma viremia with the use of combination antiretroviral therapies (cART), minor cognitive and motor disorders (MCMD) persist as a significant clinical problem in HIV-infected patients. Non-human primate models are therefore required to study mechanisms of disease progression in the central nervous system (CNS). We isolated a strain of simian immunodeficiency virus (SIV), SIVsm804E, which induces neuroAIDS in a high proportion of rhesus macaques and identified enhanced antagonism of the host innate factor BST-2 as an important factor in the macrophage tropism and initial neuro-invasion of this isolate. In the present study, we further developed this model by deriving a molecular clone SIVsm804E-CL757 (CL757). This clone induced neurological disorders in high frequencies but without rapid disease progression and thus is more reflective of the tempo of neuroAIDS in HIV-infection. NeuroAIDS was also induced in macaques co-inoculated with CL757 and the parental AIDS-inducing, but non-neurovirulent SIVsmE543-3 (E543-3). Molecular analysis of macaques infected with CL757 revealed compartmentalization of virus populations between the CNS and the periphery. CL757 exclusively targeted the CNS whereas E543-3 was restricted to the periphery consistent with a role for viral determinants in the mechanisms of neuroinvasion. CL757 would be a useful model to investigate disease progression in the CNS and as a model to study virus reservoirs in the CNS.
| Despite effective control of plasma viremia with the use of combination antiretroviral therapies, neurologic disease resulting from HIV-infection of the central nervous system (CNS) persists as a significant clinical problem. Non-human primate models are therefore required to study mechanisms of disease progression in the CNS. We generated an infectious molecular clone (CL757) of an SIV isolate from the brain of a macaque with neuroAIDS. This cloned virus induced neurological disorders in 50% of rhesus macaques infected but without rapid disease progression often seen in other commonly used animal models. Molecular analysis of tissues from macaques infected with CL757 revealed that the variants isolated from the CNS and the periphery became genetically distinct from one another. When co-inoculated with an AIDS-inducing, non-neurovirulent clone (E543-3), CL757 targeted the CNS consistent with its neurovirulence. CL757 would be a useful model to investigate disease progression in the CNS and as a model to study virus reservoirs in the CNS.
| Entry of HIV to the CNS occurs early in the course of infection and can progress to HIV-associated dementia (HAD) or HIV encephalitis (HIVE) [1]. HAD is a neurological syndrome that affects 20–30% of infected individuals in the later stages of HIV-infection and includes a range of cognitive and motor disorders, such as impaired short-term memory, reduced concentration and leg weakness. These symptoms are presented along with behavioral changes such as social withdrawal and in extreme cases, near vegetative and mute states. While the introduction of combinational anti-retroviral therapy (cART) has reduced the incidence of HAD to 10% in infected individuals, a milder form of HAD, minor cognitive motor disorder (MCMD), has become more common and affects 30% of the HIV infected population [2–5]. In this syndrome, loss of cognitive and motor functions is less severe, however, MCMD is associated with a worse prognosis for HIV infected individuals [6–8]. Clinical determination of HAD is associated with the presence of multinucleated giant cells (MNGCs) at the time of autopsy, due to productive infection of macrophages and microglial cells, a hallmark of HIVE [9–13].
Neuropathological evaluation of HAD/MCMD progression in humans is not feasible since CNS tissue sampling, other than cerebral spinal fluid (CSF) as a surrogate, is limited to end-stage disease. This creates the need for an animal model that will allow for the study of HIVE pathogenesis. Simian immunodeficiency virus (SIV) infected rhesus macaques are widely used as a model for AIDS pathogenesis. Infection of these animals with neurotropic SIV can result in SIV encephalitis (SIVE/neuroAIDS), with neuropathologic findings reminiscent of HIV encephalitis in humans, including the presence of multinucleated giant cells (MNGCs). Also, infection of macaques with SIV allows researchers access to samples of any part of the brain throughout all stages of disease progression under controlled conditions.
Currently, two nonhuman primate models tend to dominate studies evaluating the mechanism of pathogenesis of SIVE/neuroAIDS. These include the use of immunomodulation by depletion of CD8+ lymphocytes prior to or following inoculation with strains of SIV (SIVmac251/239) that otherwise are inconsistent in inducing SIVE. When CD8+ cells are depleted from rhesus macaques, disease progression is rapid, resulting in AIDS accompanied by a high incidence of SIVE within 3 to 6 months after infection [14–17]. The advantage of this model is that progression to neuroAIDS is rapid and highly reproducible. However, since the host immune system is modified, it may not directly reflect the pathogenesis of HAD/MCMD in HIV infected individuals. The second model is the co-inoculation of the uncloned, immunosuppressive virus SIVsmB670 and a neurovirulent viral clone SIVmac17E-Fr, into pig-tailed macaques [18–21]. As with the CD8+ lymphocyte depletion model, pig-tailed macaques co-inoculated with both strains also reproducibly show rapid disease progression with accompanying CNS disorders. As with the CD8-depletion model, immunomodulation is required to observe neuroAIDS, in this case, B670 induces immunosuppression to allow more efficient replication of 17E-Fr. Although co-inoculation with uncloned virus complicates the analysis of the relative importance of the components of the complex viral populations, these studies suggested that the neurovirulent clone, SIVmac17E-Fr targets the CNS. Another disadvantage of this model is that the prevalence of CNS disease in rhesus macaques is relatively low, requiring the use of pig-tailed macaques that are limited in number for large-scale experiments.
We therefore developed a model that achieves encephalitis in rhesus macaques at a high frequency without immunomodulation by repeated passage of viruses isolated from the brain of rhesus with SIVE [22]. Whereas the initial SIVsmE543-3 clone was only rarely associated with neuroAIDS and replicated inefficiently in monocyte derived macrophages (MDM), the final uncloned SIVsm804E isolate induced SIVE with relatively high frequency and replicated efficiently in MDM. Using construction of chimeric viruses of SIVsmE543-3 with portions of the env, nef and LTRs cloned from the neurovirulent SIVsm804E, we identified four amino acid substitutions in the cytoplasmic tail of gp41 that enhance replication in macrophages and were associated with enhanced BST-2 antagonism [23]. Rhesus macaques inoculated with the parental virus SIVsmE543-3 engineered to contain these four mutations exhibited higher cerebrospinal fluid (CSF) viral load than animals infected with wild type virus [23]. However, despite increased CSF viral load, this clone did not fully recapitulate the pathogenesis of the uncloned isolate [23], consistent with the hypothesis that other viral determinants were required for full neuropathogenesis. It thus was critical to develop clones fully representative of the neurovirulent strain.
In the present study, we generated a full length infectious clone from this neurovirulent isolate, SIVsm804E-CL757 (CL757). CL757 replicates efficiently in vitro in activated peripheral blood mononuclear cells (PBMCs) and monocyte-derived macrophages (MDMs), which has been reported to be a predictor of neurotropism [24–26]. Animals developed clinical symptoms of neurologic disease similar to what is seen in patients with HAD, such as loss of motor control and paralysis. Longitudinal analysis of cerebral spinal fluid (CSF) showed high viral titers in the CNS that correlated with the formation of brain lesions as seen by immunohistochemistry and SIV-specific in situ hybridization. This model also confirms studies that show compartmentalization of virus between the CNS and the periphery [27–32]. Sequences in the plasma and lymph nodes (periphery) showed more diversity of sequence than variants isolated from the brain and CSF (CNS) and showed divergent evolution from CL757 and convergence towards the parental strain SIVsmE543-3 (E543-3) [33], suggesting that CL757 is already highly adapted to the CNS environment. Co-inoculation of macaques with clone CL757 and the parental clone SIVsm-E543-3 demonstrated that CL757 specifically targets the CNS resulting in the development of neuroAIDS.
In a previous study, we identified substitutions in the cytoplasmic tail of gp41 of viruses of a neurovirulent SIV isolate, SIVsm804E that were associated with enhanced antagonism of the host restriction factor, BST-2. These substitutions conferred enhanced replication in macrophages and higher cerebral spinal fluid viral load in infected rhesus macaques [23]. However, since these animals did not develop SIVE, this virus did not appear to recapitulate the full neurovirulence of the uncloned isolate, consistent with the contribution of other genetic determinants. Therefore, we generated a series of chimeric clones from the 804E isolate, mixing and matching between three different 2-LTR constructs and four clones of the entire viral genome spanning the structural, replication and regulatory genes (clones #1, 4, 5 and 7), resulting in 12 full-length SIV clones. As shown in Table 1, clones containing the viral genomes #1 and #4 in combination with any of the LTRs (CL414, CL444, CL616, CL646, CL717 and CL747) did not replicate in either PBMCs or MDMs and were not evaluated further. CL454 and CL484 efficiently replicated in MDMs with low levels of replication in PBMC, while CL656, CL686 and CL787 replicated in PBMCs only. CL757 was the only clone to replicate robustly in both PBMCs and MDMs. Since it has been previously reported that neurotropic viruses are able to replicate in both of these cell types in vitro, CL757 was chosen for further characterization. Phylogenetic analysis of the envelope sequence of CL757 revealed that it clustered with other envelope clones derived from the SIVsm804E isolate that was its source, indicating that it was generally representative of this isolate (Fig 1A).
We previously reported that the uncloned 804E, isolated after sequential in vivo passages of virus isolated from brains of infected macaques with SIVE, replicated more efficiently in MDMs compared to its parental strain E543-3 (E543-3) [22]. Therefore, we assessed the kinetics and extent of in vitro replication of CL757 in PBMCs and MDMs. Replication kinetics were compared to the T cell tropic virus SIVmac239 (Mac239), the original parent E543-3 and the uncloned 804E source of CL757. Under the same conditions, all viruses replicated in PBMCs, however, CL757 replication efficiency was less robust than Mac239, 804E and E543-3 (Fig 1B). In contrast, CL757 showed replication kinetics similar to that of the parental uncloned 804E in MDMs, contrasting with impaired replication of Mac239 and E543-3 viruses in MDMs (Fig 1B). Replication in MDMs is a characteristic feature of neurotropic SIVs and these results indicate that CL757 is a potential candidate as a neurotropic virus. We speculated that replication in PBMCs might also be essential for neurovirulence since the virus must first establish a systemic infection. Many macrophage-tropic SIVs have been reported to be very sensitive to neutralizing antibodies (NAb) [34, 35]. We used sera from rhesus macaques (n = 4) infected for over one year with the neutralization-resistant E543-3 virus to assess the sensitivity to Nab of CL757 as compared to E543-3 and a neutralization-sensitive clone of E660-FL6 [36]. Whereas FL6 was easily neutralized by these sera, NAb titers were significantly less robust for E543 and CL757 was not neutralized, consistent with a neutralization-resistant phonotype (Fig 1C).
To assess the in vivo replication capacity of CL757, eight naïve Indian origin rhesus macaques were intravenously inoculated with 500 TCID50 of CL757 (see Table 2). Robust viral replication was observed in the plasma of all infected animals with peak viral load ranging from 105−107 viral RNA copies/mL of plasma between two and three weeks post-infection (Fig 2A). In six of the eight animals infected (H880, H881, H882, H885, H886 and H887), plasma viral RNA initially declined by two to three logs but over time gradually increased to a set point range of 105−107 viral RNA copies/mL by week 20. The remaining two animals (H883 and H884) exhibited low peak viral loads, one to two logs lower as compared to the viral RNA set points of the other six study animals (2 x 104 and 6 x 104 viral RNA copies/mL of plasma, respectively). A gradual increase in viral RNA was observed after 84 weeks post-infection in H883 such that levels eventually achieved similar end-point levels as the others animals by the time of necropsy at 96 weeks post-infection (2 x 105 copies/ml). Plasma viral load of H884 started to increase prior to H883 and reached to the same end-point level by 64 weeks post-infection (5 x 105 copies/ml) and was maintained at this level until necropsy at 100 weeks post-infection. All study animals were eventually euthanized due to development of opportunistic infections or neuroAIDS (Fig 3A) and disease progression was significantly more rapid than in a historic cohort of macaques infected with the parental E543-3 cloned virus (log rank test, p = 0.0032). However, none of the animals infected with CL757 progressed rapidly to AIDS (i.e. in <6 months), with the first animal to be euthanized surviving until 49 weeks.
All animals showed peak CSF viral loads ranging from 102 to 105 viral RNA copies/mL between two and four-weeks post-infection (Fig 2B). These peaks were controlled down to the range of 103 viral RNA copies/mL in the post-acute phases of infection. Viral loads in the CSF of four out of the eight animals infected (H880, H882, H886 and H887), ultimately reached the range of 107−108 viral RNA copies/mL (Fig 2B). Three of the four animals in this group (H880, H886 and H887) exhibited clinical symptoms of neuroAIDS (Table 1) that resulted in loss of motor control such ataxia, tremors and partial hind-limb paralysis, and other neurologic signs such as anisocoria, nystagmus, head pressing and hiding in the back of the cage. Onset of these symptoms required euthanasia according to animal care and use regulations [37]. All four of these animals exhibited brain lesions characteristic of SIVE/neuroAIDS (i.e., glial nodules, multinucleated giant cells and perivascular cuffing with macrophages and lymphocytes) by routine histopathology [38]. To confirm viral replication at the site of brain lesions, SIV-specific ISH was conducted to detect actively transcribed viral RNA. SIV RNA was detected at the site of brain lesions, indicating that brain lesions were associated with SIV replication (Fig 3B).
Despite robust replication, only four of the animals developed neuroAIDS. Some animals such as H881 and H885 developed opportunistic infections that required euthanasia. Since CSF viral load was on an upswing at the time of euthanasia, we speculated that these macaques may have eventually developed of CNS lesions if they had survived longer. Thus, although the CSF viral load of H881 increased in a similar manner to that observed in the four animals with SIVE, this animal developed a retro-orbital lymphoma at 48 weeks (most likely due to Rhesus Epstein Barr virus infection) while CSF viral load was in the early escalation phase. This necessitated euthanasia at 49 weeks post-infection. As for H885, the kinetics of increase in the CSF viral load was delayed to 32 weeks post-infection. CSF viral load of this animal reached to 104 viral RNA copies/mL at 52 weeks post-infection and was euthanized at 53 weeks post-infection due to hind limb paresis but prior to the development of SIV encephalitis. The remaining two animals with lower set point plasma viral RNA loads, H884 and H883, also showed lower CSF viral load compared to the rest. However, these macaques showed increasing viral RNA levels in plasma and CSF at much later time-points than seen in animals with neuroAIDS (92 weeks post-infection for H883 and 44 weeks post-infection for H884). Although there was no obvious pathology in the CNS, CSF viral load of H883 and H884 reached 8 x 103 and 5 x 104 viral RNA copies/mL, respectively (Fig 2B) which is right at the threshold we previously observed to be associated with SIVE [39].
Lymphocyte subset analysis revealed that CD4+ T cells decreased rapidly in the acute phase of infection (2 to 8 weeks post-infection) but showed transient recovery by 20 weeks post-infection, though overall the CD4+ T cells levels showed gradual decrease over time (Fig 2C). Loss of CD4+ T cells was more evident when we focused on CD4+ memory T cells. All animals were euthanized due to the onset of neurological and/or AIDS like symptoms (diarrhea, weight loss), which coincided with CD4+ T cell numbers of less than 200 cells/μL (Fig 2D).
We and other groups have previously reported compartmentalization between plasma, CSF and brain parenchyma of animals infected with neurovirulent SIV or in patients that were infected with HIV-1 [27–32, 38]. However, these observations came from examination of viral populations in macaques inoculated with uncloned virus containing multiple variants or where the infecting strain was not identified such as in HIV-1-infected patients. Hence it was unclear whether compartmentalization was due to the acquisition of advantageous mutations by selective pressure, or whether the infecting virus contained variants that had advantages in replicating in certain compartments. To clarify which mechanism that drives compartmentalization in the CNS, full viral envelope fragments from the plasma, axillary lymph node, CSF and the brain parenchyma were amplified from samples collected at the time of necropsy from three animals that progressed to neuroAIDS (H880, H882 and H886). Phylogenetic analysis revealed clear compartmentalization between the periphery (plasma and axillary lymph node) and the CNS (CSF and brain parenchyma) in all of the animals (Fig 4A–4C). This finding suggests that interchange of viral populations between the CNS and the periphery are relatively rare in the end stage of the disease and viruses are likely to be produced locally within each compartment. To determine if the process of compartmentalization was similar in all three animals, we constructed a phylogenetic tree that included all variants from the three animals (Fig 5). Surprisingly, viruses from the CNS of all three animals formed a unique cluster distinct from the viral sequences obtained from the periphery, regardless of the animal of origin. Therefore, viruses from the CNS of these animals were more similar to one another than to variants in autologous plasma samples. Variants from the brain and CSF had relatively shorter branch lengths, indicating they are genetically closer to each other and the inoculum CL757 compared to those variants isolated from plasma and the lymph node. We analyzed consensus amino acid substitutions in each group to determine if the difference in branch length was due to higher level of viral replication in the plasma, or if there are selective pressures driving this divergence (Table 3). Although the substitutions were generally unique in each of the animals, we found some common mutations present within both the CNS and the periphery. As shown in Table 3, fewer common amino acid substitutions (9 positions) were observed in the CNS as compared to the periphery, consistent with the branch length in Fig 5. One amino acid substitution (S217G) was observed in H880 and H886. On the other hand, 31 amino acid substitutions were observed in variants isolated from the periphery. Thirteen of those amino acid substitutions were common in two or more animals, suggestive of common selective pressure for acquisition of those mutations in the periphery. Furthermore, there were five amino acids (D79N, A130T, N429K, A829T and R847K) that reverted to the residues found in E543-3, suggesting that these positions may be important for neurovirulence. Indeed, we have previously reported that amino acid substitution at T847A in 804E is associated with enhanced antagonism against host restriction factor BST-2 [23]. Reversion of this position to original E543-3 in periphery further supports idea that BST-2 restriction is important in the CNS microenvironment and SIVsm804E gained this mutation to counter act this restriction. In addition, reversion in this position observed in variants from plasma/LN compartments suggests that this mutation was negatively selected in these compartments.
Phylogenetic analysis suggested that CL757 has advantage over its parent, E543-3 in the CNS microenvironment. To evaluate whether our neurovirulent CL757 specifically targeted the brain, we intravenously co-inoculated two animals with CL757 and E543-3 in a 1:1 ratio. Robust viral replication was observed in both infected animals with peak viremia of 1.4 x 107 and 3 x 105 viral RNA copies/mL at two weeks post-infection, respectively (Fig 6). This outcome is comparable with animals inoculated CL757 alone (Fig 2A). The set-point viral load of H842 and H843 were on par with animals that progressed to neuroAIDS (H880, H882, H886 and H887) ranging between 104−107 viral RNA copies/mL. The peak CSF viral load of H842, observed at 2 weeks post-infection (5.2 x 105 viral RNA copies/mL) was highest among all the animals in this study. On the other hand, H843 showed delayed peak CSF viral RNA load at 6 weeks post-infection (1.4 x 104 viral RNA copies/mL) (Fig 4B). CSF viral load for H842 and H843 declined to almost undetectable levels (200 viral RNA copies/mL at 12 weeks post-infection and undetectable level at 10 weeks post-infection, respectively). However, plasma viral load showed continuous increase throughout the course of disease. A significant increase in CSF viral load was observed at 76 weeks post-infection, exceeding plasma viral load in both animals (Fig 6A). Consistent with animals infected with CL757 alone, memory CD4+ T cells declined to approximately 200 cells/μL, for both animals by the time of euthanasia (Fig 6C). Viral RNA was detected at site of brain lesions in both H842 and H843, which is consistent with SIVE/neuroAIDS.
Previous studies using the pigtailed macaque model of co-inoculation of immunosuppressive SIVsmB670 with the neurovirulent 17EFr clone suggested that the clone specifically targeted the CNS but this finding was based on evaluation of a small region of env [40]. We conducted phylogenetic analysis of envelope sequences isolated from terminal plasma, CSF and the brain of H842 and H843 co-inoculated with CL757 and E543-3 (Fig 7). Consistent with results from animals inoculated with CL757 alone, compartmentalization of viral populations between the CNS and plasma was also observed in these two animals (Fig 7). More striking however was the compartmentalization of CL757-related variants, to the CNS, whereas variants from the plasma formed a group with E543-3, distinct from each other. This result clearly indicates that CL757 preferentially targets and replicates in the CNS whereas E543-3 has the advantage in replication in the periphery.
Shortly after infection, HIV becomes established in the CNS. In 20–30% of HIV-infected individuals it induces HIVE and the associated neurologic disorder HAD following a period of time that varies between individuals. Although the introduction of ART has reduced the incidence of the HAD, a milder form of neurologic disease, MCMD has become more common and is associated with a worse outcome for HIV patients. The virologic and biologic factors that determine whether an individual will develop neuroAIDS in any of its forms is not clear, nor is the extent to which the CNS represents a viral reservoir in ART-treated patients. Since neuropathological evaluation of HAD/MCMD progression is not feasible to study in humans due to the limited access to tissue samples, animal models that accurately recapitulate critical aspects of neuroAIDS in HIV-infected humans are necessary. In this study, we have developed a neuroAIDS model in rhesus macaques inoculated with either a neurovirulent virus, CL757 alone or a combination with its non-neurovirulent parent, E543-3. Unlike existing macaque models that produce a rapid onset of neuroAIDS, the disease progression in this current model closely resembles late stage development of neuroAIDS in HIV-infected patients.
These studies were initiated with the macaque passage of the molecularly cloned, AIDS-inducing SIVsmE543-3. Despite similarities in the nature of disease progression to that of HIV-1 infection in human, E543-3-infected rhesus macaques only rarely develop SIV encephalitis [22]. In an approach to establish a nonhuman primate model for neuroAIDS in rhesus macaques, we conducted sequential in vivo passage of E543-3, selecting viruses at each passage that had been isolated from the brain of animals that developed SIVE as the source of inoculum. Four such passages lead to the isolation of the uncloned virus SIVsm804E that induces neuroAIDS in a high proportion of rhesus macaques, if one excludes animals with restrictive TRIM5 α or MHC-I genotypes [22]. While this met many of the criteria we had established, disease progression in this model was variable with some animals developing a rapid disease course that is atypical of HIV neuroAIDS. Additionally the virus inoculum was quite complex leading to difficulty in molecularly tracking the virus(es) during disease progression. This made it a difficult model to dissect viral determinants of neurovirulence.
To further develop and improve this model for pathogenesis studies, we derived twelve full-length infectious viral clones from this uncloned 804E stock. Out of the 12 clones tested only one virus, CL757, was capable of replicating in both PBMCs and MDMs. Although CL757 showed a reduced capacity to replicate in PBMCs compared with its uncloned parental strain 804E, both showed similar replication kinetics and comparable levels of virus production in MDMs. Since replication of virus in macrophages in the CNS is a hallmark of neuroAIDS, CL757 was evaluated in rhesus macaques in vivo. Robust viral replication was observed in the eight inoculated animals as evidenced by high acute plasma viral RNA levels, similar to levels as we previously reported with uncloned 804E [22] and virus was detectable in the CSF of each of the animals during primary infection and reached high levels in 50% of the animals following a variable eclipse phase. Importantly, this increase was associated with development of SIVE/neuroAIDS. Most of the animals infected showed high levels of viral RNA in the CSF at the time of necropsy, although this was particularly evident in the animals that progressed to SIVE (H880, H882, H886 and H887). Three of the four (H880, H886 and H887) presented with neurological symptoms that necessitated euthanasia. These clinical symptoms of neurological disease included behavioral issues and loss of motor control, similar to symptoms in HAD/MCMD [41, 42]. Each of these animals also met the clinical definition of AIDS as assessed by depletion of CD4+ T cells below approximately 200 cells/μl, indicating that immune suppression may be essential for the development of neuroAIDS. Post-mortem examination of brain tissue of all four animals with neuroAIDS revealed widespread lesions in the white matter characterized by perivascular cuffing and glial nodules containing multinucleated giant cells expressing SIV RNA as measured by ISH [43]. We did not observe rapid disease progression in any of the animals infected with CL757 or the two animals co-inoculated with E543-3 and CL757. The more conventional disease progression contrasted with the more rapid disease observed in animals inoculated with uncloned SIVsm804E [22]. The lack of rapid disease is fortunate in that it will provide an excellent model to assess mechanisms of disease progression in the CNS in the future, and is likely more reflective of the late stage disease progression in HIV-1 infected individuals.
Due to the use of molecularly cloned virus, this study also allowed us to examine the molecular characteristics and evolution of virus in the peripheral immune system and the CNS. Previous studies in both HIV-infected humans and SIV-infected macaques show that the microenvironment of the CNS fosters viral compartmentalization [27–32, 38]. However, it is not clear if viruses become compartmentalized in the CNS due to a founder effect where specific viruses target the brain or whether compartmentalization occurs through the selection and evolution in this compartment. We previously reported viral compartmentalization in animals inoculated with uncloned virus stocks (SIVsm783Br and SIVsmH445) where the inoculum is complex and individual components are difficult to track [38, 44]. In the present study, the analysis of viral compartmentalization in macaques inoculated with CL757 allowed us to better comprehend the evolution and selective events involved in development of neuroAIDS. Additionally, using co-inoculation of two molecular clones, one neurovirulent (CL757) and the other AIDS-inducing but rarely neurovirulent (E543-3) we were able to confirm that CL757 is clearly highly adapted for replication in the CNS in contrast to its parent E543-3 strain which was generally excluded from the CNS. Phylogenetic analysis of CL757-infected animals revealed that variants isolated from CSF and the brain formed a clade distinct from variants isolated from plasma and lymph node. This result indicates distinct selective pressures between the CNS and the periphery. The CNS group had a lesser degree of divergence than that of plasma/lymph node group, suggesting that there are selective pressures against replication of CL757 in plasma/lymph node. Alternatively, greater and more sustained levels of viral replication in the plasma/lymph node compartment allowed virus to accumulate mutations. Upon closer examination, we saw mutations in the envelope of variants isolated from the periphery that appear in the parental clone E543-3 (D79N, A130T, N429K, A829T and R847K). These mutations were not observed in the variants isolated from the CNS suggesting that CL757 is highly adapted to replicate in the CNS, whereas E543-3 has advantage in the periphery. Co-inoculation of both E543-3 and CL757 confirmed this hypothesis. The variants from CNS formed one group with CL757 sequence, whereas variants from plasma and lymph node formed another group with E543-3. Thus, those amino acids that reverted back to original E543-3 may play important role in neurovirulence, which may have disadvantages for replication in plasma and lymph node. Indeed, we have previously reported that amino acid substitution at position T829A in E543-3 enhances antagonism against the host restriction factor BST-2 [23]. The fact that this amino acid substitution reverted back to the original E543-3 residue (A829T) in the plasma and lymph node indicates that enhanced BST-2 antagonism is positively selected in the CNS but negatively selected in the plasma and lymph node. The evolution of CL757 was associated with a change in biological properties. The most obvious change was the acquisition of the ability to replicate efficiently in MDMs, which has implications for neurovirulence. Our prior study demonstrated that differences in the cytoplasmic tail of gp41 (gp41 CT) of the neurovirulent variant were associated with enhanced replication in MDMs and better antagonism of BST-2 [23]. While introduction of these substitutions to E543-3 enhanced replication in MDMs and virus levels in the CSF, they did not fully recapitulate the neurovirulence of CL757, indicating that other differences are important for neurovirulence.
In summary, we have successfully isolated a molecular clone of SIV that is highly adapted to target and replicate efficiently in the CNS microenvironment and induce neurological disorders in infected rhesus macaques in high frequencies. The disease course of SIVsm804E-CL757 infected animals is conventional as opposed to those existing rapid disease models that require immunomodulation or that require the use of pigtail macaques. The CL757 model will enable us to assess mechanisms of disease progression in the CNS during the chronic phase of infection. CL757 shares the same genetic backbone with its parental E543-3, therefore this virus would be a powerful tool to assess viral determinants of neurovirulence.
This study was carried out in strict accordance with the recommendations described in the Guide for the Care and Use of Laboratory Animals of the National Institute of Health, the Office of Animal Welfare and the United States Department of Agriculture. Colony-bred rhesus macaques of Indian origin were obtained from Morgan Island rhesus monkey breeding colony, SC. All animal work was approved by the NIAID Division of Intramural Research Animal Care and Use Committees (IACUC), in Bethesda, MD (Animal Study protocol, #ASP-LMM-6). The animal facility is accredited by the American Association for Accreditation of Laboratory Animal Care. All procedures were carried out under Ketamine anesthesia by trained personnel under the supervision of veterinary staff and all efforts were made to ameliorate the welfare and to minimize animal suffering in accordance with the “Weatherall report for the use of non-human primates” recommendations. Animals were housed in adjoining individual primate cages allowing social interactions, under controlled conditions of humidity, temperature and light (12-hour light/12-hour dark cycles). Food and water were available ad libitum. Animals were monitored twice daily (pre- and post-challenge) and fed commercial monkey chow, treats and fruit twice daily by trained personnel. Early endpoint criteria, as specified by the IACUC approved score parameters, were used to determine when animals should be humanely euthanized.
|
10.1371/journal.pntd.0006151 | In vitro studies of Rickettsia-host cell interactions: Confocal laser scanning microscopy of Rickettsia helvetica-infected eukaryotic cell lines | Rickettsia (R.) helvetica is the most prevalent rickettsia found in Ixodes ricinus ticks in Germany. Several studies reported antibodies against R. helvetica up to 12.5% in humans investigated, however, fulminant clinical cases are rare indicating a rather low pathogenicity compared to other rickettsiae. We investigated growth characteristics of R. helvetica isolate AS819 in two different eukaryotic cell lines with focus on ultra-structural changes of host cells during infection determined by confocal laser scanning microscopy. Further investigations included partially sequencing of rickA, sca4 and sca2 genes, which have been reported to encode proteins involved in cell-to-cell spread and virulence in some rickettsiae. R. helvetica grew constantly but slowly in both cell lines used. Confocal laser scanning microscopy revealed that the dissemination of R. helvetica AS819 in both cell lines was rather mediated by cell break-down and bacterial release than cell-to-cell spread. The cytoskeleton of both investigated eukaryotic cell lines was not altered. R. helvetica possesses rickA, but its expression is not sufficient to promote actin-based motility as demonstrated by confocal laser scanning microscopy. Hypothetical Sca2 and Sca4 proteins were deduced from nucleotide gene sequences but the predicted amino acid sequences were disrupted or truncated compared to other rickettsiae most likely resulting in non-functional proteins. Taken together, these results might give a first hint to the underlying causes of the reduced virulence and pathogenicity of R. helvetica.
| The pathogenicity of Rickettsia helvetica has not been investigated in depth to date. In humans, seroprevalences up to 12.5% against R. helvetica have been demonstrated with forest workers being predisposed to infection. However, fulminant clinical cases are rare indicating a rather low pathogenicity compared to other Rickettsia species. We therefore investigated growth characteristics of a R. helvetica tick isolate (AS819) in two different eukaryotic cell lines with focus on ultra-structural changes of host cells during infection as determined by confocal laser scanning microscopy. Further investigations included sequencing of rickA, sca4 and sca2 genes, which have been reported to encode proteins involved in cell-to-cell spread in some rickettsiae. In contrast to what is known from other rickettsiae, R. helvetica did not spread directly from cell to cell by actin-based motility presumably due to a deletion in the predicted Sca2 protein. As Sca2 is needed for virulence our results might indicate less virulence and pathogenicity of R. helvetica isolated from ixodid ticks in Germany.
| Rickettsia (R.) helvetica is the most prevalent rickettsia found in Ixodes (I.) ricinus ticks in Germany with varying prevalence up to 17% [1–4]. The organism has mainly been considered non-pathogenic and affected patients usually show a mild disease, manifesting in non-specific fever without erythema (so-called uneruptive fever), headache, and myalgia [5]. Hence, only a few laboratories in Germany focus on R. helvetica infection as a differential diagnosis and infections might be underdiagnosed due to mild symptoms in most cases. The latter might be the reason that, to the authors’ best knowledge, no clinical case in humans has been reported from Germany so far. However, more severe clinical cases have been demonstrated in Sweden including septicemia [6], myocarditis [7] and meningitis [8]. Only recently, complex phylogenetic studies showed that R. helvetica phylogenetically was misplaced in the spotted fever group (SFG) [9]. In contrast, R. conorii subsp. conorii, a typical representative of the SFG, causes Mediterranean spotted-fever, a disease that is characterized by fever, an eschar at the site of the tick bite, and a rash spreading to the palms and soles. The disease is endemic in southern Europe but single cases have also been reported from the central and northern European mainland [10].
As many obligate intracellular bacteria, rickettsiae proliferate in the cytoplasm as dispersed, individual bacteria but may also occasionally be found in clusters and in the nucleus [11]. The primary targets for rickettsiae during infection are endothelial cells of the middle and small vessels [12–13]. Disseminated infection of the endothelium and subsequent pathophysiological effects lead to most of the clinical characteristics described for rickettsial diseases [13]. Rapid spread within host tissues is a crucial step in many infectious diseases [14]. For some Rickettsia species it has been proven that rickettsiae harness the host cell actin cytoskeleton for intracellular movement and cell-to-cell spread. This phenomenon was first reported by Teysseire et al. [15] and Heinzen et al. [16] who observed associations between R. conorii and R. rickettsii and host F-actin. Most of the SFG rickettsiae (SFGR) as well as R. typhi assemble actin tails and undergo actin-based motility mediating cell-to-cell spread and enhancing virulence [17–18]. Two actin-polymerizing proteins have been identified in SFG rickettsiae: RickA, which activates the actin-related protein-2/3 (Arp2/3) complex of the host [19–20], and surface cell antigen 2 (Sca2) which has been suggested to mimic eukaryotic formin proteins [18,21]. Cardwell and Martinez [22] identified the minimal domain within the Sca2-protein of R. conorii that is sufficient for stimulating actin polymerization. Most recently, Sca4 was identified as a secreted effector of spread independent from actin-based motility in the SFG R. parkeri [23].
So far—to the authors’ knowledge—only a single study on growth characteristic of R. helvetica in eukaryotic cells exists [24] besides the original description of R. helvetica in 1993 [25]. Based on host cell decomposition and intranuclear growth of R. helvetica, the study by Elfving et al. [24] underlined the pathogenic ability of R. helvetica. Here, we report growth characteristics of R. helvetica isolate AS819 in two different eukaryotic cell lines with focus on ultra-structural changes of host cells during infection as determined by confocal laser scanning microscopy (LSM). Further investigations included sequencing of rickA, sca4 and sca2 genes, which have been reported to encode proteins involved in cell-to-cell spread in some rickettsiae.
Eukaryotic cell lines used in this study were obtained from LGC Standards, Wesel, Germany. L929 cells (murine fibroblasts from connective tissue; up to 25 passages from the original LGC Standards culture) and Vero E6 cells (African green monkey kidney cells; up to 39 passages from the original LGC Standards culture) were grown in Minimum Essential Medium (MEM) supplemented with Gibco GlutaMAX and 1x MEM non-essential amino acids (NEAA) solution (Life Technologies GmbH, Darmstadt, Germany) and 3% fetal calf serum (FCS) at 37°C, 5% CO2. The R. helvetica AS819 isolate (seventh passage from the original culture) used for the growth studies was isolated from I. ricinus. Species identity was confirmed based on 100% ompB-sequence identity (4,848 nt, accession number MF163037) to R. helvetica type strain C9P9 (accession number AF123725). R. conorii (Moroccan isolate VR141) and R. honei (VR1472) were obtained from ATCC, Manassas, USA. All rickettsiae were cultured on Vero E6 monolayers at 32°C in MEM prepared as described above. Flasks were checked daily for detrimental alteration of the monolayer. Rickettsiae were grown until cell layers revealed plaque formation (R. conorii, R. honei) or detached from the flasks surface (R. helvetica). The amount of R. helvetica in culture supernatants was determined by quantitative gltA real-time PCR as described below. A non-infected cell control (MOCK) was carried along with every infection for the evaluation of changes in the cell layers.
L929 and Vero E6 cells (105 cells/well) seeded in 4-well BD Falcon CultureSlides (BD Biosciences, Heidelberg, Germany) were grown to confluent monolayers overnight at 37°C, 5% CO2. Prior to infection, monolayers were rinsed with FCS-free medium. Infection was performed with 100 μl of a seed culture containing 5 x 105 R. helvetica AS819-genome equivalents (corresponding to 5 genome equivalents per cell) determined by quantitative gltA real-time PCR as described below. After one hour of incubation at room temperature whilst constantly rocking, wells were filled up with 900 μl MEM supplemented with Gibco GlutaMAX NEAA and 3% FCS. Of every 4-well slide, three wells contained infected cells whilst the remaining served as a non-infected cell control (MOCK). Culture slides were incubated at 32°C, 5% CO2. Experiments were performed over a period of 20 days. A slide of each cell line was randomly selected on a daily basis and prepared for the quantification of rickettsia in the cell culture supernatant and the cell layer, respectively. Briefly, cell culture supernatants of each well were harvested individually and stored at -80°C. One ml of cell culture medium per well was added to the remaining cell layers and the slides were frozen at -80°C for one hour. Subsequently, freeze-thawed cells were scraped off, harvested and stored at -80°C until further processing. Nucleic acids were isolated from 200 μl of all samples applying the MagNA Pure LC Total Nucleic Acid Isolation Kit (Roche, Mannheim, Germany) and the MagNA Pure LC 2.0 system (Roche) according to the manufacturer’s instructions. Nucleic acids were eluted in a total volume of 50 μl. Quantification of rickettsiae was carried out by a real-time PCR targeting the single copy citrate synthase gene gltA in a Stratagene MX3000P Thermocycler (Agilent Technologies, München, Germany) as described earlier [26–27]. Five μl of nucleic acids were used as a template for the PCR. A commercially available DNA-standard (AmpTech GmbH, Hamburg, Germany) with a defined concentration (2.77 x 1010 copies/μl) was used for the quantification of Rickettsia genome equivalents. The number of R. helvetica copies/μl of template was calculated using the Stratagene Software by comparing the samples to a serial dilution of the DNA-standard (2.77 x 105 to 2.77 x 101 copies/μl).
Cells were seeded in μ-Slide 8 well ibiTreat chambers (ibidi GmbH, Martinsried, Germany). Briefly, 2.5 x 105 cells/well were grown to near confluence overnight and were infected with 5 genome equivalents per cell as described above. After an initial incubation period of 1 h at room temperature, slides were transferred to 32°C, 5% CO2. Fixation and staining procedures were performed at room temperature. At 24 hours-intervals, culture supernatants were discarded and cells were washed using phosphate-buffered saline (PBS, pH 7.2) pre-heated to 37°C. Cells were fixed for 15 min in methanol-free formalin (3.7%). Following a wash with PBS, cells were permeabilized with 0.1% Triton X-100 in PBS for 4 min. To reduce nonspecific binding and background signal, blocking was performed in PBS with 1% bovine serum albumin for 60 min. Thereafter, cells were incubated with an anti-SFG Rickettsia antibody (1:2, Fuller Laboratories, Fullerton, USA) for 30 min. After washing with PBS, a mouse monoclonal anti-alpha-tubulin antibody (1:200, Molecular Probes, Life Technologies) was added for 30 min and F-actin was stained with Alexa Fluor 568 Phalloidin (1:100, Molecular Probes, Life Technologies) in parallel. Samples were washed four times and the secondary antibody for Rickettsia staining, an Alexa Fluor 647-labeled goat anti-human IgG (1:2000, Molecular Probes, Life Technologies), was added for 30 minutes. In addition, Alexa Fluor 488-labeled goat anti-mouse IgG (1:2000, Molecular Probes, Life Technologies) was added for staining of microtubuli. Washing another four times was followed by nuclei counterstaining with DAPI (4'.6-Diamidino-2-Phenylindole, Dilactate, 1:5000, Molecular Probes, Life Technologies). Confocal images were obtained with a Zeiss LSM 710 using an EC Plan-Neofluar 40x/1.30 Oil DIC M27 objective and using the ZEN software (Zeiss, Jena, Germany). For R. helvetica-infected cells, staining of slides was performed immediately after fixation each day during the incubation period of 20 d. R. conorii served as a control but the incubation period was limited to five days after infection due to the rapid cell spread of this rickettsia within cells. Images of L929 or Vero cells infected with R. helvetica were analyzed using the software Daime [28]. The detection of nuclei and rickettsiae by the software was optimized by comparing manual counts and software-based counts for three pictures each and the following parameters were chosen: Nuclei stained with DAPI were detected using threshold detection for objects larger than 500 pixels. Rickettsiae were detected using edge detection for objects larger than 20 pixels. As it was not possible to visually separate single rickettsiae in highly infected host cells at later time points, we decided to use the signal area instead of the number of objects in order to quantify the infection progress. The signal area corresponding to rickettsiae in at least 450 host cells was analyzed in images from five, ten, fifteen and twenty days post infectionem (p.i.) and the proportion of the area of rickettsiae to the area of cell nuclei was calculated for each time point.
Confluent L929 and Vero E6 cells grown in 6-well plates were washed two times with FCS-free Medium. One ml of R. helvetica culture supernatant (undiluted and serially diluted 10−1–10−4) was added per well. One well contained an uninfected control. Plates were incubated at room temperature for 1 h while constantly rocking. Double concentrated M199 (Gibco Thermo Fisher Scientific, Waltham, USA) containing 10% FCS was mixed with an equal volume of pre-heated 2% Agarose (Agarose LE, Biozym, Hessisch Oldendorf, Germany) and each well was then filled with 4 ml of agarose overlay. Plates were incubated for 21 d at 32°C, 5% CO2. Formalin (3.7%) fixation (room temperature, 1 h) was done on days 7, 14, and 21 p.i.. After removal of the agarose plugs the cells were stained using crystal violet (1%) in 20% ethanol (30 min, room temperature). After staining, the plates were washed and examined. In parallel the BSL-2 Rickettsia R. honei was used as a positive (i.e. plaque-forming) control.
Amplification of the partial rickA gene was conducted by PCR using primers described by Balraj et al. [29]. PCR conditions were established using DNA from R. conorii VR141. Direct sequencing of PCR products was carried out by GATC Biotech AG sequencing service (Konstanz, Germany). Determination of open reading frames and subsequent amino acid (aa) alignments with corresponding sequences retrieved from the GenBank database (http://www.ncbi.nlm.nih.gov/) were performed using the software package BioEdit v. 7.2.5 [30]. The BioEdit Sequence Alignment Editor, Version 7.2.5 and the implemented ClustalW, Version 1.4 [31] were applied for sequence analyses. In addition, sequencing results of rickA were complemented by data obtained from whole genome pyrosequencing of R. helvetica AS819 using the Roche 454 GS-FLX platform (data not yet published). 377,211 shotgun reads (135,929,920 bases) were assembled using GS De Novo Assembler version 2.3, GS Reference Mapper version 2.3, and DNASTAR SeqMan Ngen version 10.1.0. On average, the coverage of the R. helvetica AS819 plasmid was 146-fold and the coverage of the genome was 34-fold. We predicted CDSs using the RAST prokaryotic genome annotation server (http://rast.nmpdr.org/rast). RNAs were identified using tRNAscan-SE v. 1.23 (tRNA), aragorn v. 1.2.34 (tRNA, tmRNA), and RNammer v.2.1 (rRNA). Database searches were done using BLAST and infernal. Hereby sca2 gene sequences were obtained.
The growth of R. helvetica AS819 in two eukaryotic cell lines is summarised in Fig 1A and 1B. Results were obtained from triplicates of infected cultures per cell line and time point. After two to four days of lag, a continuous increase of R. helvetica genome equivalents (given as copies/μl) was seen in both cell lines resulting in a maximum of 6 x 105 copies/μl at day 19. Propagation in Vero E6 cells (Fig 1A) revealed a slight increase of intracellular R. helvetica DNA copies from day four to eleven after infection. A 100-fold increase was observed from day eleven to day twelve after infection. In L929 cells (Fig 1B), DNA copies of intracellular R. helvetica tripled from day five to day six and a sharp increase of DNA was seen from day seven to day ten resulting in an approximately 1,000-fold rise. A maximum of 100-fold difference between intracellular and extracellular DNA copies/μl was measured in Vero E6 cells (Fig 1A). In contrast, less than 10-fold difference was found using L929 cell lines (Fig 1B).
Rapid cell-to-cell spread was not seen in R. helvetica (isolate AS819)-infected cell monolayers. LSM analyses revealed that the percentage of infected cells was rather constant over a period of several days indicating little cell-to-cell spread. Five to ten days p.i., infected cells were scattered throughout the monolayers (Figs 2 and 3). At that time, up to ten rickettsiae were countable within the cytoplasm of a single cell. From day fifteen until the end of the experiment, the number of infected host cells increased. However, infected cells remained in clusters indicating that R. helvetica initiated mainly new infection of adjacent cells. R. helvetica build up in large numbers in the cytoplasm of single cells were visible. No intranuclear bacteria were seen.
A quantification of rickettsiae per nucleus area over time revealed higher numbers of R. helvetica AS819 within L929 cells compared to Vero E6 cells during the whole infection experiment (Fig 4). In both cell lines the number of rickettsiae per cell nucleus doubled from time point to time point until 15 days after infection. The experiment could be pursued until day 20 p.i. using Vero E6 cells and Fig 4 shows an additional slight increase from day fifteen to day 20 after infection. The detachment of L929 cells increased after day fifteen p.i., hence, due to the low number of attached cells LSM analyses were impossible after that point in time.
In neither cell line actin polymerization due to R. helvetica AS819 was detected in contrast to the R. conorii–infected cell lines (Fig 5). Further, in the latter a prominent plaque formation was visible on day four p.i. indicating rapid cell-to-cell spread. The analysis of images of the MOCK-infected controls did not yield any signals corresponding to rickettsiae.
No plaque formation was seen in R. helvetica-infected cell lines at 7 d p.i. (Fig 6A and 6B), 14 p.i. and 21 d p.i. (Fig 6C and 6D). As noticed before, the detachment of L929 cells increased with prolonged incubation time (Fig 6D). In contrast, plaque formation was readily observed in both cell lines infected with R. honei at 7 d p.i. (Fig 7A and 7B).
The PCR targeting rickA and subsequent pyrosequencing resulted in a nucleotide sequence of 1,677 nt (accession number MF163038) with highest similarity (i.e. 90%; 1,528 nt/1,694 nt, 40 gaps) to the complete coding sequence of the rickA gene of R. raoultii strain DnS14 (accession number EU340900). The open reading frame encoded a deduced 559 aa sequence. Using the RAST annotation server, this sequence was identified as hypothetical Wiscott-Aldrich Syndrome Protein (WASP)-like protein.
A protein BLAST search revealed 100% identity (559/559 aa) to a hypothetical protein of the R. helvetica type strain C9P9 (accession number WP010421970) followed by 83% similarity (465/560 aa) to the Arp2/3 complex-activating protein RickA of R. felis (accession number WP039594871). Hypothetical RickA in R. helvetica AS819 revealed an N-terminal domain for binding monomeric actin (G-actin binding site) and several proline (P)-rich repeats were counted within the protein sequence (S1 Fig). In addition, the commonly called WCA region composed of a WASP-homology 2 (WH2) region, a central (C region), and an acidic domain was identified in R. helvetica AS819. As described for partial RickA of different other Rickettsia species, human WASP and N-WASP, a conserved motif (ФXXФXXФXXXRXXФ) was found in the C region (S1 Fig), with Ф representing an aliphatic amino acid, X any residue, and R an arginine [32]. In addition, R. helvetica AS819 possesses two WH2 regions (S1 Fig) which has also been described for several other rickettsiae.
Analyses of nucleotide fragments obtained by pyrosequencing also resulted in 1,917 nt (accession number MF163040) that revealed 97% (1,863 nt/1,918 nt, 13 gaps) similarity to the partial protein PS 120 (D) gene sequence of R. helvetica type strain C9P9 (accession number AF163009) but 99% (1,901/1,918 nt, 13 gaps) to the partial sca4 nucleotide sequence of R. asiatica strain IO-1 (accession number DQ110869). Four partial R. helvetica sca4 gene sequences found in GenBank were identical (767/767 nt, accession number KR150775) or revealed 99% similarity (accession numbers KT825971, KT825970, FJ358501) to the sequence of R. helvetica AS819. The open reading frame resulted in a 639 aa sequence that shared 95% (599/632 aa, 4 gaps) similarity to the partial protein PS 120 sequence of R. helvetica type strain C9P9 (accession number AAL23857) and revealed 99% identity (630/637 aa, 4 gaps) to the partial Sca 4 sequence of R. asiatica strain IO-1 (accession number AAZ83584). Within the aa sequence two stretches were recognized that resemble vinculin-binding sites (Fig 8).
Furthermore, a 4,884-nucleic acid-stretch (accession number MF163039) with highest identity (99%, 4,858/4,888 nt, 7 gaps) to the R. helvetica C9P9 complete sca2 gene sequence (accession number AY355375) was obtained. This sequence contains also multiple stop codons and therefore seems to be a pseudogene. Therefore, the predicted aa sequence is disrupted compared to other rickettsiae (S2 Fig) most likely resulting in a non-functional protein.
The pathogenicity of R. helvetica has not been investigated in depth to date. Its main vector, I. ricinus, is considered as a generalist with an extraordinarily broad host spectrum including mammals, birds, and reptiles [33]. Hence, potential transmission of R. helvetica to humans seems likely. In humans, seroprevalences up to 12.5% against R. helvetica have been demonstrated with forest workers being predisposed to infection [34–37]. However, fulminant clinical cases are rare indicating a rather low pathogenicity compared to other rickettsiae.
There has been only one study describing the life cycle, growth characteristics and host cell response of R. helvetica in a Vero cell line [24] besides the first description of this species [25]. Concurrent to the results by Elfving et al. [24] we observed a short lag phase of up to four days after infection of the monolayers. However, in contrast to that study [24] R. helvetica AS819 grew rather slowly. One possible explanation might be that host cell fragments in our seed culture competed with intact host cells for rickettsial attachment thereby decreasing uptake efficiency as has been described for R. prowazeki seeds [38]. Elfving et al. [24] also reported a lag phase but used suspensions of lysed cells. Moreover, the culture supernatant of our seed culture may have contained an unknown amount of dead or late-growth-phase rickettsia that were no longer in an active growth state [38]. The latter would lead to a lag phase in the intracellular growth [38] as was noticed in our experiments. The percentage of viable R. helvetica AS819 was not assessed hence, the copy numbers calculated for the seed may have included a large amount of dead organisms. The differences between the Swedish and our study might also be attributed to the different culture techniques used in the respective study: conventional culture in the study at hand versus shell vial centrifugation technique used by the Swedish colleagues [24]. The latter technique has been described to increase infectivity [39] but does not resemble the host-pathogen-interaction during natural infection. No intranuclear R. helvetica AS819 was detected in the first description [25] and in our experiments, which is in addition in contrast to the study by Elfving et al. [24] and might also result from the centrifugation step used in their study. We showed that R. helvetica AS819 grew constantly but not rapidly after a phase of adaptation in both tested Vero E6 and L929cell lines, which was confirmed by the LSM investigations. LSM revealed that the dissemination of R. helvetica AS819 in both cell lines was rather mediated by cell break-down and bacterial release than cell-to-cell spread as has also been described for R. prowazekii [40]. This was indicated by the irregularly infected cell layers. R. helvetica AS819 did not produce cytopathogenic effects in Vero E6 and L929 cells which is in agreement with the first description of this species [25]. In addition, actin polymerization due to R. helvetica AS819 was not detected in both cell lines which has also been described for R. helvetica elsewhere [25]. This adds to the previous suggestion that R. helvetica lacks intracellular motility which makes it unlikely to invade nuclei [25,41]. Moreover, plaque formation in Vero E6 and L929 cells was absent after 21 d of infection suggesting that R. helvetica does not spread from cell to cell. Interestingly, this is in contrast to Rolain et al. [42] who reported formation of small plaques after up to 8 d of incubation using the same cell lines. This difference might be due to strain-specific variation in growth characteristics and virulence as has been described for different strains of R. rickettsii [43–45]. Moreover, the number of serial passages of cell lines and rickettsia can influence growth characteristics [42–43,46].
RickA-mediated nucleation of actin plays a role in the intracellular spread of some species [16,47]. Although R. helvetica AS819 possesses a gene encoding for RickA, a bacterial actin nucleator most closely related to WASP/N-WASP-family proteins [19], no host actin polymerization (HAP) was observed. This has also been described for R. raoultii where rickA expression is not sufficient to promote actin-based motility [29]. Furthermore, R. felis and R. parkeri possess genes encoding full-length RickA but lack spread by HAP [48]. A disruption of the rickA coding sequence as has been shown in R. peacockii and REIS [49–50] was not confirmed in R. helvetica AS819 by our sequence analyses. In addition to RickA, Sca2 has been suggested to be necessary for the actin-based motility of rickettsiae [18]. Cardwell and Martinez [22] demonstrated that the first third of the Sca2 passenger domain is highly conserved among SFG rickettsia with the exception of a 39 amino acid deletion. They showed that the deletion of residues 309 to 347 resulted in complete abolishment of actin assembly in R. conorii [22]. Sca2 aa-sequence analysis from R. helvetica AS819 revealed a disrupted aa-sequence resulting in a non-functional protein. This is in concordance to the R. helvetica type strain C9P9. In this strain, sca2 has been deemed a pseudogene with one or more fragments that don’t span the complete protein [48]. Most likely, this might be responsible for the lack of actin tails observed in R. helvetica AS819-infected cell lines. Sca2 also seems to play an important role in the initial bacterial-host interaction and Sca2 of R. conorii mediates both adhesion and invasion of mammalian cells in vitro [22]. However, our R. helvetica isolate AS819 did not exhibit an appreciable defect in adherence or invasion of Vero and L929 cells in vitro. This might be attributed to other highly conserved proteins which may compensate for the lack of functional Sca2 [22]. Only recently, Sca4 was identified as another protein from SFG rickettsia that promotes spread [23]. Specifically, Sca4 of R. parkeri binds to the cell-adhesion protein vinculin and inhibits its activity thereby reducing intercellular tension forces [23,51]. For R. helvetica isolate AS819 a 639 aa-stretch was identified that revealed 95% similarity to Sca4 of R. helvetica C9P9. For the latter species Sca4 has been supposed to be a truncated protein [48] which most certainly might also be the case for the strain investigated in this study.
Cell-to-cell spread is one crucial step in the intracellular life cycle of several pathogens including rickettsiae. Actin-based motility contributes to cell-to-cell spread and dissemination within the host. In contrast to other rickettsiae, R. helvetica isolate AS819 did not spread directly from cell to cell by actin-based motility presumably due to a deletion in the predicted Sca2 protein. As Sca2 is needed for virulence [14] our results suggest less virulence and pathogenicity of R. helvetica isolated from ixodid ticks in Germany.
|
10.1371/journal.pgen.1007157 | The Arabidopsis SUMO E3 ligase SIZ1 mediates the temperature dependent trade-off between plant immunity and growth | Increased ambient temperature is inhibitory to plant immunity including auto-immunity. SNC1-dependent auto-immunity is, for example, fully suppressed at 28°C. We found that the Arabidopsis sumoylation mutant siz1 displays SNC1-dependent auto-immunity at 22°C but also at 28°C, which was EDS1 dependent at both temperatures. This siz1 auto-immune phenotype provided enhanced resistance to Pseudomonas at both temperatures. Moreover, the rosette size of siz1 recovered only weakly at 28°C, while this temperature fully rescues the growth defects of other SNC1-dependent auto-immune mutants. This thermo-insensitivity of siz1 correlated with a compromised thermosensory growth response, which was independent of the immune regulators PAD4 or SNC1. Our data reveal that this high temperature induced growth response strongly depends on COP1, while SIZ1 controls the amplitude of this growth response. This latter notion is supported by transcriptomics data, i.e. SIZ1 controls the amplitude and timing of high temperature transcriptional changes including a subset of the PIF4/BZR1 gene targets. Combined our data signify that SIZ1 suppresses an SNC1-dependent resistance response at both normal and high temperatures. At the same time, SIZ1 amplifies the dark and high temperature growth response, likely via COP1 and upstream of gene regulation by PIF4 and BRZ1.
| Ambient temperature is a major actor in plant immunity and in growth regulation. Foremost, high temperature (>27°C) is known to block plant defence responses. High temperature also alters the plant morphology by inducing elongation growth, which facilitates plant ‘cooling’. This process is called thermomorphogenesis. Importantly, the SUMO E3 ligase SIZ1 suppresses plant immunity at normal conditions (22°C), but its role in immunity at high temperature was unknown. SIZ1 was recently shown to sumoylate and activate the ubiquitin E3 ligase COP1, a key player in thermomorphogenesis affecting the accumulation and/or stability of key transcription factors for this process (PIF4 and HY5). At high temperature PIF4 suppresses SNC1-dependent growth defects and auto-immunity for the snc1-1 mutant. We report that part of the SNC1-dependent auto-immune response is retained and activated in the siz1 mutant at high temperature resulting in enhanced resistance to Pseudomonas. In addition, we find that SIZ1 controls the thermomorphogenesis response and it affects expression of a substantial subset of PIF4 and BZR1 gene targets in response to high temperature. Our data imply that SIZ1 acts upstream of the PIF4/BZR1 hub. Combined the data highlight that SIZ1 has a dual role in the trade-off between SNC1-dependent immunity and growth at elevated temperature, where the latter aspect potentially runs via COP1.
| Ambient temperature is a major factor that affects plant growth and development, but also plant immunity [1,2]. In particular, the temperature range of 16-32ºC modulates the output of many plant immune receptors. For example, the tobacco N (Necrosis) gene fails to trigger resistance against Tobacco mosaic virus (TMV) at 30°C, while conferring resistance at 23°C [3]. This is accompanied by the loss of the hypersensitive response (HR) above 27°C. This HR includes a localized cell death that appears to be associated with recognition of pathogen effectors resulting in effector-triggered immunity (ETI) [4–7]. Multiple examples of high temperature suppression of ETI have been described for the TNL-type of immune receptors (Toll Interleukin-1 receptor [TIR], NB-LRR-type) [2], including the tobacco immune receptor N against Tobacco mosaic virus (TMV) [7,8], but also resistance mediated by the Arabidopsis immune receptor RPS4, which recognizes the avirulence protein AvrRPS4 from Pseudomonas, is suppressed at high temperature [9]. Finally, SNC1 (Suppressor of npr1-1, constitutive 1) dependent auto-immunity in the gain-of-function mutant snc1-1 is suppressed at high temperature [10]. Auto-immunity in the snc1-1 mutant was caused by hyperaccumulation of a mutant variant of SNC1 resulting in a dwarf stature of the mutant plant with curly leaves at 22°C [11]; At 28°C this auto-immune phenotype of snc1-1 is fully suppressed yielding plants with wild type rosettes without any macroscopic lesions or microscopic cell death. Importantly, HR activation by SNC1 required nuclear localization of SNC1, which appeared to be compromised when plants were kept at 28°C [6,7,12].
In non-infected plants, SNC1 levels are tightly controlled at both the transcript and protein level to prevent spurious immune signalling [13]. The expression of SNC1 is, for example, indirectly negatively regulated by the plasma membrane-localized protein BON1 (Bonzai 1) [14], but also the protein levels of SNC1 are regulated e.g. by the immune adaptor SRFR1 (Suppressor of RPS4-RLD 1) [15,16], several protein folding chaperones [17], and the F-box protein CPR1 (Constitutive expressor of Pathogenesis-related (PR) proteins 1) [11,18]. Mutations in the corresponding genes (e.g. snc1-1, bon1, srfr1-4 and cpr1-2) cause SNC1-dependent auto-immunity (hereafter SNC1auto-I). SNC1auto-I relies on EDS1 and PAD4 (Enhanced disease susceptibility 1, Phytoalexin-deficient 4) [19]. Upon recognition of biotrophic pathogens, EDS1 translocates from the cytoplasm, where it is sequestered by the related protein PAD4, to the nucleus [20–23]. Nuclear localization of EDS1 is necessary for transcriptional reprogramming to trigger SA biosynthesis and other plant defence responses.
Strikingly, high temperature suppression of auto-immunity depends for the snc1-1 mutant on the central growth regulator PIF4 (Phytochrome Interacting Factor 4), a transcription factor (TF) that is essential for thermomorphogenesis at 28°C [24]. This implies that plant growth is prioritized over SNC1-dependent auto-immunity at 28°C via transcriptional regulation. High ambient temperature increases PIF4 activity by controlling both its transcript levels and protein levels in a diurnal dark/light cycle [25]. This process is directly affected by relocalization of the ubiquitin E3 ligase COP1 (Constitutive Photomorphogenesis 1) to the nucleus in dark conditions. In the nucleus COP1 targets key regulators of both PIF4 protein activity and PIF4 gene expression for degradation [26]. Recent data highlight that COP1 is not only essential for the dark-induced growth response, but also at high ambient temperature in a normal diurnal dark/light cycle [27].
Here we studied auto-immunity in a mutant of the Arabidopsis SUMO E3 ligase SIZ1. Auto-immunity of siz1 highly resembles SNC1auto-I [28,29], i.e. the mutant shows enhanced resistance to Pseudomonas infection due to high levels of SA, its rosette adopts a very similar morphology (including lesions and spontaneous cell death) as the SNC1auto-I mutants, and this auto-immune phenotype depends on PAD4. Auto-immunity in the siz1 mutant is likely caused by the absence of sumoylation on one or more of its substrates, as the sumo1/2KD knock-down mutant also displays auto-immunity [29]. SIZ1 is the major SUMO E3 ligase in Arabidopsis [30], affecting SUMO conjugation of many substrates including pivotal regulators of growth [31–33]. For example, COP1 is a direct substrate of SIZ1 and its sumoylation enhances the intrinsic ubiquitin E3 ligase activity of COP1 [34,35].
As PIF4 controls the high temperature-mediated recovery of snc1-1 auto-immunity and SIZ1 controls the activity of a key regulator of PIF4, namely COP1, we assessed here (i) whether the siz1 auto-immune phenotype requires a functional SNC1 gene copy at normal and high temperature. Moreover, we tested (ii) if loss of SIZ1 function suppresses the COP1/ PIF4 mediated growth response at high temperature and in dark conditions. We found that siz1 auto-immunity is sustained at 28°C resulting in enhanced resistance to bacteria, which depended on both SNC1 and EDS1. The dwarf stature of siz1 also hardly recovered at 28°C. Moreover, we found that siz1 shows a compromised thermosensory growth response, which was independent of SNC1 and PAD4. This positive regulatory role of SIZ1 in growth regulation was suppressed by the TF HY5 (Elongated hypocotyl 5) at 22°C, while it depended on COP1 function at 28°C (and in dark conditions). HY5 is a direct substrate for COP1 targeted protein degradation. Finally, we found that high temperature induced transcriptome changes are both attenuated and delayed in the siz1 and sumo1/2KD mutants and that a substantial subset of the affected genes are known genomic targets for PIF4 binding and regulation.
A hallmark of SNC1auto-I is a dwarf stature and curled leaves. These morphological defects disappear when SNC1auto-I mutants like cpr1-2, bon1, snc1-1, and srfr1-4 are grown at 28°C, adopting a wild type stature (Fig 1A and 1B). Here we tested if also for siz1 these morphological defects are rescued when it grows at high temperature. In contrast to the four aforementioned SNC1auto-I mutants, we observed that siz1 remains significantly smaller than the wild type control at 28°C (Fig 1A and Fig 1B, compare group ‘cd’ with group b). At 22°C, the rosette weight of siz1 was indistinguishable from these four SNC1auto-I mutants (Fig 1B, group ‘d’). Previous work by others had shown that the auto-immune phenotype of these SNC1auto-I mutants depends on (i) a functional gene copy of PAD4 and EDS1, and (ii) accumulation of the defence hormone SA [10,16,36,37]. Likewise, Lee and co-workers demonstrated that the siz1 phenotype (partially) depends on PAD4 and SA accumulation [28], but the role of EDS1 remained unknown. Since EDS1 is the major nuclear actor of the PAD4/EDS1 hub [22,38] and SIZ1 is considered to primarily act in the nucleus [39], we examined if siz1 auto-immunity depends on EDS1. The siz1 growth defect partially recovered when it was crossed with the eds1-2 mutation in the Col-0 background, but this recovery did not significantly differ from the recovery seen for the double mutants siz1 pad4 and siz1 NahG (a transgene encoding salicylate hydroxylase that effectively prevents SA accumulation by converting it to catechol) at 22°C (Fig 1C and 1E; all post hoc group ‘c’). We also crossed siz1 with a mutant for SID2 (Salicylic acid induction deficient 2), which encodes the key enzyme for SA synthesis in plant immunity [40]. As seen by others for other auto-immune mutants [41], introduction of the sid2 mutation did not rescue the siz1 growth defect seen at 22°C (Fig 1C and 1E, group d). Importantly, at 28°C none of the siz1 double mutants showed any additional growth recovery compared to siz1 alone (Fig 1D and 1E, group c). This suggests that the small growth recovery of siz1 seen at 28°C (Fig 1E, from only ‘d’ at 22°C to ‘cd’ at 28°C) is potentially linked to suppression of its auto-immune phenotype, which in turn would depend on EDS1/PAD4 and SA accumulation.
Hence, we assessed if other hallmarks of the SNC1auto-I phenotype are also partially rescued when siz1 is grown at 28°C. We looked at spontaneous cell death, expression of defence-related genes (PR1, PR2, and SNC1), and accumulation of the encoded PR proteins. Both spontaneous cell death and PR1 expression are known (i) to strongly depend on EDS1/PAD4 and SA accumulation, and (ii) to be suppressed at 28°C in the aforementioned SNC1auto-I mutants. Spontaneous cell death was fully suppressed when siz1 was grown at 28°C (Fig 1F). At 22°C spontaneous cell death was lost in the double mutants siz1 pad4, siz1 eds1 and siz1 NahG (Fig 1F), indicating that EDS1/PAD4 and SA accumulation are required for the spontaneous cell death in siz1. At 22°C expression of PR1 and PR2 was also strongly up-regulated in siz1 compared to the control (Col-0) and expression of both genes required EDS1, PAD4 and SA accumulation (Figs 2E, S1A and S1B). At 28°C, PR1 expression was completely suppressed in siz1, but PR2 expression partially remained (S1B Fig). This situation was reflected in their protein levels, i.e. PR1 levels were high in siz1 at 22°C while undetectable at 28°C (Fig 1G). In contrast, PR2 levels were elevated in siz1 both at 22°C and 28°C albeit to a lower level at 28°C. In the case of the four SNC1auto-I mutants, PR1 and PR2 did not accumulate when these mutants were grown at 28°C (Figs 1G and 2D). Thus, the siz1 auto-immune response is (partially) temperature sensitive, but it does not simply mimic the ‘classic’ behaviour of SNC1auto-I mutants.
As elevated expression of SNC1 triggers auto-immunity at 22°C [42], we measured SNC1 expression in siz1. SNC1 expression proved to be induced by nearly 5-fold in siz1 at 22°C (S1C Fig), suggesting that an increase in SNC1 transcript levels could be causal for the siz1 dwarf stature and auto-immunity. To determine if the SNC1 gene is indeed required for the siz1 phenotype at 22°C/28°C, we crossed siz1 with a loss-of-function mutant of SNC1, snc1-11 (SALK_04705). This mutant has a T-DNA insertion in the first exon, which results in a severely truncated transcript [42]. When grown at 22°C, the siz1 snc1-11 double mutant displayed a small but significant growth recovery compared to siz1 (Fig 2A and 2B; group ‘c’ and ‘d’, respectively), which is more apparent when the plants are flowering (S2 Fig). However, in our conditions the snc1-11 mutant itself also displayed a small but significant increase in biomass compared to the wild type control (Col-0) at 22°C (Fig 2B). More importantly, both siz1 and the siz1 snc1-11 double mutant largely kept their dwarf stature when grown at 28°C. This is striking, as the growth defects of the SNC1auto-I mutants cpr1-2, bon1 and srfr1-4 recovered strongly (to wild type levels) when the snc1-11 mutation was introduced in these mutants by crossing [10,15,18]. The increase in SNC1 transcript levels can, therefore, not be the main or sole cause of the dwarf stature of siz1.
Nonetheless, spontaneous cell death was fully suppressed in siz1 snc1-11 at 22°C (Fig 2C), while PR2 and to a lesser extent PR1 still accumulated in siz1 snc1-11 at 22°C (Fig 2D). Also at 28°C PR2 still accumulated to some extent in siz1 snc1-11, similar to siz1 (Fig 2D). The PR1 and PR2 protein levels were again mirrored by their gene expression levels (Fig 2E), i.e. at 22°C the expression of PR1 was roughly 50% in siz1 snc1-11 in comparison to siz1, which in both cases was fully suppressed when these two mutants were grown at 28°C. On the other hand, PR2 expression remained detectable when both mutants were grown at 28°C. Also the (truncated) transcript of SNC1 still accumulated to higher levels in siz1 snc1-11 than in siz1. For snc1-11, 2–3 samples showed up-regulation of PR1 and PR2, while the remaining samples 5 samples showed hardly any up-regulation suggesting that the latter samples reflect the general trend.
Increased SNC1 protein levels are known to trigger auto-immunity [11]. SNC1 levels are negatively controlled by the HSP90/SGT1/SRFR1 chaperone-complex of which some components were reported to be SUMO substrates [43,44]. We therefore examined whether siz1 auto-immunity was attenuated when mutants for SGT1a, SGT1b, and RAR1 were introduced by crossing. Introduction of these mutants in siz1 (i.e. siz1 rar1, siz1 sgt1aKO and siz1 sgt1beta3) partially compromised cell death induction (S3A Fig), while it hardly enhanced rosette growth in these siz1 chaperone double mutants (S3B Fig). Hence, the chaperones contribute to the siz1 phenotype, but they are not essential for spontaneous cell death. Clearly, the siz1 auto-immune phenotype partially depends on SNC1, but not all of the elements of the auto-immune phenotype disappear when SNC1 is non-functional.
As the PR1 levels were down in siz1 at 28°C, we tested if enhanced resistance of siz1 to the pathogen Pseudomonas syringae pv. syringae strain DC3000 (PstDC3000) is compromised at high temperature. In order to inoculate similar looking plants, all plants were grown at 28°C and half of the plants was shifted to 22°C twenty-four hours prior to the inoculation. In this way extreme differences in rosette size, morphology, or tissue structure had no impact on the disease assay (compare the plants grown at 28°C in Fig 1A). The 24 hours pre-incubation at 22°C was sufficient to re-activate auto-immunity in the SNC1auto-I mutants tested (cpr1-2, bon1, snc1-1) resulting in reduced susceptibility to PstDC3000 (Fig 3A, post hoc groups ‘cd’ and ‘d’). As expected, the three tested SNC1auto-I mutants (cpr1-2, bon1, and snc1-1) were as susceptible as the wild type control (Col-0) at 28°C (Fig 3B). However, siz1 displayed enhanced resistance to PstDC3000 both at 22°C and 28°C (Fig 3A and 3B). This resistance was compromised in siz1 at 22°C when PAD4, EDS1 or SNC1 were mutated (Fig 3A). At high temperature, only siz1 pad4 retained enhanced resistance to PstDC3000, while siz1 eds1-2 and siz1 snc1-11 were both as susceptible as the wild type control (Fig 3B). This means that enhanced resistance of siz1 to the pathogen PstDC3000 at 28°C was still dependent on EDS1 and SNC1.
In the case of snc1-1, high temperature suppression of immunity and restoration of growth were both reported to depend on PIF4 [24]. Therefore, we also tested if the pif4-2 mutant showed altered resistance to PstDC3000 at 22°C/28°C. The pif4-2 plants showed a clearly compromised thermomorphogenesis response at 28°C, i.e. (i) the hypocotyl length was reduced (Fig 4A and 4B), (ii) the rosette showed no hyponasty and (iii) the leaf blades and petioles failed to elongate in comparison to Col-0. However, the pif4-2 mutant was as susceptible to PstDC3000 as the wild type control (Col-0) at either temperature in our conditions (Fig 3).
As snc1-1 auto-immunity is inhibited by PIF4 at high temperature [24], the enhanced immunity of siz1 to PstDC3000 at 28°C might also be due to reduced PIF4 function. In line with this notion, we found that siz1 and the sumo1/2KD mutant both showed reduced hypocotyl elongation at 28°C in normal diurnal dark/light cycles (Fig 4A, light; 4B, compare 22C L with 28C L), implying that SIZ1 and the two archetype SUMO proteins, SUMO1 and SUMO2 (hereafter SUMO1/2), act as positive regulators of thermomorphogenesis similar to PIF4 (pif4-2 was included as control for the loss of thermosensitive hypocotyl elongation; Fig 4A and 4B). SIZ1 and SUMO1/2 were both also needed for skotomorphogenesis (dark-induced hypocotyl elongation) at 22°C and 28°C (Fig 4A, dark; 4B, compare 22C L with 22C D). The compromised dark and high temperature growth responses were both independent of PAD4 and SNC1, as they still occurred to same extent in siz1 pad4 and siz1 snc1-11 (Fig 4B). This means that not the auto-immune phenotype of siz1 is responsible for the compromised thermo/skotomorphogenesis, but rather that SIZ1 itself acts as positive regulator of these growth responses. In support of this notion, we confirmed that the SNC1auto-I mutants cpr1, bon1, and srfr1-4 display a normal thermomorphogenesis response (S4 Fig), indicating that PIF4 function is unaffected in them. Moreover, the sumo1/2KD consistently showed a stronger reduction in hypocotyl elongation than siz1 nearing pif4-2 at the 28°C in a normal dark/light cycle (Fig 4B, 28°C L).
The mutants siz1 and sumo1/2KD also displayed a strong reduction in hypocotyl elongation when they were kept in the dark at 22°C and 28°C (Fig 4B; panels 22C D, 28C D). As SIZ1 stimulates COP1 activity and the nuclear function of COP1 is activated in the dark [34,35], we examined whether loss of SIZ1 function could enhance the thermo/skotomorphogenesis phenotype of a strong but not lethal COP1 mutant, cop1-4 [45]. Hypocotyl elongation was indeed more reduced in siz1 cop1-4 than in cop1-4 alone in dark conditions at 22°C and 28°C (Fig 4C; panels 22C D, 28C D). Thus, COP1 is critical for the thermosensory growth response–as recently reported [27], while SIZ1 appears to primarily enhance this response (as further detailed below).
In light conditions, the TF HY5 is known to inhibit hypocotyl elongation by inhibiting PIF4 expression [25]. COP1 targets HY5 for proteasomal degradation when COP1 is active in the nucleus. We found that SIZ1 function is needed for the full hypocotyl elongation of the HY5 loss-of-function mutant hy5-215 in a diurnal light/dark cycle at 22°C (Fig 4C, panel 22C L). This means that in a diurnal light/dark cycle at 22°C the stimulatory role of SIZ1 on hypocotyl growth is masked by the inhibitory role of HY5. We also compared the rosette size and morphology of siz1 cop1-4 and siz1 hy5-215 with the single mutants at both temperatures (S5 Fig). At 22°C siz1 cop1-4 and siz1 hy5-215 both adopted a siz1 rosette size/morphology. At 28°C growth was recovered for siz1 hy5-215, but to a lesser extent than for siz1. In contrast, siz1 cop1-4 failed to respond to the high temperature and this mutant still closely resembled cop1-4 mutant (having a compact rosette with hardly any petioles and no hyponasty; S5A Fig). This is consistent with a model in which COP1 primarily conveys the thermosensory growth response and that SIZ1 amplifies the output of this response.
As biosynthesis of the hormones gibberellic acid (GA3) and the brassinosteroids is needed for the temperature induced hypocotyl elongation [46], we checked if the positive regulatory role of SIZ1 and SUMO1/2 in thermomorphogenesis requires these two hormones. First, we inhibited GA3 or BR biosynthesis by adding paclobutrazol (PAC) or propiconazole (PPZ), respectively. Irrespective of the genetic background, we found that biosynthesis of both hormones was essential for the temperature-induced hypocotyl elongation in the lines tested including the residual elongation in pif4-2 (Fig 4D and 4E). GA3 is known to reduce the abundance of the DELLAs by triggering their degradation [47]. In turn the DELLAs restrain cell growth by reducing protein abundance of the PIFs (including PIF4) and the TF BZR1 (Brassinazole resistance 1) [48,49]. A combined treatment of 28°C+GA3 resulted in increased hypocotyl elongation for each of the four tested lines compared to the 28°C control (-) (Fig 4D). However, hypocotyl elongation was still impaired for siz1, sumo1/2KD and pif4-2 in the combined treatment 28°C+GA3 (Fig 4D). This implies that the positive role of SIZ1 on temperature-induced hypocotyl growth is independent of DELLA accumulation. The combined treatment of 28°C plus the brassinosteroid Brassinolide (28°C+BL) triggered a small but significant increase in hypocotyl elongation in the control (Col-0) plants compared to the mock treatment (Fig 4E,—vs. BL). However, the SUMO mutants (siz1 and sumo1/2KD) showed no additional response to the combined treatment 28°C+BL (Fig 4E). Strikingly, the pif4-2 mutant did respond to the BL treatment (from post hoc group D to C), suggesting that in siz1 and sumo1/2KD brassinosteroid signalling is apparently already at its maximum physiological level.
To elucidate how SIZ1 and SUMO1/2 conjugation affect high temperature-induced gene expression, we grew siz1 pad4 and the sumo1/2KD pad4 mutants for two weeks at 22°C and then shifted them to 28°C (4 hrs after light onset) to trigger a temperature induced transcriptional response. To avoid that constitutive (auto-)immune signalling impedes the thermosensory transcriptional response at t = 0, we performed the experiment in the pad4 background, which largely blocked siz1 auto-immunity at 22°C (i.e. the enhanced accumulation of PR1 and PR2, spontaneous cell death and the increased resistance to PstDC3000 are suppressed in siz1 pad4; Figs 1F, 1G and 3A), but it only partially restored the dwarf stature. Importantly, increased resistance to PstDC3000 was not lost in siz1 pad4 at 28°C, similar to siz1 (Fig 3B). The plants were sampled at the shift to 28°C (day 0) and 24hrs (day 1) and 96 hrs (day 4) after the shift. Catala et al. had previously shown that siz1 shows a strong up-regulation of defence-related genes (like PR genes and immune receptors), while genes involved in BR biosynthesis/signalling are down-regulated [50]. We first determined which genes are differentially expressed at 22°C in siz1 pad4 in comparison to the control (pad4). We found that a small set of genes encoding for TNL immune receptors, Receptor-like kinases (RLKs), and Receptor-like proteins (RLPs) remained up-regulated in siz1 pad4 in comparison to the control (pad4) at 22°C (S1A Table). SNC1 or immune receptors of the CNL type (Coiled-coil NB-LRR-type) were not amongst the up-regulated genes in the microarray data. Real time PCR revealed that SNC1 was roughly two-fold induced in siz1 pad4 (close to the cut-off value for differential gene expression), while SNC1 showed no up regulation in siz1 eds1 or siz1 NahG (S1C Fig). As SNC1 was 5-fold induced in siz1 at 22°C (S1C Fig), we conclude that this requires feedback regulation via EDS1 and SNC1.
There was no broad up-regulation of TF families linked to plant immunity (WKRY, TGA or MYC family) in siz1 pad4 at 22°C. Likewise, PR genes like PR2, PR3, or PR4 were no longer strongly up-regulated in siz1 pad4 at 22°C. The genes involved in BR biosynthesis and signalling were also no longer collectively down-regulated except for two genes, which encode for two rate-limiting enzymes of the Brassinosteroid (BR) biosynthesis pathway (DWF4 or DWARF 4; and BR6OX2 or BRASSINOSTEROID-6-OXIDASE 2) [51–53]. This suggests that the BR levels might be reduced in siz1 pad4. In agreement with this, we found that the TFs BEE1, BEE3, and TCP1 are down-regulated in siz1 pad4 (S1B Table). BEE1 and -3 are two closely related bHLH TFs that act as early response TFs required for the full BR response [54]. TCP1 encodes a TF that directly positively regulates the expression of DWF4 [55]. Combined, these data argue that the siz1 pad4 phenotype may be (partially) due to BR-deficiency.
We then selected the set of thermosensitive genes by identifying the genes that are differentially expressed (DEGs, q ≤ 0.01) in pad4 in response to the shift to 28°C (comparing day 1 to day 0, day 4 to day 1, and day 4 to day 0). The DEGs were clustered based on their expression profile and their expression dynamics was revealed by plotting their standardized expression values in a clustered heat map (Fig 5A, red-to-blue). To detect differences in the gene expression profiles of siz1 pad4 and sumo1/2KD pad4 we plotted the same gene expression heat maps for the two mutants while retaining the gene clustering (Fig 5A). We also plotted the difference in gene expression (Δ) between the mutants (siz1 pad4 and sumo1/2KD pad4) and pad4 (brown-to-cyan heat maps). Fig 5A reveals that overall the gene expression profiles of the thermosensitive genes do not differ strongly between the two SUMO conjugation mutants and the control pad4 (blue-to-red heat maps). In other words, most of the thermosensitive genes also respond to the shift to 28°C in siz1 pad4 or sumo1/2KD as they do in pad4.
However, most of the thermosensitive genes appear to show an attenuated response in siz1 pad4 and sumo1/2KD pad4 at day 1 and/or 4. For example the up-regulated genes (changing from red at day 0 to blue at day 4 in the heat maps) show less expression in siz1 pad4 and sumo1/2KD pad4 than pad4 at day 4 (brown colour in the ‘ΔExpr (mut-WT)’ heat maps). Likewise, the down-regulated genes (shift from blue at day 0 to red at day 4) show increased expression in siz1 pad4 and sumo1/2KD pad4 at day 4 (cyan colour in the ‘ΔExpr (mut-WT)’ heat maps). To confirm this notion, we selected for each time point the DEGs in siz1 pad4 and sumo1/2KD pad4 in comparison to pad4. A large set of these DEGs was shared between the two mutants (siz1 pad4 and sumo1/2KD pad4), as can be seen in the VENN diagrams (Fig 5B). Strikingly, the largest number of DEGs was obtained for both mutants at day 1 rather than at day 4. To visualize the dynamic response of these DEGs in response to high temperature, we plotted in a scatter plot the fold change in expression of these DEGs for siz1 pad4 and sumo1/2KD pad4 (both y-axis) versus pad 4 (x-axis) (by separately combining the DEGs for the different time points for the two mutants). The left panel in Fig 5C and 5D depicts the change in expression from day 0 to day 1, while the right panel depicts the change from day 0 to day 4. This revealed that primarily in the control (pad4) at day 1 the expression of the DEGs changed due to the increase in temperature, while in siz1 pad4 and sumo1/2KD pad4 these genes largely failed to respond at this time point (Fig 5C and 5D, panel day 1–0). This is best seen in the global expression profiles (top and right side of the scatter plot) revealing a double hump in pad4, while the expression profile displays a single Gaussian curve around zero for both SUMO mutants. In contrast, at day 4 we find a positive correlation for the change in expression of all DEGs (Pearson R = 0.73; linear regression) with a slope = 0.61 for siz1 pad4 versus pad4. This means that at day 4 the DEGs responded in siz1 pad4 to the high temperature, but their response was overall attenuated. A similar situation is seen for sumo1/2KD pad4 at day 4 (Pearson R = 0.79; slope = 0.87). Thus, SIZ1 and SUMO1/2 both appear to control in a similar manner both the timing and the amplitude of the temperature-induced transcriptional response.
We then examined if the direct genomic targets of the TFs PIF4/BZR1/ARF6 are differentially expressed in siz1 pad4 and sumo1/2KD pad4. The direct genomic targets of these tree TFs, which form a trimeric transcriptional hub, were obtained from published chromatin-immunoprecipitation (ChIP) datasets of these TFs [56,57]. As shown in Fig 5E, nearly 25% of the genomic targets of these three TFs was differentially expressed in siz1 pad4 during the course of the temperature shift experiment. This overlap was very significant with p-values of 3.07e-21 (PIF4), 2.11e-8 (BZR1), 9.24e-11 (ARF6) using a hypergeometric test (based on 26859 annotated probes; TAIR9). The overlap was still significant but less strong for sumo1/2KD pad4 (with an overlap of ±12%, Fig 5E) and p-values of 1.17e-6 (PIF4), 6.92e-4 (BZR1), 1.46e-4 (ARF6). Thus, there is a significant enrichment for the genomic targets of these three TFs amongst the DEGs in both our mutants in response to shift to temperature 28°C (Fig 5E). The change in expression of these genomic targets of these three TFs in the mutants versus the control (pad4) mirrored largely the global pattern seen for all the DEGs combined (S6 and S7 Figs). Thus, the response of the misexpressed genomic targets of PIF4, BZR1, and ARF6 in the siz1 pad4 and sumo1/2KD pad4 mutants follows the same trend as the global response (i.e. their expression is largely delayed till day 4 and the response remains attenuated at day 4). This corroborates our hypothesis that the PIF4-dependent high-temperature growth response is compromised in siz1 and sumo1/2KD. We also looked at the genomic targets of the ‘cold’ regulator HY5 that binds to and competes (at low temperature) for the same genomic targets as PIF4 [58,59]. The HY5 genomic targets largely failed to respond in siz1 pad4 at day 1, while at day 4 their response was largely attenuated in siz1 pad4 compared to pad4 (S8C and S8D Fig). This effect on the expression of the HY5 genomic targets was less clear for sumo1/2KD pad4 (S9C and S9D Fig). While examining the list of DEGs we noted that many SAUR (Small auxin up RNA) genes were present among the top of the gene lists. PIF4 is known to regulate auxin biosynthesis via the SAUR family [60]. The differentially expressed SAUR genes showed a strong deregulation in siz1 pad4 and sumo1/2KD pad4 at both time points, with very distinct global expression profiles in the mutants versus the control (pad4) (S8E, S8F, S9E and S9F Figs). Combined, our data revealed that the siz1 pad4 and the sumo1/2KD pad4 mutants display a delayed and attenuated transcriptional response to high temperature (in comparison to pad4), which runs in part over the PIF4/BZR1 transcriptional hub.
Here, we describe an interconnected dual role for SIZ1 and SUMO1/2 conjugation in the switch between plant immunity and high temperature induced growth (as summarized in the model of Fig 6). Our data unveil that both SIZ1 and SUMO1/2 conjugation are positive regulators of thermo- and skotomorphogenesis upstream of the PIF4/BZR1 growth regulation hub. In this hub, BZR1 is activated by the hormone BL, while PIF4 is activated by dark conditions and high ambient temperature. In line, these two TFs share a large number of genomic targets that are synergistically regulated by them [56]. We find that loss of SIZ1 and SUMO1/2 both delays and attenuates this transcriptional response to high temperature affecting many targets of PIF4 and BZR1. This suggests that SIZ1 activity acts as a positive regulator of PIF4 function in thermomorphogenesis and that PIF4 function is apparently compromised/inhibited in siz1 at high temperature (Fig 6, siz1-2). Importantly, the PIF4 protein abundance is positively regulated by COP1 E3 ligase activity [58], while COP1 activity is stimulated by SIZ1-dependent sumoylation (Fig 6, wild type route c.) [34,35]. Our data unveil that COP1 is essential to convey this high temperature signal, as recently reported by others [27], while SIZ1 enhances the high temperature and dark signal. This role of SIZ1 in thermo/skotomorphogenesis is distinct from its reported role on cell elongation due to constitutive defence signalling [61], as hypocotyl elongation was still compromised at high temperature when PAD4 or SNC1 were mutated. Likewise, we noted that the rosette of siz1 pad4, siz1 eds1, and siz1 NahG remained compact at 28°C (without strong petiole elongation or hyponasty as seen for Col-0).
Interestingly, part of the siz1 auto-immune phenotype is sustained at high temperature resulting in enhanced resistance to bacteria (Fig 3). This enhanced resistance still required SNC1 and EDS1 function at 28°C (Fig 3, Fig 6 wild type route a.). The latter is relevant, as both SNC1 and EDS1 immune signalling depend on their nuclear localization, while SNC1 nuclear localization is impaired at high temperature [7,12,22,62]. High temperature suppression of snc1-1 auto-immunity and concomitantly rescue of its growth phenotype requires PIF4 function [24] (Fig 6 wild type route b.). SNC1-dependent auto-immunity, including enhanced resistance to the bacterial pathogen Pseudomonas, is normally fully suppressed in the mutants bon1, crp1-2 and snc1-1 at 28°C, resulting in normal rosette growth (e.g. Figs 1–3) [7,63]. However, siz1 fails to resume normal growth at 28°C and this is independent of PAD4/EDS1, SNC1 or SA accumulation. This implies that the ‘high temperature’ signal is not properly conveyed in siz1. At the same time, SIZ1 suppresses expression of a small subset of immune receptors at 22°C, even when PAD4 is mutated. It remains an open question if elevated expression of one of these immune receptors (S1A Table) is causal for the auto-immune phenotype of siz1, rather than the misexpression of SNC1.
Biochemically, SUMO conjugation was already implied as a regulator of photomorphogenesis [34,35]. Our data suggest that the role of SIZ1 in thermomorphogenesis is mechanistically independent of light sensing, as hypocotyl elongation in siz1 was also reduced in the dark. Previous works had indicated that sumoylation of phyB allows PIF5 to bind its target promoters resulting in root growth stimulation. These authors demonstrated that sumoylation of the Pfr state (red light activated state) of phyB suppresses the interaction between phyB and PIF5, the closest homologue of PIF4 [32,64,65]. Our GA3 treatment experiment also suggests that SIZ1 controls thermomorphogenesis response independent of DELLA accumulation (Fig 4D). The DELLAs control the stability of the PIFs, while they themselves are also controlled by sumoylation [31,49]. Other (putative) sumoylation substrates implicated in PIF4 function are ELF3 (Early flowering 3) [43,66], HFR1 [67] and LAF1 [68], HY5 and HY5-like (HYL) [69]. The role of sumoylation has not yet been determined for ELF3. Both HFR1 and LAF1 are also targets for COP1-mediated degradation. The link between their degradation and sumoylation remains to be studied. Nevertheless, it is evident that (i) SUMO conjugation acts at multiple levels as a regulator of growth and that (ii) certain COP1 substrates are also targets for sumoylation.
Finally, we found that several actors in BR biosynthesis and signalling are still down-regulated (DWF4, BEE1, BEE3, and TCP1) in siz1 pad4. CESTA, a close homologue of BEE1 and BEE3, is another SUMO substrate that directly binds to BEE1 to control BR biosynthesis [70]. Catala and co-workers had previously reported that from the nearly 1600 differentially expressed genes in siz1 (>two-fold change), eleven down-regulated genes were known to be critical for BR biosynthesis and signalling [50]. In addition, they found in their genome-wide expression analysis that both PIF4 and PIF5 were underexpressed in siz1 [50]. These data warrant further research on the role of sumoylation on BL signalling and biosynthesis.
To conclude, SIZ1 and SUMO1/2 both act as important positive regulators of growth, while SIZ1 also acts as negative regulator of an SNC1-dependent immune response at high temperature. SIZ1 thus plays an interdependent dual role in growth and immunity at elevated ambient temperature.
The genetic resources for this research were wild type Arabidopsis (Arabidopsis thaliana) ecotype Col-0, siz1-2 [71], cop1-4 [45], cpr1-2 [18], bon1-1 [10], hy5-215 [72], snc1-1 [19], srfr1-4 [15], pad4-1 [73], eds1-2 (backcrossed in Col-0) [74], sid2-1 [40], 35Spro::NahG [75], snc1-11 (SALK_047058) [10], sgt1a-3 [16], sgt1b(eta3) [76], rar1-21 [77], pif4-2 [78], and sumo1/2KD [aka sum1-1 amiR-SUMO2 line B][29,79]. The double mutants pad4 siz1, NahG siz1 [28], cop1-4 siz1-2 and hy5-215 siz1-2 [35] are described elsewhere. Arabidopsis plants were grown under white light with 120 μmol m-2 sec-1 under short-day (SD) light conditions (11 hr light, 13 hr dark) at 22°C or 28°C on a compost/perlite soil mixture. After crossing, the plants were genotyped according to the primer combinations and primer sequences presented in the S2 and S3 Tables, respectively.
The fresh rosette weight of plants (minimum 8) grown individually in single pots was measured. The rosette was sampled from 5-week-old plants grown in parallel at 22°C or 28°C. Statistical analyses were made using two-way ANOVA (genotype, temperature, interaction GxT) followed by Tukey post hoc test in Prism7. Significantly different groups are indicated by letters.
For immunoblot analysis, seedlings or leaf material was homogenized in liquid nitrogen, thawed on ice in extraction buffer (10% glycerol, 50 mM K2HPO4/KH2PO4 pH 7.5, 150 mM NaCl, 1 mM EDTA, 2% w/v polyvinylpolypyrrolidone K25, 1× protease inhibitors (Roche), 1% v/v Nonidet P-40, 0.1% SDS and 5 mM DTT), and centrifuged for 10 min at 13,000g. The supernatant was mixed 1:1 with 2× SB (125 mM Tris-HCl pH 6.8, 4% SDS, 20% v/v glycerol, and 100 mM DTT), and the samples were boiled for 10 min. Proteins were separated on 15% SDS-PAGE and blotted onto Polyvinylidene fluoride (Immobilon-P, MIllipore) membranes. Secondary immunoglobulins conjugated to horseradish peroxidase were visualized using ECL Plus (GE Healthcare). Primary antibodies against PR1 (αPR1) were described previously [80] and αPR2 was obtained from Agrisera (#AS12 2366, ~35kDA). Incubation of both primary and secondary antibodies were done in Tris-buffered saline with 0.05% Tween-20 (TBST) followed by three rinses of 10 minutes in TBS. Equal protein loading was confirmed for the samples by Ponceau S staining of the membranes and when needed the loaded total protein amounts were standardized using BCA protein analysis on the total protein extracts prior to protein loading of the gels. The primary antibodies αPR1, αPR2 and the secondary antibody Goat-anti-Rabbit HRP (Fisher) were used at 1:5000, 1:2000 and 1:5000 dilutions, respectively.
For the gene expression analysis, total RNA was extracted from 100–200 mg of leaf material of 5-week-old plants grown at 22/28°C using TRIzol LS reagent (Fisher). The RNA was treated with DNase (ThermoFisher) according to the supplier’s protocol and RNA concentrations were determined by measuring the Abs(260) on a Nanodrop. cDNA was synthesised from 1 μg total RNA using RevertAid H reverse transcriptase in the presence of the RNAse inhibitor Ribolock (both ThermoFisher) following the supplier’s protocol. All biological samples were measured in technical replicate with 3–4 biological replicates per experiment. The PCR amplification was followed using Hot FIREPol EvaGreen qPCR (Solis Biodyne) in a QuantoStudio3 (ThermoFisher). Gene expression was normalized using two genes: Actin2 (At3g18780) and beta-Tub4 (At5g44340). The primers used are given in the S2 Table. The Ct values were corrected for primer efficiencies. All expression data were analysed using the pipeline in qBASE+ (Biogazelle).
Cold-stratified (3 days at 4°C) sterilized seeds (~50 per line) were placed on vertical plates with 1/2 MS medium supplemented with 1% w/v sucrose and 1% w/v Daishin agar (Duchefa). Seeds were irradiated with white light for 6 hrs to promote germination and then incubated in the specified light/temperature conditions for 5 days. The used seeds were fresh and from the same seed harvest. Seedlings were scanned and the hypocotyl lengths were measured using ImageJ (http://rsb.info.nih.gove/ij). Sensitivity to the Gibberellin biosynthesis inhibitor Paclobutrazol (Pac, Duchefa) and the hormone Gibberellic acid (GA3, Duchefa) was analysed by growing the seedlings on 0.5 μM PAC or 10 μM GA3, respectively. Likewise, sensitivity to the Brassinosteroid biosynthesis inhibitor Propiconazole (PPZ, Sigma-Aldrich) or the hormone 24-epiBrassinolide (BL, #b1439, Sigma-aldrich) was analysed by adding 2 μM PPZ or 0.1 μM BL to the plates, respectively.
Pseudomonas syringae pv. tomato DC3000 (PstDC3000) [81] (carrying the empty vector pVSP61) was freshly grown overnight at 28°C with 200 rpm in 10 mL Kings B broth [82] supplemented with rifampicin (50 μg/mL) and kanamycin (40 μg/mL) to reach an OD600 of ~0.9–1.2. Directly prior to infiltration, the bacterial suspensions were spun down, washed with 10 mM MgSO4, and resuspended at OD600 = 0.0002 (1×105 CFU/mL) in 10 mM MgSO4 for syringe leaf infiltrations. For the Pst disease assays the plants were germinated and grown at 28°C constant temperature (with 11L/13D) for 5 weeks in soil. Twenty-four hours prior to inoculation (9:00 am), one batch of plants was moved to 22°C (SD) and both plant sets were placed in propagators to increase humidity (>90%). The two plant batches were simultaneously infiltrated at 22°C and 28°C using the same bacterial suspension. Upon infiltration the plants were left to dry for 1.5–2 hrs after which they were again covered with lids for 72 hours to increase humidity (>90% relative humidity). Humidity and temperature was followed using a data logger inside the propagators for the duration of the experiment. Leaf discs were taken 1 hour after dipping (t = 0) at both temperatures and 72 hrs post infiltration (t = 3). At least 6 plants were infiltrated per condition. In total 8 samples were taken for each condition combining 2–3 leaf discs with a diameter of 5 mm. Leaf discs were taken from different leaves and only ‘mature’ fully elongated rosette leaves were sampled. The first-formed round shaped leaves were excluded from tissue sampling. The sampled intact leaves were surface-sterilized prior to taking leaf discs (10 sec dip in 70% ethanol followed by two washes with sterile water). The disease assays were performed with at least two independent replicates with similar results.
The rosette leaves were stained with a 1:1 mixture (v/v) of ethanol and lactic acid–phenol–trypan blue solution (2.5 mg mL−1 trypan blue, 25% v/v lactic acid, 25% phenol, 25% glycerol, and water) and boiled for 5 min. For destaining, the trypan blue solution was replaced with a chloral hydrate solution (2.5 g mL−1 in water), as described [83].
The siz1 pad4, sumo1/2KD pad4, and pad4 plants were grown on soil in SD conditions at 22°C for 2 weeks and then transferred to 28°C at noon (t = 0). Leaf samples were taken in triplicate for total RNA extraction at t = 0, 24 hrs (1d), and 96 hrs (4d). Total RNA was purified using the RNAeasy mini kit (QIAGEN). The RNA quality was examined by monitoring Abs(260/280) and the Abs(260/230) ratios. Total RNA (100 ng) was amplified using the GeneChip WT PLUS kit (Affymetrix) generating biotinylated sense-strand DNA targets. The labelled samples were hybridized to Arabidopsis Gene 1.1 ST arrays (Affymetrix). Washing, staining and scanning was performed using the GeneTitan Hybridization, wash, and stain kit for WT Array Plates, and the GeneTitan Instrument (both Affymetrix).
All arrays were subjected to a set of quality control checks, such as visual inspection of the scans, checking for spatial effects through pseudo-color plots, and inspection of pre- and post-normalized data with box plots, ratio-intensity plots and principal component analysis. Normalized expression values were calculated using the robust multi-array average (RMA) algorithm [84]. The experimental groups were contrasted to test for differential gene expression. Empirical Bayes test statistics were used for hypothesis testing [85] using the Limma package in R 3.2.1 (http://cran.r-project.org/), and all p-values were corrected for false discoveries according to Storey and Tibshirani [86]. Downstream statistical analyses (e.g. hypergeometric tests on enrichment) were performed in Python using the Scipy.stats module (https://scipy.org/scipylib/). The microarray data were deposited in Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/) under the accession number GSE97641 and Github (DEGs and scripts used to prepare Fig 5; https://github.com/LikeFokkens/Siz1_immunity-vs-growth_temperature).
|
10.1371/journal.ppat.1003697 | MicroRNA-155 Promotes Autophagy to Eliminate Intracellular Mycobacteria by Targeting Rheb | Mycobacterium tuberculosis is a hard-to-eradicate intracellular pathogen that infects one-third of the global population. It can live within macrophages owning to its ability to arrest phagolysosome biogenesis. Autophagy has recently been identified as an effective way to control the intracellular mycobacteria by enhancing phagosome maturation. In the present study, we demonstrate a novel role of miR-155 in regulating the autophagy-mediated anti-mycobacterial response. Both in vivo and in vitro studies showed that miR-155 expression was significantly enhanced after mycobacterial infection. Forced expression of miR-155 accelerated the autophagic response in macrophages, thus promoting the maturation of mycobacterial phagosomes and decreasing the survival rate of intracellular mycobacteria, while transfection with miR-155 inhibitor increased mycobacterial survival. However, macrophage-mediated mycobacterial phagocytosis was not affected after miR-155 overexpression or inhibition. Furthermore, blocking autophagy with specific inhibitor 3-methyladenine or silencing of autophagy related gene 7 (Atg7) reduced the ability of miR-155 to promote autophagy and mycobacterial elimination. More importantly, our study demonstrated that miR-155 bound to the 3′-untranslated region of Ras homologue enriched in brain (Rheb), a negative regulator of autophagy, accelerated the process of autophagy and sequential killing of intracellular mycobacteria by suppressing Rheb expression. Our results reveal a novel role of miR-155 in regulating autophagy-mediated mycobacterial elimination by targeting Rheb, and provide potential targets for clinical treatment.
| microRNA-155 (miR-155) plays an essential role in regulating the host immune response by post-transcriptionally repressing the expression of target genes. However, little is known regarding its activity in modulating autophagy, an important host defense mechanism against intracellular bacterial infection. Mycobacterium tuberculosis is a hard-to-eradicate intracellular pathogen that infects approximately one-third of the global population, and causes 1.5 million deaths annually. The present study explores a novel role of miR-155 in the host response against mycobacterial infection. Our data demonstrates that mycobacterial infection triggers the expression of miR-155, and the induction of miR-155 in turn activates autophagy by targeting Rheb, a negative regulator of autophagy. miR-155-promoted autophagy accelerates the maturation of the mycobacterial phagosome, thus decreasing the survival of intracellular mycobacteria in macrophages. These findings contribute to a better understanding of the host defense mechanisms against mycobacterial infection, providing useful information for development of potential therapeutic interventions against tuberculosis.
| Mycobacterium tuberculosis (M. tuberculosis) is a hard-to-eradicate intracellular pathogen [1] that infects approximately one-third of the global population, and causes 1.5 million deaths annually [2]. However, only 10% of latent infections lead to active tuberculosis, indicating the importance of host immune defense against mycobacterial infection [2]. The anti-mycobacterial immune system is mainly dependent on cellular immunity mediated by macrophages and T lymphocytes [2]. Macrophages belong to the first line of anti-mycobacterial immune defense and recognize the invading mycobacteria by virtue of various pattern recognition receptors (PRRs) [3]. Macrophages also function to secrete inflammatory cytokines and to present bacterial peptide to T lymphocytes, thus leading to a rapid activation of the adaptive immune response [3]. Although a host can deploy a multitude of immune defense mechanisms against M. tuberculosis, the bacteria is capable of surviving and persisting within host macrophages because of its repertoire of evading the host immune response [4]. For instance, M. tuberculosis can limit the acidification and maturation of mycobacterial phagosomes to escape degradation by lysosomal hydrolases, preventing subsequent antigen presentation [5], [6]. In turn, the host also evolves unique ways to combat intracellular pathogens, such as initiating autophagy to reverse the mycobacteria-induced inhibition of phagosome maturation [7], [8].
Autophagy is an evolutionarily conserved process which is involved in maintaining cytoplasmic homeostasis by degrading damaged organelles or misfolded proteins [9], [10]. The autophagic cascade is initiated by the engulfment of cytoplasmic cargoes by an autophagosome, which then fuses with a late endosome to form the autolysosome, exposing the inner compartment to lysosomal hydrolases for degradation [11]. Several markers of autophagy are well characterized, such as autophagy-related gene 7 (Atg7) and microtubule-associated protein light chain 3 (LC3). Atg7 acts as an E1-activating enzyme taking part in membrane elongation, while LC3 undergoes conversion of LC3-I into its lipidated form LC3-II and is specifically located on the autophagosome until being degraded during autolysosome maturation [12]. Accumulating evidence demonstrates that autophagy is a crucial defense mechanism against a variety of intracellular pathogens, including Shigella flexneri [13], Salmonella typhimurium [14], Listeria monocytogenes [15] and M. tuberculosis [7]. Induction of autophagy overcomes the trafficking block imposed by M. tuberculosis, thus increasing the acidification and maturation of mycobacterial phagosomes, and inhibiting mycobacterial survival in macrophages [7].
Previous studies have demonstrated that autophagy can be activated by physiological signals (e.g., starvation), pharmacological agonists (e.g., rapamycin), or immunological stimuli, such as TLR ligands and cytokines (e.g., IFN-γ and TNF) [12]. The mammalian target of rapamycin (mTOR) plays a key role in inhibiting autophagy [16], and its activity is enhanced by Ras homologue enriched in brain (Rheb) [17]. Activation of autophagy is essential for initiating a protective response and adapting cells to metabolic stresses and immunological challenges [18]. However, excessive autophagic activation, derailed autophagic trafficking or imbalanced degradation may produce pathogenic conditions, leading to cellular toxicity and death [19]. Therefore, autophagy must be tightly regulated. Although signaling pathways governing autophagy remain to be fully delineated, recent evidence suggests that several microRNAs (miRNAs) may participate in modulating autophagy by directly targeting autophagy-related genes, such as Beclin1 and Atg4 [20].
As a class of small non-coding RNAs, miRNAs are highly conserved between different eukaryotic species and function as key regulators of gene expression at the post-transcriptional level by targeting mRNAs for translational repression or degradation [21]. It is reported that several miRNAs fine-tune the innate and adaptive immune responses to mycobacterial infection. For example, miR-125b decreases TNF production by directly targeting its 3′UTR, whereas miR-155 enhances TNF production by increasing TNF mRNA half-life [22]. miR-144* expression is elevated in active TB patients and functions to inhibit TNF and IFN-γ production and T cell proliferation [23]. miR-29 suppresses IFN-γ production in natural killer cells, and CD4+ and CD8+ T cells by directly targeting IFN-γ mRNA, thus contributing to the host susceptibility to mycobacterial infection [24]. However, the role of miRNAs in autophagy-mediated intracellular bacterial elimination remains unclear.
In the present study, we investigated the potential role of miR-155 in modulating autophagy and bacterial clearance in macrophages. Our study demonstrated that miR-155 expression was significantly induced after mycobacterial infection in vivo and in vitro. Overexpression of miR-155 promoted autophagy and the maturation of mycobacterial phagosomes in macrophages, thus facilitating the elimination of intracellular mycobacteria. More importantly, we identified Rheb as a novel functional target of miR-155 in eliminating intracellular mycobacteria. These findings provide a better understanding of host defense mechanisms in mycobacterial infection.
To determine the expression level of miR-155 in vivo in response to mycobacterial infection, miR-155 expression was tested in the lungs of M. tuberculosis H37Rv (H37Rv) infected-BALB/c mice. miR-155 expression in the lungs of H37Rv-infected mice was approximately 2.5-fold higher than in normal uninfected animals (Fig. 1A, p<0.001). We further examined mycobacteria-induced miR-155 induction in cultured murine bone marrow-derived macrophages (BMDMs) and macrophage-like RAW264.7 cells. miR-155 expression was increased in M. bovis BCG (BCG)-challenged murine BMDMs in a time-dependent manner (Fig. 1B). Moreover, in RAW264.7 cells, miR-155 expression was gradually increased by BCG and M. tuberculosis H37Ra (H37Ra) infection in a time- and dose-dependent manner (Fig. 1, C–F).
To determine the role of miR-155 during mycobacterial infection, we next examined its effects on mycobacterial survival by colony-forming unit (CFU) assay. RAW264.7 cells were transiently transfected with miR-155 mimic or inhibitor, and then challenged with BCG, H37Ra or H37Rv at an MOI of 10. Our results showed that miR-155 significantly reduced the survival of intracellular BCG in RAW264.7 cells at 1 h, 6 h, 1 d, 2 d and 3 d postinfection (Fig. 2A, all p<0.05). Moreover, the survival of intracellular H37Ra (Fig. 2B) and H37Rv (Fig. 2C) in RAW264.7 cells transfected with miR-155 mimic were also decreased at the indicated timepoints postinfection (H37Ra: p<0.01 at 3 d postinfection, and p<0.05 at other timepoints; H37Rv: p<0.05 at 1 h postinfection, and p<0.01 at other timepoints).
Additionally, transfection with miR-155 inhibitor increased bacterial survival of BCG in RAW264.7 cells at 1 h, 6 h, 1 d, 2 d and 3 d postinfection (Fig. 2D, p<0.05 at 1 h and 2 d postinfection, and p<0.01 at other timepoints). Similarly, the survival rate of H37Ra (Fig. 2E) and H37Rv (Fig. 2F) also was enhanced in miR-155 inhibitor-treated RAW264.7 cells at the above indicated timepoints postinfection (H37Ra: p<0.01 at 3 d postinfection, and p<0.05 at other timepoints; H37Rv: p<0.05 at 1 d postinfection, and p<0.01 at other timepoints). To test whether miR-155 affects mycobacterial phagocytosis in macrophages, we transfected RAW264.7 cells with miR-155 mimic or inhibitor, and then challenged with fluorescein Texas red-labeled BCG at an MOI of 10 for 1 h. Flow cytometry was performed to measure mycobacterial phagocytosis in macrophages. Our results showed that neither overexpression nor inhibition of miR-155 had an effect on macrophage-mediated mycobacterial phagocytosis (Fig. 2, G and H).
We next explored the mechanisms by which miR-155 promotes the intracellular mycobacterial elimination in macrophages. To this end, we examined the role of miR-155 on the maturation of mycobacterial phagosomes, which is often blocked by mycobacteria to allow their escape from lysosomal bactericidal mechanisms [25]. RAW264.7 cells were challenged with fluorescein Texas red-labeled BCG, and lysosomes were monitored by using immunostaining for lysosomal marker CD63 or staining with DQ Green BSA (DQ-Green, a fluorogenic substrate for proteases). Our results showed that miR-155 enhanced the fusion of BCG phagosomes and lysosomes by nearly two fold, as calculated by the percentage of co-localization of BCG with CD63- or DQ-Green-positive lysosomes (Fig. 3, A–D, both p<0.01). Rapamycin, an inducer of autophagy, was used as a positive control, and treatment increased the percentage of co-localization of BCG with lysosomes by approximately two fold (Fig. 3, A–D, both p<0.01). Taken together, these findings suggest that miR-155 inhibits the survival of intracellular mycobacteria by promoting the maturation of mycobacterial phagosomes.
It has been reported that the induction of autophagy promotes the maturation of mycobacterial phagosomes and the elimination of mycobacteria within infected macrophages [7]. To test the hypothesis that miR-155 induces autophagy in macrophages to enhance the maturation of mycobacterial phagosomes, we evaluated the autophagic activity in RAW264.7 cells using Western-blot and fluorescence microscopy to test the processing of LC3 (conversion from LC3-I to LC3-II) and the number of LC3 puncta, respectively. Real-time PCR data showed that transfection with miR-155 mimic significantly increased the expression level of miR-155 (Fig. 4A), whereas transfection with miR-155 inhibitor markedly decreased miR-155 expression in RAW264.7 cells (Fig. 4B), and with or without BCG challenge. Western-blot results showed that transient transfection with miR-155 mimic activated autophagy in normal uninfected and BCG-infected RAW264.7 cells, as suggested by the increased the amount of LC3-II (Fig. 4C), while transfection with miR-155 inhibitor reduced the amount of LC3-II in RAW264.7 cells before and after BCG challenge (Fig. 4D).
The dynamics of LC3-II accumulation during the process of autophagy depends on both the conversion rate of from LC3-I to LC3-II, and the degradation rate of LC3-II by autolysosomes [12]. To further confirm the role of miR-155 on macrophage autophagy activity, we employed bafilomycin A1, an antagonist of vacuolar H+ ATPase, to prevent luminal acidification and autophagosomal cargo degradation. Our results showed that in the presence of bafilomycin A1, transfection with miR-155 mimic in RAW264.7 cells increased the amount of LC3-II (Fig. 4E), while inhibition of miR-155 decreased the amount of LC3-II (Fig. 4F). Furthermore, autophagic activity was examined in RAW264.7 cells stably expressing GFP-LC3, in which the punctate form (type II) of the autophagy marker LC3 can be directly viewed by confocal microscopy. Overexpression of miR-155 led to the redistribution of GFP-LC3 from diffuse to punctate pattern (Fig. 5, A and B, p<0.01). Additionally, the effect of miR-155 on the distribution of endogeneous LC3 was detected by immunofluorescent staining. Results showed that miR-155 promoted the formation of LC3 puncta (Fig. 5, C and D, p<0.01), which supports our Western-blot data. In addition, we confirmed the modulation of autophagy by miR-155 in normal uninfected macrophages by using an autophagic organelle-specific fluorescent dye monodansylcadaverine (MDC). The results of confocal microscopy indicated that the number of MDC-positive autophagic vacuoles was significantly increased in RAW264.7 cells after transient transfection with miR-155 mimic (Fig. 5, E and F, p<0.01). As a potent autophagy inducer, rapamycin markedly augmented the number of LC3 puncta and MDC-positive autophagic vacuoles (Fig. 5, A–F). These data together demonstrate that miR-155 elevates the autophagic response in macrophages.
Next, we analyzed the formation of autophagosomes containing BCG in macrophages using confocal microscopy. Our results showed that overexpression of miR-155 enhanced the co-localization of BCG with GFP-LC3-positive autophagosomes in a stable GFP-LC3-expressed RAW264.7 cell line (Fig. 6, A and B, p<0.01). miR-155 overexpression also elevated the co-localization of BCG with endogeneous LC3 autophagosomes (Fig. 6, C and D, p<0.01). Similarly, transfection with miR-155 elevated the co-localization of BCG with MDC-positive autophagic vacuoles in RAW264.7 cells (Fig. 6, E and F, p<0.01). Induction of autophagy with rapamycin markedly increased the co-localization of BCG with LC3 puncta (Fig. 6, A–D) or MDC-positive autophagic vacuoles (Fig. 6, E and F, p<0.01).
To determine whether miR-155 enhances the elimination of intracellular mycobacteria via autophagy, we blocked autophagy by using 3-methyladenine (3-MA). Transfection with miR-155 mimic significantly enhanced autophagy and reduced the viability of BCG in RAW264.7 cells, and these effects were partly reversed by treatment with the autophagy inhibitor 3-MA, as indicated by Western-blot and CFU data (Fig. 7, A and B). RAW264.7 cells also were transfected with specific siRNAs against Atg7, to block the autophagic response. Transfection with Atg7 siRNA dramatically decreased protein expression levels of Atg7 and the amount of LC3-II, indicating the efficacy of Atg7 knockdown and autophagy inhibition (Fig. 7C). More importantly, the miR-155-induced autophagy was markedly inhibited in the RAW264.7 cells co-transfected with Atg7- vs control-siRNA (Fig. 7C). And silencing of Atg7 attenuated the miR-155-mediated mycobactericidal activity, when compared to the control treatment (Fig. 7D). Collectively, these results demonstrate that miR-155 promotes the elimination of intracellular mycobacteria by activating autophagy in macrophages.
To identify the specific target of miR-155 that modulates autophagy, bioinformatics analysis was performed with TargetScan (http://www.targetscan.org/). We found that Rheb, which inhibits autophagy via mTOR, displayed a potential seed match for miR-155 in its 3′-untranslated region (3′UTR) (Fig. 8A). To elucidate whether miR-155 represses Rheb by directly interacting with its 3′UTR, we generated psiCHECK-2-REPORT luciferase constructs containing the 3′UTR of Rheb with the putative miR-155 binding site (WT-Rheb), and used psiCHECK-2-REPORT luciferase constructs containing the 3′UTR of Rheb with a mutation at the putative miR-155 binding site (mut-Rheb, UUA to AAU) as controls (Fig. 8A). RAW264.7 cells were co-transfected with control or miR-155 mimic together with these reporter constructs, followed by assessment of luciferase activity at 24 h after transfection. Overexpression of miR-155 repressed the expression of luciferase fused to the WT Rheb 3′UTR (p<0.001), but failed to repress the expression of luciferase fused to the Rheb 3′UTR containing a mutated miR-155 seed sequence (Fig. 8B).
To explore whether miR-155 represses endogenous Rheb, RAW264.7 cells were transfected with control or miR-155 mimic, and protein levels of Rheb were measured by Western-blot. Rheb protein levels were dramatically decreased in RAW264.7 cells transfected with miR-155- vs control-mimic (Fig. 8C), and increased in RAW264.7 cells after transfection with miR-155- vs control-inhibitor (Fig. 8D). However, in both gain- and loss-of-function studies, miR-155 had no effect on the levels of Rheb mRNA expression (Fig. 8, E and F). Altogether, these results indicate that miR-155 post-transcriptionally represses the expression of Rheb by directly interacting with its 3′UTR seed region.
The next series of experiments were designed to explore whether miR-155 promotes autophagy-induced bacterial elimination by targeting Rheb. We first tested the effect of Rheb on autophagy in RAW264.7 cells. Transient transfection with a plasmid expressing Rheb effectively increased protein levels of Rheb and downregulated the amount of LC3-II (Fig. 9A). Furthermore, blocking autophagy flux with bafilomycin A1 effectively inhibited the degradation of LC3-II, and in the presence of bafilomycin A1, overexpression of Rheb also decreased the amount of LC3-II, when compared with treatment with control plasmid (Fig. 9A), indicating that Rheb overexpression dramatically inhibits autophagy in macrophages. To further determine whether miR-155 enhances the elimination of intracellular mycobacteria by targeting Rheb, RAW264.7 cells were co-transfected with miR-155 mimic and Rheb-expressing plasmid, and then confocal microscopy and CFU assay were performed to detect the maturation of mycobacterial autophagosomes and intracellular bacterial load, respectively. Confocal microscopy results showed that overexpression of miR-155 enhanced the co-localization of mycobacterial phagosomes and MDC-positive autophagic vacuoles in RAW264.7 cells after co-transfection with a control vector (Fig. 9, B and C, p<0.05), suggesting that miR-155 promotes the formation of mycobacterial autophagosomes. However, miR-155-mediated formation of mycobacterial autophagosome in RAW264.7 cells was abrogated by co-transfection with Rheb-expressing plasmid (Fig. 9, B and C, p<0.05). Furthermore, CFU assay results indicated that overexpression of miR-155 significantly decreased the bacterial load of intracellular BCG by 25% and 40% at 1 h and 6 h postinfection, respectively (Fig. 9D, both p<0.01), whereas Rheb overexpression significantly inhibited miR-155-mediated mycobactericidal activity in RAW264.7 cells (Fig. 9D, p<0.05). Together, these results suggest that miR-155 promotes autophagy-mediated elimination of intracellular mycobacteria by targeting Rheb.
Autophagy has been demonstrated to play an essential role in the host immune response against mycobacterial infection. Nonetheless, the molecular basis involved in autophagy-mediated mycobacterial clearance remains largely unclear. Here we report a novel role of miR-155 in regulating autophagy and mycobacterial elimination in macrophages by targeting Rheb, which may provide a better understanding of the host anti-mycobacterial response.
Emerging evidence has shown that miR-155 is mainly expressed in activated macrophages [26], dendritic cells [27], B [28] and T lymphocytes [29]. miR-155 expression is up-regulated by a variety of inflammatory mediators and pathogens [26], [30]. Our study shows that miR-155 expression is enhanced both in vivo and in vitro after mycobacterial infection, which is consistent with previous study showing the miR-155 induction in response to avirulent M. smegmatis [22] and M. bovis BCG challenge [31]. Recent studies have revealed that M. tuberculosis purified protein derivative (PPD) induces a high expression of miR-155 in peripheral blood mononuclear cells (PBMCs) from active TB patients vs healthy donors [32]. These data implicate a potential correlation of miR-155 with mycobacterial infection.
Studies have revealed that miR-155 participates in various biological processes including oncogenicity, haemotopoiesis and inflammatory responses [33]. In vivo studies using miR-155 deficient mice have demonstrated that miR-155 is required for normal immune function of T and B lymphocytes as well as dendritic cells [34]. It is also reported that miR-155 promotes the development of T helper 1 (Th1) and Th17 cell subsets [35], and attenuates the Th2 cell response by targeting c-Maf, a potent transactivator of the IL-4 promoter [36]. Moreover, miR-155 contributes to the classical activation (M1 polarization) of macrophages and iNOS expression by targeting CCAAT/enhancer binding protein β (CEBPβ), a hallmark of alternatively activated (M2) macrophage [37]. These studies together implicate a potent role of miR-155 in the cellular immune response, which has been demonstrated as the major arm of host anti-mycobacterial defense.
As one of the most important cell types in the anti-mycobacterial immunity, macrophages function as the predominant responder cell against M. tuberculosis infection. The cells uptake and kill mycobacteria and by initiating an inflammatory response, but also provide a preferred hiding and replication site for mycobacteria [38]. Therefore, precisely modulating of macrophage activity is crucial in the protective immunity against mycobacterial infection.
The exact role of miR-155 in macrophages during mycobacterial infection has yet to be elucidated. Studies have demonstrated that miR-155 enhances TNF production in human macrophages in response to mycobacteria [22], and TNF in turn elevates macrophage activity to eliminate intracellular mycobacteria [39], [40], indicating that miR-155 might be required in the macrophage-mediated immune defense against infection. In the present study, RAW264.7 cells were infected with mycobacteria at an MOI of 10 for 1 h, and then incubated for another 1 h, 6 h, 1 d, 2 d, and 3 d, after removing the extracellular bacteria. Our results demonstrate that miR-155 promoted the maturation of mycobacterial phagosome and decreased the survival of intracellular BCG at all tested timepoints, indicating that miR-155 promoted macrophage-mediated bacterial elimination. However, it is reported that in some cases, miR-155 may display a different role in modulating mycobacterial survival. For example, Ghorpade et al. reported that overexpression of miR-155 in RAW264.7 cells promoted apoptosis and increased intracellular bacterial load at 96 h postinfection (including a 24 h infection at an MOI of 10 plus a 72 h postinfection incubation) [31]. Ghorpade's observation is inconsistent with others showing that induction of apoptosis in infected host cells reduces the viability of intracellular mycobacteria [41], but could be explained by Lee's findings showing that high numbers of intracellular mycobacteria trigger macrophage necrosis that could promote mycobacterial survival [42]. In Ghorpade's study, the percentage of viable cells was markedly decreased to 10% after transfection with miR-155 overexpressing plasmid plus BCG infection, thus greatly increasing the actual MOI during infection.
Autophagy has been proved to be a crucial element of the innate immune response against intracellular pathogens, including mycobacteria [43]. In the present study, we demonstrate that miR-155 induces the processing of LC3 and the accumulation of LC3 puncta in both BCG-challenged and unchallenged RAW264.7 cells, which indicates that miR-155 accelerates the autophagic response in macrophages and reduces intracellular bacterial load. Moreover, we observed that inhibition of autophagy with 3-MA or silencing Atg7 reduced miR-155-mediated autophagy and bacterial killing. In addition, we found that in control mimic treated RAW264.7 cells, treatment with 3-MA dramatically increased the survival of BCG at 6 h, though the reduction of the level of LC3-II is modest. This may be explained by other reports showing that 3-MA promotes the survival of mycobacteria by suppressing nitric oxide production [44].
Autophagy not only elevates the delivery of mycobacterial phagosomes for mycobacterial killing [7], but also enhances the presentation of mycobacterial antigens (e.g., Ag85B) to induce a protective CD4+ T lymphocyte response [45]. Both in vivo and in vitro studies have shown that induction of autophagy effectively increases the intracellular killing of M. tuberculosis [46], suggesting that targeting autophagy may be a potential therapeutic strategy for TB treatment. Interestingly, our results showed that inhibition of autophagy by 3-MA or Atg7 siRNA attenuated, but not fully blocked the miR-155-mediated mycobactericidal activity at 6 h postinfection (Fig. 7, B and D), indicating that in addition to autophagy, other alternative mechanisms may be involved in the miR-155-mediated bacterial elimination. Studies have demonstrated that macrophages also employ an oxygen-dependent system to control intracellular pathogens [47], [48], [49]. Our unpublished data showed that overexpression of miR-155 did not influence iNOS expression or nitric oxide production, but slightly enhanced the production of reactive oxygen species in BCG-infected RAW264.7 cells (data not shown), suggesting that miR-155 may exert antimycobacterial function partially through enhancing ROS production.
Studies have demonstrated that miR-155 implements different functions in various physical, pathological and experimental conditions by repressing distinct targets, including, but not limited to, TGF-β activated kinase 1/MAP3K7 binding protein 2 (TAB2) [27], suppressor of cytokine signaling 1 (SOCS1) [50], PU.1 [51], CEBPβ [52], Src homology-2 domain-containing inositol 5-phosphatase 1 (SHIP-1) [53], myeloid differentiation primary response gene 88 (MyD88) [54], phosphoinositide 3-kinase (PI3K) p85 [55]. Most of these miR-155 targets are involved in modulating inflammatory response. Among these targets of miR-155, PI3K p85 plays an important role in regulating autophagy. However, our unpublished data from luciferase assay and Western blot demonstrated that miR-155 did not bind to the 3′UTR of p85, and had no influence on p85 expression in RAW264.7 cells (data not shown), which differs from other reports showing that miR-155 targets p85 in human B-cell lymphoma cells [55]. Studies have demonstrated that different targets may be involved in different systems [33], [56]. For example, miR-125b is reported to promote hepatocellular carcinoma cell apoptosis by targeting Bcl-2, an anti-apoptotic protein [57], while miR-125b inhibits apoptosis of Hela cells and human immortalized myelogenous leukemia K562 cells by targeting pro-apoptotic proteins, Bak1, Mcl1 and p53 [58].
In the present study, we identified that Rheb is a novel target for miR-155. Rheb functions as a negative regulator of autophagy, by directly interacting with mTOR [17] and increasing the mTOR activity [59]. Rheb activity is regulated by upstream kinases, such as Akt, AMP-activated protein kinase, and glycogen synthase kinase-3β, while little is known regarding the mechanism involved in modulating Rheb expression. It is reported that Rheb expression is increased in hepatocytes after Hepatitis C virus infection [60], but reduced in RAW264.7 cells after treatment with hydrogen peroxide [61]. Our study indicates that miR-155 post-transcriptionally down-regulates the expression level of Rheb, thus activating autophagy in macrophages. Moreover, CFU results suggest that overexpression of Rheb reduces the ability of miR-155 to induce intracellular bacterial clearance, indicating that Rheb is the major target of miR-155 in modulating the autophagic response and mycobacterial elimination in macrophages.
A recent study showed that Shigella infection results in sustained amino acid starvation and mTOR inhibition, thereby triggering the autophagy response in host cells for innate host defense, whereas Salmonella only induces a transcient amino acid starvation, favoring mTOR reactivation and leading to autophagy escape [62]. These findings reflect the intimate link between host metabolism and autophagic response against invading pathogens. Our study demonstrates that mycobacteria-induced miR-155 targets Rheb, a positive regulator of mTOR signaling, which is crucial in sensing and responding to cellular nutrients and stress [63], suggesting that miR-155 may have a potential role in modulating the metabolic signaling in host cells.
Collectively, the present study demonstrates that the miR-155 is induced by mycobacterial infection, and promotes autophagy in macrophages by targeting Rheb, conferring protection against infection with intracellular mycobacteria. Our study unravels an important role of miR-155 in autophagy regulation and mycobacterial elimination, which may provide useful information for developing potential therapeutic interventions against tuberculosis.
All animal experiments were performed in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals, and the experimental procedures were approved by the Medical Ethics Committee of
Sun Yat-sen University Zhongshan School of Medicine and the Biosafety Management Committee of Sun Yat-sen University (No. 2012-33).
Eight-week-old BALB/c and C57BL/6 (B6) mice were purchased from Sun Yet-sen University Animal Supply Center. Middlebrook 7H9 broth medium and Middlebrook 7H10 agar were purchased from BD Difco Laboratories (Sparks, MD). The CD63 antibody was obtained from Santa Cruz Biotechnology (Santa Cruz, CA). The LC3 antibody was purchased from Novus Biologicals (Littleton, CO). Antibodies against Atg7 and Rheb were obtained from Cell Signaling Technology (Beverly, MA). The β-actin antibody, MDC, bafilomycin A1, 3-methyladenine, rapamycin and DMSO vehicle control (0.2%) were obtained from Sigma-Aldrich (St. Louis, MO). The Texas Red and DQ-Green dyes were purchased from Invitrogen (Carlsbad, CA).
Murine macrophage-like RAW264.7 cells (ATCC; TIB-71) were cultured in DMEM supplemented with 10% fetal bovine serum (FBS) and 100 U/ml penicillin, 100 µg/ml streptomycin (GIBCO, Invitrogen). BMDMs were prepared by culturing bone marrow from the femurs and tibiae of 6- to 8-week old B6 mice in DMEM containing 10% FBS, 2 mM L-glutamine, 1 mM sodium pyruvate, 100 U/ml penicillin, 100 µg/ml streptomycin, and 10% L929 conditioned medium. Non-adherent cells were removed after 24 h and cultured for 7 days. M. bovis BCG strain 19015, M. tuberculosis H37Ra strain 25177 and M. tuberculosis H37Rv strain 25618 were purchased from the American Type Culture Collection (ATCC), and mycobacteria were grown in Middlebrook 7H9 broth medium or on 7H10 agar plates supplemented with OADC and cultured in a standard tissue culture incubator at 37°C with an atmosphere of 5% CO2 and 95% air.
M. tuberculosis H37Rv was homogenized to generate a single cell suspension. BALB/c mice were intraperitoneally injected with M. tuberculosis H37Rv (5×106 CFU/per mouse). After 6 weeks, lungs were removed from sacrificed mice. Total RNA was isolated using TRIzol reagent (Invitrogen) according to the manufacturer's recommendations. Animal experiments were performed in accordance with the approval of the Scientific Investigation Board of Sun Yat-sen University (Guangdong, China).
pEX-GFP-LC3 was from Addgene (#24987), which was deposited by Isei Tanida [64]. The 3′UTR and cDNA sequence of mouse Rheb was amplified by reverse transcription-PCR and cloned into psiCHECK-2 (Promega, Madison, WI) and pMSCV-neo (Clontech, Mountain View, CA), respectively. The mutation in the 3′UTR of mouse Rheb was generated with the QuikChange Site-Directed Mutagenesis Kit from Stratagene (La Jolla, CA) following the manufacturer's protocol. Negative control (NC) siRNA was purchased from Invitrogen. Atg7 siRNA was purchased from Dharmacon/Thermo Fisher Scientific (Waltham, MA).
RAW264.7 cells (at approximately 50% confluence) were transiently transfected with 30 nM control or miR-155 mimic (Applied Biosystems, Foster City, CA); 50 nM control or miR-155 LNA-inhibitor (EXIQON, Vedbaek Denmark); or 1.6 µg plasmid; or 40 pmol siRNA, using Lipofectamine 2000 (Invitrogen) according to the manufacturer's instructions.
RAW264.7 cells were infected with M. bovis BCG or M. tuberculosis H37Ra or H37Rv at an MOI of 10. After 1 h incubation at 37°C, the infected cells were washed extensively with PBS to remove extracellular mycobacteria, and the infected cells were incubated for another 1 h, 6 h, 1 d, 2 d, or 3 d, and then lysed in 1 ml of distilled water. Quantitative culturing was performed using 10-fold serial dilutions. Aliquots of each dilution were inoculated in triplicate on Middlebrook 7H10 agar plates with OADC. Plates were incubated for 3 weeks, and colonies were counted.
Phagocytosis was assayed by flow cytometry, as described by others [65]. Briefly, BCG was incubated with Texas Red (Invitrogen) at room temperature for 2 h, protected from light, and then gently rinsed with PBS. Then RAW264.7 cells were challenged with Texas Red -labeled BCG at an MOI of 10. After 1 h incubation, cells were washed three times with cold PBS and centrifuged to remove extracellular bacteria, and then analyzed by flow cytometry using a Beckman Coulter EPICS XL/MCL (Beckman Coulter Inc., Fullerton, CA) instrument.
RAW264.7 cells (at approximately 50% confluence) were cultured in 24-well plates one day prior to transfection. psiCHECK-2 luciferase reporter plasmids (Promega, Madison, WI) containing either a wild-type or mutated Rheb 3′UTR were co-transfected with control or miR-155 mimic into RAW264.7 cells with Lipofectamine 2000 (Invitrogen). Cells were harvested 24 h later, and luciferase activity was assessed with the Dual-Luciferase Reporter Assay System (Promega) following the manufacturer's protocol.
RAW264.7 cells were transfected with pEX-GFP-LC3 using Lipofectamine 2000 according to manufacturer's instructions. Stable transfectants were selected for 3 weeks with 1 mg/ml G418 and maintained in 0.2 mg/ml G418. The transfectants were confirmed by Western-blot, and the formation of LC3 puncta in response to rapamycin was assessed by fluorescence microscopy (data not shown).
Total RNA was isolated using TRIzol reagent (Invitrogen) according to the manufacturer's recommendations. For miRNA, the expression levels of miR-155 were detected using a TaqMan microRNA kit (Applied Biosystems) and normalized to small nuclear RNA (Rnu6). For mRNA, first-strand cDNA synthesis was performed using RevertAid First Strand cDNA Synthesis Kit (Thermo Fisher Scientific, Waltham, MA). For RT-PCR, the expression of Rheb was assessed by PCR amplification using a standard protocol. Amplified products were fractionated by 1% agarose gel electrophoresis and visualized by ethidium bromide staining. The primer sequences used for PCR were: Rheb, 5′- ATGCCTCAGTCCAAGTCCCGGA AG-3′ (forward) and 5′- TCACATCACCGAGCACGAAGA -3′ (reverse); β-actin, 5′-GATTACTGCT CTGGCTCCTAGC -3′ (forward) and 5′- GACTCATCGTAC TCCTGCTTGC -3′ (reverse).
Cells were washed three times with ice-cold PBS and then lysed in lysis buffer containing 1 mM phenylmethylsulfonyl fluoride, 1% (vol/vol) protease inhibitor cocktail (Sigma), and 1 mM DTT. Equal amounts (20 µg) of cell lysates were resolved by SDS-PAGE and then transferred to PVDF membranes. Membranes were blocked in 5% non-fat dry milk in PBST and incubated overnight with the respective primary antibodies at 4°C. The membranes were incubated at room temperature for 1 h with appropriate HRP-conjugated secondary antibodies and visualized with Plus-ECL (PerkinElmer, Shelton, CA) according to the manufacturer's protocol.
For immunofluorescence experiments, cells were grown on collagen-precoated glass coverslips in 24-well plates. Cells were transiently transfected with control or miR-155 mimic (30 nM) for 24 h and then infected with Texas Red-labeled BCG (MOI 10) for 1 h. Cells were fixed with 4% paraformaldehyde followed by membrane permeabilization using 0.2% Triton X-100. Cells were blocked with 5% BSA and incubated with primary and then secondary antibodies before mounting. In fluorescence experiments, RAW264.7 GFP-LC3 cells were transfected with control or miR-155 mimic using Lipofectamine 2000 and either left uninfected or infected with Texas Red-labeled BCG (MOI 10) for 1 h. MDC staining was performed by adding MDC (50 µM) to cells and incubating at 37°C for 30 min. The cells were fixed in 4% paraformaldehyde for 10 min and viewed by confocal microscopy (Zeiss Axiovert, LSM710).
Unpaired Student's t test or one-way analysis of variation was used to determine the significance of the results from real-time RT-PCR experiments and CFU assays, the quantification of GFP-LC3 puncta, MDC-positive autophagic vacuoles and the co-localization of BCG with lysosomes or autophagosomes. Data were considered statistically significant at p<0.05.
microRNA-155: [ Mus musculus]. NCBI Gene ID: 387173.
Rheb: [Mus musculus]. NCBI Gene ID: 197744.
Atg7: [Mus musculus]. NCBI Gene ID: 74244.
|
10.1371/journal.pgen.1006837 | A conserved role for the ESCRT membrane budding complex in LINE retrotransposition | Long interspersed nuclear element-1s (LINE-1s, or L1s) are an active family of retrotransposable elements that continue to mutate mammalian genomes. Despite the large contribution of L1 to mammalian genome evolution, we do not know where active L1 particles (particles in the process of retrotransposition) are located in the cell, or how they move towards the nucleus, the site of L1 reverse transcription. Using a yeast model of LINE retrotransposition, we identified ESCRT (endosomal sorting complex required for transport) as a critical complex for LINE retrotransposition, and verified that this interaction is conserved for human L1. ESCRT interacts with L1 via a late domain motif, and this interaction facilitates L1 replication. Loss of the L1/ESCRT interaction does not impair RNP formation or enzymatic activity, but leads to loss of retrotransposition and reduced L1 endonuclease activity in the nucleus. This study highlights the importance of the ESCRT complex in the L1 life cycle and suggests an unusual mode for L1 RNP trafficking.
| Long interspersed nuclear elements (LINEs) are a class of retrotransposable elements that mutate mammalian genomes. LINEs have been highly successful in the human genome, multiplying to over 800,000 copies. The LINE-encoded replication machinery is also used by other retrotransposons, and in total, has been responsible for the generation of over 1/3 of human DNA sequence. To replicate, a LINE mRNA forms a ribonucleoprotein particle (RNP) with its proteins. This RNP eventually enters the nucleus to integrate a cDNA copy of itself into chromosomes. The events between RNP formation and successful integration are difficult to study and largely unknown. Here we show that the ESCRT complex plays a conserved role in LINE retrotransposition in both yeast and humans. ESCRT is a membrane budding complex involved in cellular trafficking and membrane budding/fusion. Our results imply that membranes play an integral part of LINE replication, and ESCRT may be required for RNP trafficking towards the nucleus.
| Long Interspersed Nuclear Elements (LINEs) are an ancient class of non-long terminal repeat (non-LTR) retrotransposable elements widely dispersed among eukaryotes. These elements can be categorized into distinct clades based on homology of conserved domains [1]. The L1 clade is of particular interest because its namesake element, L1, is widespread throughout mammalian genomes. L1 has been enormously successful in populating the human genome, comprising at least 17% of human DNA [2]. In addition, other human retrotransposons such as Alu and SVA depend on the L1 machinery to replicate [3–5]. When the sequences of these L1 “parasites” are taken into account, greater than 30% of the human genome has been produced by the L1 retrotransposition machinery [2].
The structure of a typical active, full-length L1 is shown in Fig 1A. L1 has two open reading frames (ORFs) which encode two proteins called ORF1p and ORF2p. ORF1p is a homotrimeric non-sequence specific RNA binding protein that is thought to play a structural role in the L1 ribonucleoprotein particle (RNP) formation [6–11]. ORF1p also has nucleic acid chaperone activity in vitro [12]. The amino acids required for the chaperone activity are required for L1 retrotransposition [13], but how they contribute to L1 replication remains unclear. ORF2p encodes endonuclease and reverse transcriptase activity, both of which are important for retrotransposition [14–16].
To replicate, the bicistronic L1 mRNA is transcribed, translated, and complexes with L1-encoded proteins (ORF1p and ORF2p) to form a cytoplasmic L1 RNP [17,18] (Fig 1B). A cytoplasmic L1 RNP is presumed to gain access to the nucleus by an unknown mechanism. Once in the nucleus, ORF2p-encoded endonuclease activity nicks a chromosome, and ORF2p reverse transcriptase uses this nick as a primer to copy L1 mRNA at the new chromosomal site [19,20]. Together, these enzymatic events are referred to as target-primed reverse transcription (TPRT). How the TPRT complex is resolved has yet to be established. This entire process of L1 replication (transcription through integration) is called L1 retrotransposition or “L1 activity”. Sequencing efforts have shown that up to 30% of genomic structural variation between human individuals is due to L1 retrotransposition [21,22], and L1 retrotransposition is also a source of DNA damage [23,24] and novel disease alleles [25,26]. Thus, at a minimum, L1s are endogenous mutagens that have played a major role in the evolution of our genome. Recent studies have shown that the relationship between L1s and their mammalian host cells is even more complex, and that L1s may play a major unrecognized role in health and development. L1 overexpression is associated with infertility, cancer, neurological disorders, and aging [24,27–39], although whether L1 plays a biologically significant role in these processes is inconclusive.
The streamlined simplicity of an L1 element belies the complexity of LINE retrotransposition, suggesting that LINE RNPs utilize host factors to assist in their replication. Multiple proteomic studies and candidate based approaches have identified host factors that interact with and restrict L1 activity, such as the piwi-interacting RNA machinery [24,27,28], the APOBEC3 family [40–45], hnRNPL [46], and the antiviral factors MOV10 and ZAP [47–50]. However, the list of positive host factors (factors that facilitate L1 retrotransposition) with demonstrated specific and direct interaction with the L1 proteins is remarkably short. Currently we are aware of only one such positive factor that binds directly to L1 proteins. PCNA binds to ORF2p and is hypothesized to be recruited in the nucleus during or after L1 cDNA synthesis [49]. Another positive factor, the poly(A) binding protein PABPC1, binds the L1 RNP via RNA [51].
To facilitate the identification of positive LINE host factors, we previously developed a budding yeast model of LINE retrotransposition [52]. In this model, Zorro3, an L1-clade member from Candida albicans (Fig 1A), was redesigned for the ability to retrotranspose in Saccharomyces cerevisiae. We found that Zorro3 retrotransposes in a manner mechanistically similar to mammalian L1, and therefore we hypothesized that the S. cerevisiae genome encodes host factors required for LINE retrotransposition. In the current study, we describe a genetic screen using the Zorro3 model to identify host factors for LINE retrotransposition. We report the identification of the endosomal sorting complex required for transport (ESCRT) as a critical factor for successful LINE retrotransposition. ESCRT physically interacts with Zorro3, and the importance of ESCRT for retrotransposition is retained with human L1. Human ORF1p contains a conserved ESCRT-interacting domain which is dispensable for RNP formation and enzymatic function, but important for retrotransposition. Disruption of the ESCRT-L1 interaction alleviated L1 endonuclease-mediated toxicity. Together, our data reveals a surprising potential role for membrane budding in the trafficking of L1 RNPs.
To identify host factors important for LINE retrotransposition, we screened the S. cerevisiae yeast knockout collection [53] for strains unable to support robust Zorro3 retrotransposition. Each strain was transformed with a Zorro3 retrotransposition reporter plasmid (Fig 1C), then assayed for retrotransposition. This assay is based on previously described retrotransposition assays [52,54,55] and depends on the splicing of an artificial intron out of a reporter gene (mMET15AI) placed in the Zorro3 3’ untranslated region. A functional MET15 reporter is reconstituted only after Zorro3 transcription, splicing, and integration. Knockout strains were first qualitatively screened for lack of retrotransposition, then the identified retrotransposition-defective strains were quantitatively assayed for retrotransposition frequency (Fig 1D). Out of 4,819 knockout strains screened, 56 strains had a severe (> = 90%) decrease in retrotransposition as compared to a wild-type control strain (S1 Table). Of these 56 strains, 9 were deleted in a gene encoding an endosome associated protein, and 6 of these proteins belong to the endosomal sorting complex required for transport (ESCRT), a 27-fold enrichment of ESCRT genes as compared to the entire genome. Because our screen only identified 6 out of the 19 ESCRT proteins, we next individually tested deletion strains of all ESCRT genes with the quantitative retrotransposition assay, and found that virtually all ESCRT strains had a defect in retrotransposition (Fig 1E, S2 Table). The two exceptions, IST1Δ and VTA1Δ, have been reported as non-essential for ESCRT function [56,57]. In the early stages of our screen, we also transformed a plasmid expressing the lacZ gene under the control of the GAL1 promoter into selected knockout strains with potentially altered retrotransposition frequency. We assayed β-galactosidase activity in these strains to determine whether the altered retrotransposition frequencies could be explained by effects on the GAL1 promoter (S3 Table). For some knockout strains (e.g. SNO1, NPR1, YML095C-A, CUE1, YNL190W) reduction of β-galactosidase activity tracked closely with retrotransposition activity. In the case of the four ESCRT knockouts examined, β-galactosidase activity was lower, but could not entirely explain the >90% reduction in retrotransposition activity in each case (S3 Table). In sum, our screen suggests that ESCRT is required for efficient LINE retrotransposition in budding yeast.
ESCRT has roles in cell abscission, formation of intraluminal vesicles, nuclear envelope reformation, and viral budding from the plasma membrane [58–65]. The commonality between these topologically equivalent processes is membrane fusion, a process ESCRT facilitates by constricting and sealing the membrane “neck” made in each case (Fig 2A). Because LINE RNPs have not been known to traffic through membranes, and ESCRT has multiple important cellular roles, we wondered whether ESCRT and Zorro3 directly interact, or if the retrotransposition defects present in ESCRT deficient strains were simply indirect genetic effects. To test this, we asked whether ESCRT proteins physically associate with Zorro3 proteins.
We HA-epitope tagged and overexpressed 6 yeast ESCRT proteins (Snf7p, Vps4p, Vps23p, Bro1p, Snf8p, and Vps20p) in the S. cerevisiae strain JHY148 [52], which expresses Zorro3 (Fig 2B). Snf7p, Vps4p, Vps23p, Bro1p, and Snf8p associated with Zorro3 ORF1p in both ORF1p and HA immunoprecipitations (Fig 2B). These associations were not disrupted in the presence of RNase A, suggesting that the interactions were not RNA mediated (S1 Fig). In the strain without a plasmid (JHY148 alone) or with empty vector (pGAL-HA, S1 Fig), HA pulldowns did not co-immunoprecipiate Zorro3 ORF1p, demonstrating that the interactions were dependent on the presence of HA-tagged ESCRT proteins. When the same experiment was performed in JHY146, a Zorro3-free strain isogenic to JHY148, the Zorro3 ORF1p antibody only pulled down Bro1p (Fig 2C), suggesting that while the Bro1p interaction is non-specific, the other ESCRT/ORF1p interactions are Zorro3 specific. Thus, components of the ESCRT complex physically interact with Zorro3.
ESCRT is a conserved complex in eukaryotes, so we next tested the effects of reduced levels of human ESCRT proteins on human L1 retrotransposition. We chose three human genes (ALIX, CHMP6, and CHMP3) orthologous to yeast ESCRT proteins that 1) had a dramatic effect on Zorro3 retrotransposition in yeast and 2) only have one known human ortholog (Fig 3A). We transfected short interfering RNAs (siRNAs) into HeLa cells, which depleted endogenous ALIX, CHMP6 or CHMP3 protein to levels undetectable by western blot (Fig 3B). Using a standard retrotransposition assay [15,66], these ESCRT-depleted cells were then tested for their ability to support human L1 retrotransposition. In parallel, the cells were also transfected with a control plasmid, pcDNA3, containing the neomycin resistance gene, to determine whether depletion of ESCRT genes limited the ability to form G418 resistant colonies. To take into account possible effects of ESCRT knockdown on cell growth and/or viability, retrotransposition induced G418 resistant colonies were normalized to pcDNA3 induced G418 colonies. Knocking down ALIX, CHMP6, or CHMP3 led to a 67%, 94%, and 81% reduction, respectively, in normalized L1 activity (Fig 3C, S4 Table). ESCRT knockdown does not reduce steady state L1 ORF1p levels (S2 Fig), suggesting that the retrotransposition defects are not due to alteration of L1 expression. To ensure that the reduction in L1 activity was not caused by siRNA off target effects, we designed ALIX, CHMP6, and CHMP3 cDNAs with synonomous mutations conferring resistance to their respective siRNA used in Fig 3C. When empty vector was cotransfected in the L1 assay, siRNA knockdown of ALIX, CHMP6, and CHMP3 led to 82%, 95%, and 87% reduction of L1 retrotransposition (Fig 4D, S5 Table), similar to the results in Fig 3C. However, cotransfection of the appropriate siRNA-resistant cDNA in the L1 assay mitigated the effects of ESCRT siRNAs (Fig 3D, S5 Table). In the presence of resistant ALIX cDNA, siALIX reduced relative L1 activity to 54% from 70%, a reduction of only 23%. In the presence of resistant CHMP6 cDNA, siCHMP6 reduced relative L1 activity to 7% from 32%, a reduction of 78%. In the presence of resistant CHMP3 cDNA, siCHMP3 increased relative L1 activity to 69% from 61%. Overexpression of CHMP6 is known to cause cytotoxicity [67], which is consistent with the low levels of G418 colony formation when CHMP6 is overexpressed (Fig 3D, resistant CHMP6 cDNA + scramble siRNA). This makes the results of the CHMP6 rescue ambiguous as compared to the rescue with ALIX and CHMP3 (Fig 3D). In total, our results suggest an important role for ESCRT proteins during human L1 retrotransposition.
Some enveloped viruses encode short amino acid motifs called “late domains”, which recruit ESCRT to the budding virus at the plasma membrane [68]. Without ESCRT, viral buds are not pinched off the membrane, and remain attached to the cell [69]. We recognized that the consensus sequence of a known viral late domain, YPXnL, corresponds to a motif in human L1 ORF1p (Fig 4A). This motif is conserved among many mammalian L1s (Fig 4B), but the biological function of this L1 motif has not been previously assigned. However, introducing multiple alanine substitutions (YPAKLS->AAAALA) in the motif is known to abolish 99.9% of L1 activity [15]. Although this reduction in retrotransposition activity could be explained, in part, to a disrupted ESCRT/ORF1p interaction, the YPAKLS->AAAALA mutation also leads to poor steady state ORF1p expression and severe defects in L1 RNP formation ([9] and Fig 4C). Thus, the reduction in L1 activity could simply be attributed to the instability of ORF1p and/or the inability to form RNPs.
We took advantage of previous YPXnL studies to make a separation of function ORF1p mutant that can still form functional L1 RNPs. The known binding partner of the YPXnL late domain is the human protein ALIX, which serves as an adapter protein to recruit ESCRT [69]. The structures of HIV and EIAV YPXnL motifs bound to ALIX have been solved and revealed critical interactions required for ALIX/ YPXnL binding [70]. The YPXnL tyrosine extends into a hydrophobic pocket in ALIX, and mutating this tyrosine eliminates late domain binding. The corresponding Y282 tyrosine within the L1 ORF1p YPXnL motif is conserved in mammalian ORF1p, and is present on the external surface of the protein in a position that is presumably accessible for binding to ALIX [11]. We mutated this tyrosine to alanine (Y282A) in L1 ORF1p. In contrast to the YPAKLS->AAAALA mutant, the Y282A mutant produces steady state ORF1p and RNPs indistinguishable from wild type (Fig 4C), suggesting that the Y282A mutant does not alter steady-state levels of L1 RNPs. In addition, we examined RNPs for reverse transcriptase activity using the L1 element amplification protocol (LEAP) assay, the best current biochemical assay for functional L1 RNPs [71]. RNPs isolated from wild type or Y282A transfected cells generated similar LEAP products, whether analyzed by gel electrophoresis or direct DNA sequencing (Fig 4D and S3 Fig). We found no evidence of reverse transcriptase priming upstream of the L1 poly(A) tail (S3 Fig). Thus, based on currently available assays, RNPs produced by a Y282A mutant appear to be functionally “normal”. However, the Y282A mutant exhibited a profound defect (~90% reduction) in retrotransposition (Fig 4E, S6 Table). Co-immunoprecipitation of ALIX with L1 ORF1p was dependent on the YPXnL motif, as the Y282A mutant abolished the ORF1/ALIX interaction (Fig 4F). These co-IPs were performed in the presence of RNase A, suggesting that this interaction is not RNA dependent. The strength of the wild type ORF1p/ALIX interaction varied from experiment to experiment (S4 Fig). In sum, our data support the importance of the L1 YPXnL late domain for L1/ESCRT interaction and L1 retrotransposition.
The interaction between ESCRT and L1 proteins strongly suggests that the L1 RNP buds into or interacts with a membrane at some point during L1 replication. Several factors confound our ability to directly pinpoint which membranes, if any, are relevant to L1 biology. L1 retrotransposition is inefficient–even under conditions of massive L1 overexpression over multiple days, typically less than 1% of cells will harbor a retrotransposition event [15,72–75]. In addition, although the L1 mRNA/proteins must enter the nucleus to replicate, most reports suggest that the vast majority of L1 protein is detected in the cytoplasm [18,49,76]. In our hands, we see similar cytoplasmic localization of L1 ORF1p, without obvious co-localization with known cellular organelles (S5 Fig). With the enormous background of L1 particles sequestered in the cytoplasm, it is currently impossible to visually identify the location of the rare “active” RNPs.
We therefore blocked membrane associated functions that we hypothesized could play a role in retrotransposition. ESCRTs are used for budding into endosomes and lysosomes, both of which lead to cargo deposition and degradation in the lysosome. Although it is counterintuitive that lysosome targeting would be required for retrotransposition, the close association of ESCRT with the endosomal membrane system led us to test the requirement of functional lysosomes for L1 activity. We blocked lysosome function using amantadine or chloroquine, drugs that neutralize the acidic lysosome environment [77]. L1 activity was unaffected in amantadine treated cells, and increased in chloroquine treated cells (Fig 5A). Cells treated with amantadine or chloroquine demonstrated the previously described swelling of lysosomes when visualized by confocal microscopy [78–80], indicating that the inhibitors were acting as expected (Fig 5B). This suggests that lysosomal function is not required for L1 retrotransposition. We also reasoned that the L1 RNP might be associated with endoplasmic reticulum (ER) membrane. L1 entry into the ER could be required for glycosylation or transport of the L1 RNP. We treated cells with kifunensine or D-mannojirimycin, inhibitors of N-linked glycosylation [81,82]. These drugs reduced glycosylation and production of the mature, endoglycosidase H resistant form of the normally heavily glycosylated ICAM-1 (Fig 5C), indicating an effective glycosylation block. L1 retrotransposition was not significantly affected by this block (Fig 5A). Finally, we treated cells with Exo1 or golgicide A, potent golgi inhibitors that cause golgi collapse and arrest of soluble and membrane associated protein secretion [83,84]. Neither of these drugs had a dramatic effect on L1 retrotransposition frequency (S5 Fig), suggesting that glycosylation or other pathways downstream of the ER are not essential for L1 replication. However, this does not rule out the possibility that L1 RNPs enter the ER for another purpose, such as formation of a shielded vesicular compartment analogous to viral replication compartments [85], or retrograde L1 RNP transport towards the nucleus, the ultimate destination for successful L1 retrotransposition.
Because of the importance of ESCRT in vesicular traffic, and the assumed requirement of L1 RNPs to traffic to the nucleus to replicate, we wondered whether the ESCRT/L1 interaction is necessary for L1 nuclear entry. In our hands, we detect a very low amount of L1 protein in the nucleus when we induce L1 expression (S6 Fig) in tissue culture cells. The small amount of L1 ORF1 that we detect in the nucleus hinders our ability to identify “active” L1 RNPs from the overwhelming amount of background L1 expression. Thus, we chose to use a functional assay as a proxy for nuclear entry of L1 particles. Expression of full-length, replication competent L1 induces the formation of γ-H2AX foci, a marker for DNA double stranded breaks or replication stress [23]. The L1-induced γ-H2AX foci are dependent on L1 endonuclease activity, suggesting that chromosome cleavage by L1 proteins leads to foci formation. As these are nuclear events, we presume that if γ-H2AX foci are formed, enzymatically active L1 particles have entered the nucleus. We used a tetracycline inducible system [86] to induce L1 expression. In this system, full-length L1 is under the control of a Tet-On promoter, allowing L1 expression upon treatment with doxycycline. When we induced L1 expression in tet-HeLa cells, we found L1-dependent upregulation of γ-H2AX by western blot (Fig 6A). This γ-H2AX upregulation was eliminated when we induced expression of L1 with an endonuclease active site mutation (D145A), suggesting that the γ-H2AX upregulation was dependent on L1 endonuclease activity. In a similar manner, we found that when we induce expression of ORF1 Y282A mutant L1, the γ-H2AX upregulation is reduced (Fig 6B). Because the Y282A mutation presumably disrupts the L1/ESCRT interaction, this result is consistent with the hypothesis that the L1/ESCRT interaction is a prerequisite for nuclear L1 endonuclease activity. This further implies that, in the absence of ESCRT, L1 RNPs are impaired for nuclear entry. Alternatively, it is possible that the Y282A mutant blocks endonuclease activity independent of nuclear entry.
Non-LTR retrotransposons have persisted in eukaryotic genomes for hundreds of millions of years, and with few exceptions, are transmitted vertically [1]. This close relationship between retrotransposons and their hosts, and the small number of L1-encoded proteins, suggests that L1 likely exploits host proteins to complete a round of retrotransposition. The two L1-encoded proteins, ORF1p and ORF2p, provide RNA binding, endonuclease, reverse transcriptase, and nucleic acid chaperone activities–these activities are hypothesized to play a role during TPRT. How L1 completes other aspects of its life cycle is still unclear. Although many L1 restriction factors have been discovered [48–50], additional L1 or host encoded activities required for successful retrotransposition are not well understood, with just a handful of these host factors identified. Disruption of the non-homologous end joining (NHEJ) pathway leads to modest decreases in L1 retrotransposition [87], providing genetic evidence suggesting that the NHEJ pathway may assist resolving TPRT intermediates. However, there is no physical evidence for interaction between NHEJ and L1 complexes. In addition, L1 expression generates DNA damage [24,28,88]; hence, it is not surprising that loss of NHEJ in the presence of DNA breaks leads to a decreased colony forming ability. PCNA has recently been found to physically interact with ORF2p via a canonical PCNA-interacting protein (PIP) box motif [49]. The PIP box is required for the ORF2p interaction, and reduction of PCNA expression by RNA interference leads to lower retrotransposition activity, suggesting a positive role for PCNA in L1 retrotransposition. The L1/PCNA interaction requires endonuclease and reverse transcriptase activity, suggesting that the interaction occurs during or after TPRT.
Although significant progress has been made in defining the L1 interactome, the cytoplasmic events leading to active L1 retrotransposition have remained mysterious. The poly(A) binding protein PABPC1 binds L1 RNPs, presumably through the L1 mRNA poly(A) tail [51]. Reduction of PABPC1 by RNA interference results in lower levels of functional L1 RNPs and reduced retrotransposition. This is consistent with the requirement of a poly(A) tail for L1 retrotransposition [89]. Once made, how a functional L1 RNP traffics to the nucleus to initiate TPRT is not known. L1 RNPs are localized in a diffuse pattern with discrete cytoplasmic foci when visualized by immunofluorescence [18,49,76,90]. The cytoplasmic foci colocalize with stress granule markers. However, even under conditions of massive overexpression and production of L1-containing stress granules, retrotransposition frequency is low (<1% in colony forming retrotransposition assays) and nuclear localization of L1 RNPs is rare. Stress granules are degraded by autophagy [91], and disruption of genes critical for autophagy result in increased L1 expression and retrotransposition activity [92]. Together, this evidence suggests that L1 cytoplasmic foci may simply be RNPs that are sequestered for degradation. The rare L1 RNPs that are trafficking to the nucleus may be impossible to monitor with current methodology.
In this study, we provide evidence that the ESCRT complex interacts with LINE RNPs and is critical for robust LINE retrotransposition, including human L1 retrotransposition. The L1 ORF1p/ALIX interaction depends on an ORF1p YPXnL late domain. In viruses, YPXnL motifs have been shown to directly bind with ALIX [69, 70], suggesting that L1 ORF1p also directly interacts with ALIX via this motif. The established role of ESCRT in membrane budding/fusion suggests that active L1 RNPs may be intimately associated with membranes at some point during a successful retrotransposition event. Membrane association of LINE RNPs is a relatively new concept first suggested by the discovery of esterase domains encoded by the ORF1p of some members of the CR1 non-LTR retrotransposon clade [93]. These esterase domains are classified as lipolytic acetylhydrolases, and more recently the esterase domain of the purified zebrafish LINE element ZfL2-1 ORF1p was shown to be enzymatically active and able to bind phospholipids and liposomes in vitro [94]. An N-terminally truncated version of purified human ORF1p (which lacks an identifiable esterase domain) also co-migrated with lipids in a lipid floatation assay, suggesting that membrane interaction may be universally conserved for LINEs [94]
Although we do not have the ability to visually distinguish L1 RNPs that are in the process of retrotransposition from the predominant background of sequestered L1 RNPs, we can speculate how ESCRT assists L1 replication based on this study, combined with current knowledge. The rarity of horizontal transmission of non-LTR retrotransposons suggests that normal L1 replication does not involve functional L1 RNPs leaving the cell to “infect” other cells (and/or species). Thus, we feel it is unlikely that ESCRT functions to release L1 particles from the plasma membrane. Our data also suggest that membrane trafficking related processes such as glycosylation, secretion, and lysosomal degradation are also dispensable for L1 retrotransposition. When we disrupt the ESCRT/L1 interaction, L1 endonuclease activity in the nucleus is reduced. Although this reduction in endonuclease activity in the nucleus suggests an impairment in the nuclear entry of L1 RNPs, it is also possible that L1 RNPs can enter the nucleus without ESCRT, but are unable to cut DNA. Plausible models for the role of ESCRT during L1 retrotransposition are that ESCRT is involved in the formation of a membrane bound compartment for RNP maturation, akin to the ESCRT-dependent replication compartments of some viruses [85], or ESCRT helps active L1 RNPs traffic to the nucleus by enabling the RNPs to cross a membrane barrier (Fig 7). For example, L1 RNPs could bud into the outer nuclear membrane, or the contiguous endoplasmic reticulum, to form perinuclear or intracisternal vesicles. Fusion of these vesicles with the inner nuclear membrane could deposit L1 RNPs into the nucleus. Although this would be an unusual mechanism for nuclear entry, a similar ESCRT-dependent process in the opposite direction, called nuclear egress, is used by herpesviruses and endogenous RNPs to exit the nucleus [95,96]. In addition, nuclear envelope reformation after mitosis is an ESCRT-dependent process that is topologically equivalent to membrane bud neck constriction [63,64]. Recruitment of L1 RNPs to the nuclear membrane by ESCRT after mitosis could be a route for L1 nuclear entry. Finally, ESCRT could assist L1 retrotransposition in a non-membrane dependent manner. However, this would require a currently undefined function of ESCRT. Distinguishing between these various models will require technological advances in tracking the cellular location of actively retrotransposing L1 particles.
The yeast knockout collection [53] was obtained from the ATCC (GSA-5) and was in the BY4741 background (MATa his3Δ1; leu2Δ0; met15Δ0; ura3Δ0) [97]. Construction of strains JHY146 (MATα his3Δ200 ura3-167 GAL+ lys2::mHIS3AI) and JHY148 (MATα his3Δ200 ura3-167 GAL+ lys2::Zorro3mHIS3AI) have been described previously [52].
HeLa cells were a gift from John Moran [15]. Tet-HeLa cells were from Clontech (Mountain View, CA). Tet-293 cells were a gift from Jef Boeke [51]. All cells were grown in Dulbecco’s Modified Eagle Media (Invitrogen #11965118) supplemented with 10% fetal bovine serum (Invitrogen, #16000069) and 1% Penicillin/Streptomycin (Invitrogen, #15140122). Puromycin resistant plasmids were selected with 1 μg/mL puromycin (Invivogen #ant-pr-1) and maintained with 0.25 μg/mL puromycin. Tet-inducible proteins were induced with 500 ng/mL doxycycline (Fisher Scientific #BP2653). Cells were grown at 37°C unless otherwise specified.
Oligonucleotides used in this study are listed in S9 Table.
All cloned PCR and site directed mutagenesis products described below were completely sequenced. Sequence files of all plasmids are available upon request.
p425Zorro3mMET15AI (Zorro3 retrotransposition plasmid)—The yeast MET15AI retrotransposition marker was made by first PCR amplifying the MET15 gene from the yeast strain BY4742 using primer JH308/JH311, then TOPO cloning the amplicon into pCR2.1 to make pCRMET15. An artificial intron was amplified from pSCZorro3mHIS3AI [52] using the primers JH251/JH255, digested with PvuII/SnaBI, and cloned into the AleI site of pCRMET15 to make pCRMET15AI. The MET15AI marker was removed from pCRMET15AI with XhoI, blunted with T4 DNA polymerase/dNTPs, and cloned into the T4 DNA polymerase blunted SalI site of pSCZorro3 [52] to make pSCZorro3mMET15AI. Oligonucleotides JH28/JH29 were annealed then digested with RsaI/SacI. pRS425 [98] was digested with XhoI, blunted with T4 DNA polymerase, then further digested with SacI. Oligonucleotides JH28/JH29 were annealed then digested with RsaI/SacI, then cloned into the digested pRS425 vector to make pRS425FE. The EagI/FseI GAL1-Zorro3mMET15AI fragment from pSCZorro3mMET15AI was cloned into the EagI/FseI of pRS425FE to make p425Zorro3mMET15AI. Standard site directed mutagenesis techniques were used to make p425Zorro3RTmutmMET15AI, which is identical to p425Zorro3mMET15AI but with D674A, D675A missense mutations in ORF2p.
p415GAL-lacZ was made by cloning the SmaI/HindIII lacZ fragment from pCMV-lacZ (Clontech) into the SmaI/HindIII sites of p415Gal1 [99].
pGAL-HASNF7, pGAL-HAVPS4, pGAL-HAVPS23, pGAL-HABRO1, pGAL-HASNF8, pGAL-HAVPS20 –These plasmids are based on the p426GAL1 vector [99], with a URA3 selection marker, 2 micron replication origin and GAL1 promoter. Using standard PCR cloning procedures, a 3x HA tag was inserted, along with each ORF amplified from BY4741 genomic DNA.
pJM101L1.3, pJM105L1.3, and pJM110 were gifts from John Moran and Haig Kazazian [15,66,100]. These plasmids were derived from a pCEP4 backbone (Invitrogen).
pJM101L1.3Y282A was made site directed mutagenesis of pJM101L1.3 using standard cloning techniques.
pTetPuro was made by digesting pMT302 (gift from Jef Boeke [49]) with NotI/BstZ17I, blunting, and religation.
pTetL1.3 was made by ligating a NotI/BstZ17I fragment from pJM101L1.3 into the NotI/BstZ17I sites of pMT302.
pTetL1.3–282 was made by ligating a NotI/BstZ17I fragment from pJM101L1.3Y282A into the NotI/BstZ17I sites of pMT302.
pTetL1.3–145 was made by ligating a PmlI/BbvCI fragment from pQF288 (a gift from Jef Boeke [16]) and a BbvCI/BamHI fragment from pTetL.13 into the PmlI/BamHI sites of pTetL1.3.
pEF was made by digesting pTracerEF A (Invitrogen) with SphI/BspQI, blunting, and religating.
pEFALIX-2HA was made by filling in annealed JH-IM54/JH-IM55 by PCR, then amplifying this product with JH-IM56/JH-IM57 to make tandem HA tags. ALIX cDNA was amplified from mammalian gene collection clone MGC-17003 using primers JH-IM50/JH-IM61. The amplified ALIX cDNA and tandem HA tags were fused together in a PCR with JH-IM50/JH-IM57. The product was digested with EcoRI and cloned into EcoRI/EcoRV digested pEF. Although this strategy intended to add 3 HA tags, in practice it added 2 HA tags. cDNA expression is under the control of the strong constitutive elongation factor-1α promoter.
pEFCHMP6-2HA was made by filling in annealed JH-IM54/JH-IM55 by PCR, then amplifying this product with JH-IM56/JH-IM57 to make tandem HA tags. CHMP6 cDNA was amplified from mammalian gene collection clone MGC-19477 using primers JH-IM44/JH-IM58. The amplified CHMP6 cDNA and tandem HA tags were fused together in a PCR with JH-IM44/JH-IM57. The product was digested with MfeI/XbaI and cloned into EcoRI/XbaI digested pEF.
pEFCHMP3-2HA was made by filling in annealed JH-IM54/JH-IM55 by PCR, then amplifying this product with JH-IM56/JH-IM57 to make tandem HA tags. CHMP3 cDNA was amplified from mammalian gene collection clone MGC-11028 using primers JH-IM53/JH-IM59. The amplified CHMP3 cDNA and tandem HA tags were fused together in a PCR with JH-IM53/JH-IM57. The product was digested with MfeI/XbaI and cloned into EcoRI/XbaI digested pEF.
pEFALIXs19465mut (cDNA resistant to the s19465 siRNA) was made by amplifying the ALIX cDNA from MGC-17003 using a series of fusion PCRs with primer sets JH-IM50/JH-IM88 and JH-IM87/JH-IM62, digesting with EcoRI/EagI, and cloning into the EcoRI/EagI sites of pEF.
pEFCHMP6s28474mut (cDNA resistant to the s28474 siRNA) was made by amplifying the CHMP6 cDNA from MGC-19477 using a series of fusion PCRs with primer sets JH-IM77/JH-IM81 and JH-IM74/JH-IM67, digesting with EcoRI/XbaI, and cloning into the EcoRI/XbaI sites of pEF.
pEFCHMP3s35990mut (cDNA resistant to the s35990 siRNA) was made by amplifying the CHMP3 cDNA from MGC-11028 using a series of fusion PCRs with primer sets JH-IM78/JH-IM82 and JH-IM75/JH-IM68, digesting with MfeI/XbaI, and cloning into the EcoRI/XbaI sites of pEF.
pGEX6p2Z3ORF1 –Zorro3 ORF1 was amplified with primers JH256/JH257, digested with BamHI/SalI, and ligated into the BamHI/XhoI sites of pGEX-6P-2 (GE Healthcare).
pGEX-3’hORF1 –synthetic human L1 [101] was amplified with primer JH1004/JH1005, digested with EcoRI/BamHI, and ligated into the EcoRI/BamHI sites of pGEX-6P-2.
G01 (rabbit polyclonal anti-Zorro3 ORF1)–The plasmid pGEX6p2Z3ORF1 was transformed into BL(DE3)plysS competent bacteria and GST-Z3ORF1p expression was induced with IPTG. GSTZ3ORF1p was purified with a glutathione sepharose column, followed by FPLC on a Superose 6 10/300 GL column. Fractions of the highest purity and concentration were sent to Proteintech (Chicago, IL) for antibody production. Characterization of G01 is shown in S7 Fig.
JH73 and JH74 (rabbit monoclonal anti-L1 ORF1)–The plasmid pGEX-3’hORF1 was transformed into BL(DE3)plysS bacteria and GST-3’hORF1 was purified as described for GSTZ3ORF1p. Purified GST-3’hORF1 was sent to Epitomics (now Abcam) for rabbit monoclonal antibody production. Hybridoma supernatants were screened by western blot, immunoprecipitation, immunofluorescence, and immunohistochemistry of paraffin embedded samples for the best clones for all applications. Antibody was purified from hybridoma supernatant with Protein A Agarose (Pierce). Characterization of JH73 and JH74 is shown in S8 Fig. Additional characterization is shown in [49].
Commercially available antibodies used: anti-HA affinity matrix (Sigma #11815016001), mouse anti-HA clone 12CA5 (Sigma #11583816001), rat anti-HA clones 3F10 (Sigma #11867423001), rabbit polyclonal anti-ALIX (Bethyl #A302-938A), mouse monoclonal anti-CHMP3 clone F-1 (Santa Cruz #sc-166361), rabbit polyclonal anti-CHMP6 clone FL-201 (Santa Cruz #sc-67231), rabbit polyclonal anti-H3 (Abcam #ab1791), mouse monoclonal anti-γ-H2AX clone JBW301 (EMD Millipore #05-636-I), mouse monoclonal anti-ICAM-1, clone 15.2 (Santa Cruz #sc-107), mouse monoclonal anti-tubulin clone DM1A (Sigma #T6199), rabbit monoclonal anti-EEA1 clone C45B10 (Cell Signaling Technology #3288), rabbit monoclonal anti-Rab7 clone D95F2 (Cell Signaling Technology #9367), mouse monoclonal anti-LAMP2 clone H4B4 (Abcam #ab25631), mouse monoclonal anti-calnexin clone AF18 (ThermoFisher #MA3-027), rabbit monoclonal anti-AIF clone D39D2 (Cell Signaling Technology #5318), mouse monoclonal anti-RCAS1 clone D9 (Santa Cruz #sc-398052), rabbit polyclonal anti-EDC4 (Cell Signaling Technology #2548), rabbit monoclonal anti-eIF3H clone DC91 (Cell Signaling Technology #3413).
Lithium acetate transformation was used to transform all strains of the yeast knockout collection with the plasmid p425Zorro3mMET15AI. For controls, BY4741 was transformed with p425Zorro3mMET15AI or p425Zorro3RTmutmMET15AI. Transformants were selected on SC-leu plates. Three clones from each transformation were used to make 1.5 cm2 patches on SC-leu+galactose plates. After 6 days at 23°C, the patches were replica plated to SC-met plates. Strains that exhibited a qualitatively noticeable decrease in Zorro3 retrotransposition were selected for quantitative Zorro3 retrotransposition assays [52]. Briefly, transformants were inoculated in 3 mL SC-leu cultures and incubated for 3 days at 23°C on a roller drum. After incubation, 200 μL of each culture was plated on an SC-met plate, and 10 μL of a 1:1000 dilution was plated on a YPD plate. Retrotransposition frequency was calculated as (SC-met colonies)/(YPD colonies x dilution factor).
Yeast strains were transformed with p415GAL-lacZ and selected on SC-leu plates. The transformed strains were grown overnight in SC-leu media with glucose, washed with H2O, then resuspended and incubated overnight in SC-leu with galactose. Cell concentrations of cultures were normalized to a constant OD600. 1 mL of each normalized culture was used to assay for β-galactosidase activity using a previous described method [102]. Briefly, the pelleted culture was resuspended in 500 μL of Z buffer. 50 μL of 0.1% SDS and 50 μL chloroform were added and the solution was vortexed. 100 μL of a 4 mg/mL stock of o-nitrophenol-β-D-galactoside (ONPG) was added and incubated at 37 degrees for 15 minutes. The reaction was quenched with 500 μL of 1 M Na2CO3. 100 uL of each supernantant was transferred into a 96 well plate, and OD420 was measured with a plate reader. Wild-type BY4741 transformed with p415GAL-lacZ and grown in glucose was used as to measure background activity and was subtracted from all other values. β-galactosidase activity of BY4741 transformed with p415GAL-lacZ was set to 1, and for each experiment, the activity of knockout strains were normalized to this value.
Strains JHY146 or JHY148 were transformed with pGAL-HASNF7, pGAL-HAVPS4, pGAL-HAVPS23, pGAL-HABRO1, pGAL-HASNF8, or pGAL-HAVPS20 and selected on SC-ura plates. Zorro3 and the respective ESCRT proteins were induced in SC-ura+galactose cultures for 16 hours, and lysed in a BioSpec mini bead beater with 25 mM Tris-Cl pH 7.4, 150 mM NaCl, 0.5% Triton X-100, 0.5% Na-deoxycholate, 1 mM EDTA, 4% glycerol, 1% protease inhibitor cocktail (Sigma #P2714), 1 mM PMSF, 10 mM MgCl2, and acid washed glass beads. Where indicated, 0.5 μg/mL RNase A was added to the IPs. Antibody G01 was used for IP and western blot of Zorro3 ORF1p, anti-HA affinity matrix was used for HA IP, and anti-HA clones 12CA5 or 3F10 were used for HA western blot.
To knockdown expression of ESCRT proteins and test L1 retrotransposition, we modified an established retrotransposition assay [66]. 2 x 105 HeLa cells/well were seeded in 6-well plates. 24 hours after seeding, cells were transfected with 1 μg of L1 reporter plasmid (pJM101L1.3 or pJM105L1.3) and 15 pmol of silencer select siRNA (Ambion—s19465 for ALIX, s29474 for CHMP6, s35990 for CHMP3, or scramble negative control) using Lipofectamine 2000 (Invitrogen) according to the manufacturers protocol. 24 hours post-transfection, G418 selection (600 μg/mL) was started and continued for 11–12 days. G418-resistant colonies were fixed and stained with 250 μl/well of modified Giemsa stain (Sigma #GS500). To assess whether knockdown affected transfection efficiency or viability, we performed parallel co-transfections with 10 ng pcDNA3 (Invitrogen) in place of the L1 reporter plasmid. G418-resistant colonies from the retrotransposition assay were normalized to the respective pcDNA3 colony forming assays to give “normalized L1 activity”. For siRNA rescue retrotransposition experiments, co-transfections were performed as described above with the following amounts of nucleic acids: 0.5 μg L1 reporter construct, 0.5 μg siRNA resistant cDNA or empty vector, and 15 pmol siRNA.
To test the effectiveness of siRNA knockdown, we performed separate experiments where we co-transfected HeLas with pEFALIX-2HA, pEFCHMP6-2HA, or pEFCHMP3-2HA and the relevant siRNA or scramble negative control siRNA. Knockdown was assessed by western blot with anti-HA 12CA5 antibody.
To perform standard L1 retrotransposition assays (in the absence of siRNA knockdown), 2.5 x 105 HeLas/well were seeded in 6-well plates. 24 hours after seeding, cells were transfected with 1 μg L1 reporter plasmid and 3 μL XtremeGENE9 (Roche) according to the manufacturer’s protocol. 24 hours post-transfection, G418 selection (600 μg/mL) was started and continued for 11–12 days. G418-resistant colonies were fixed and stained with 250 μl/well of modified Giemsa stain (Sigma #GS500). To assess transfection efficiency, extra wells were set aside for each condition and analyzed for ORF1p expression by western blot. In cases where inhibitors were added to the retrotransposition assays, the inhibitors were added to the media 1 hour post-transfection. The final concentration of drugs used were as follows: 500 μM amantadine (Sigma #A1260), 10 μM chloroquine (Sigma #C6628), 25 μM kifunensine (Sigma #K1140), 1 mM D-mannojirimycin (Sigma #D9160), 2.5 μM golgicide (Sigma #G0923), 75 μM Exo1 (Sigma #E8280).
RNP isolation and LEAP assays were based on previous published work [71]. Briefly, to isolate L1 RNPs, 2 x 106 HeLa cells were plated in a T75 cell culture flask. Cells were transfected 24 hours later with 20 μg of L1 plasmid per flask. Three days post-transfection, medium was replaced by DMEM supplemented with 0.2 mg/mL hygromycin B and exchanged daily. On day 7 post-transfection, one T-75 flask was seeded with 2 x 106 untransfected HeLa cells to be used as a negative control. On day 10 post-transfection, transfected and untransfected cells were harvested and whole-cell lysates were prepared and subjected to ultracentrifugation as described previously [71].
The LEAP reaction was performed as described previously [71] with the following changes: a 1 μL aliquot of the pelleted samples was incubated with 50 mM Tris-HCl (pH 7.5), 50 mM KCl, 5 mM MgCl2, 10 mM DTT, 0.4 mM LEAP primer, 20 U RNaseOut (Invitrogen), 0.2 mM dNTPs and 0.05% Tween 20 in a final volume of 50 μL for 1 hour at 37°C. After incubation, 1 μL of LEAP reaction was used in a standard 30 μL PCR using AmpliTaq DNA polymerase (Applied Biosystems) and 0.4 mM of each linker PCR primer and L1.3 end primer according to the manufacturer’s protocol. PCR conditions as previously described [71]. The complete reaction was visualized on a 2% agarose gel with Ethidium Bromide.
Tet-293 cells were transfected with pTetPuro, pTetL1.3, or pTetL1.3–282. 24 hours post-transfection, cells were selected with puromycin. 100 mm plates of selected cells were transfected with 21 μg of pEFALIX-2HA in N,N-bis(2-hydroxyethyl)-2-amino-ethanesulfonic acid (BES) buffer as previously described [103]. L1 expression was induced with doxycycline, and after 48 hours of induction cells were lysed using a Precellys 24 homogenizer (Bertin Instruments) and acid-washed beads (Sigma #G8772) in 25 mM HEPES pH 7.4, 200 mM NaCl, 1 mM MgCl2, 1 mM CaCl2, 1 x protease inhibitor cocktail (Sigma #S8830), 40 units DNase I, 2.5 mg RNaseA, and 10% glycerol. L1 ORF1p was immunoprecipitated with JH74. Western blot analysis was performed with JH73 (ORF1p) and anti-ALIX or 12CA5 (HA).
HeLa cells were seeded at 2 x 105 cells/dish in 35 mm glass bottom dishes (MatTek #P35G-0-10-C). 24 hours after seeding, cells were treated with 10 μM chloroquine or 50 μM amantadine. After 24 hours, fresh media/drug was added along with 1 μM acridine orange, and incubated at 37°C for 1 hour. The media was washed and replace with PBS, then directly imaged on an Olympus FV1000 confocal miscroscope with a FITC filter.
Tet-HeLa cells were transfected with pTetPuro, pTetL1.3, pTetL1.3–282, or pTetL1.3–145 and transfected populations were selected with puromycin. In selected cells, L1 expression was induced with doxycycline, and whole cell lysates were analysed by western blot for H3 and γ-H2AX at 72 hours.
10,000 cells containing the plasmid pTetL1.3 were seeded on coverslips treated with 0.01% poly-L-lysine in 24-well plate. Approximatively 16 hours after seeding, L1.3 expression was induced with 0.5 ug/ml doxycycline. Cells were fixed 48 hours after induction with 4% paraformaldehyde in PBS pH 7.4 for 15 minutes at room temperature (RT). After fixation, cells were washed 3 times with PBS pH7 .4. For calnexin, RCAS1, EEA1, RAB7 and AIF staining, cells were permeabilized with PBS pH 7.4 containing 0.1% Triton X-100 for 15 minutes at RT. Nonspecific interactions were blocked with 10% normal donkey serum (NDS) in PBS for 1 hour at RT. Staining with primary antibody, diluted in PBS pH 7.4 and 2% NDS was done overnight at 4°C. The following dilutions for primary antibody were used: anti-ORF1p clone 4H1 (mouse monoclonal antibody) 0.75 μg/ml, anti-ORF1p clone JH74 (rabbit monoclonal antibody) 0.15 μg/ml; anti-calnexin 1:1000; anti-RCAS1 1:200; anti-EEA1 1:200; anti-RAB7 1:100; anti-AIF 1:400. After primary antibody incubation, cells were washed 3 times with PBS pH 7.4. Secondary antibodies were used at 1:1000 dilution in PBS pH7.4 with 2% NDS for 1–2 hours at RT. To detect anti-ORF1p antibodies, we used Alexa Fluor 488 Donkey anti-Mouse IgG (ThermoFisher # A21202) or Alexa Fluor 488 Donkey anti-Rabbit IgG (ThermoFisher # A21206). To detect antibodies against organelle markers, we used Alexa Fluor 594 Donkey anti-Mouse IgG (ThermoFisher # A21203) or Alexa Fluor 594 Donkey anti-Rabbit IgG (ThermoFisher #A21207). After incubation with secondary antibody, cells were washed 2 times with PBS pH 7.4, incubated with 1 μg/ml DAPI in PBS pH 7.4 for 5 minutes and washed once with PBS pH 7.4. Coverslips were mounted on glass slides with ProLong Gold antifade reagent (Life Technologies #P36930). For LAMP2 and eIF3h staining, cells were incubated with PBS pH 7.4 with 0.3% Triton X-100 and 5% NDS for 1 hour at RT to permeabilize cells and block nonspecific protein interactions. Staining with primary antibody was done overnight at 4°C in PBS pH 7.4 with 0.3% Triton X-100 and 1% BSA. Anti-LAMP2 antibody was used at a 1:100 dilution and anti-eIF3h antibody at a 1:400 dilution. Cells were washed 3 times with PBS pH 7.4 and incubated with secondary antibodies diluted to 1:500 in PBS pH 7.4 with 0.3% Triton X-100 and 1% BSA for 1–2 hours at RT. After secondary antibody, cells were washed, DAPI stainied and mounted as described above.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.